# Don't Replace Me — full text index Source: https://dontreplace.me/ · Generated: 2026-09-16 ## Homepage Don't Replace Me — A Survival Guide to the AI Apocalypse You might get replaced. ...unless you read this first. AI Assessment Read the Intro Free by someone who uses AI to build things. What this book is Every third post on your feed says AI is replacing everyone. Usually right before someone sells you a course. Meanwhile your company still can't get people to update their Slack status. The guy selling the masterclass hasn't shipped software in four years. And the headline that scared you this morning was written to sell ads, not inform you. Here's what's real: AI replaces tasks, not people. But if someone in your field picks up the tools and you don't, you'll notice. Not next year. This quarter. This is the honest version. No hype. No panic. Just what's changing, what isn't, and what to do Monday. From the book A few of the 24 rules. 200+ pages. 24 chapters. here's a taste. Rule #1 Fear Is a Subscription Model Unsubscribe from the panic. Rule #5 It's Not Smart. It's Fast. Confidence is unrelated to accuracy. Rule #8 Taste Can't Be Downloaded The robot doesn't know what good looks like. You do. Rule #18 Build a Moat The moat is everything. Keep filling it. Free chapter The chapter that kills the panic. Subscribe to the author's newsletter. You get the introduction chapter as an ebook right away, plus occasional field notes on staying useful while the robots figure themselves out. Send Me the Chapter intro chapter lands immediately. newsletter is worth reading. What's inside Seven parts. One arc. from "wow, I'm fucked" to "oh, I'm fine" I Don't Panic In which we kill the fear before it kills your career II Know Your Enemy What AI actually is, explained without making you want to die III What They Can't Take from You Your unfair advantages, and why you should stop undervaluing them IV Your First 72 Hours Stop reading about AI and start using it V Field Survival What happens when you actually start using this stuff at work VI Fortify Your Position Playing the long game while everyone else panics VII The New World Turns out, it's not an apocalypse after all Who wrote this Dmitry Kargaev Designer turned AI engineer. Among the top 0.1% of first ChatGPT users. Builds everything with AI now. deeflect.com “ I'm one of the biggest AI haters because of IP theft, deepfakes, fake news and other stuff. I still decided to use it and harness it instead of running away from it. ” Not selling a course. Not selling a masterclass. If I do, you have my permission to throw this book at me. The book A Survival Guide to the AI Apocalypse. 200+ pages. 24 chapters. Available now. Get the Book available now Paperback · Hardcover · Ebook × AI Readiness Assessment How much do you know? How exposed is your job? --- ## AI A/B Testing Prompts: Run Better Experiments URL: https://dontreplace.me/articles/ai-ab-testing-prompts-run-experiments-without-fooling-yourself AI promptsA/B testingexperimentationproduct analyticsgrowth AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. August 21, 2026 Updated August 21, 2026 A/B testing sounds scientific because it has letters, percentages, and dashboards. Then someone peeks after six hours, finds a green number, and ships the variant because the quarterly target is looking hungry. AI A/B testing prompts can help structure an experiment brief, challenge a mushy hypothesis, draft variant directions, identify confounders, and summarize de-identified results. They cannot make broken tracking trustworthy, turn a tiny sample into evidence, guarantee causality, or decide whether a conversion lift is worth harming customers. AI can organize the lab notebook. Humans still design the experiment, validate the data, review the statistics, protect users, and own the decision. These ten templates are built for product managers, growth teams, marketers, designers, analysts, and founders who want useful experiments rather than screenshot theater. For the broader numbers work, pair them with AI data analysis prompts. When the result affects a launch, use a real go/no-go decision process instead of asking a chatbot to bless it. What A/B testing can actually tell you A controlled A/B test compares outcomes for groups exposed to different experiences. When assignment, instrumentation, duration, and analysis are sound, it can estimate whether a specific change caused a measurable difference for the tested population during the tested period. That sentence has more caveats than a software license because experiments are easy to contaminate. A useful test needs: One decision: what the team will do differently after learning the result. A falsifiable hypothesis: a predicted change, audience, mechanism, and outcome. Reliable assignment: eligible units are assigned consistently without leakage. Validated instrumentation: events mean what the dashboard claims they mean. A primary metric: chosen before results, not after browsing every chart. Guardrails: measures that catch damage hidden behind the primary win. An appropriate sample and duration: reviewed by someone statistically qualified. A predeclared decision rule: ship, iterate, continue, or stop under defined conditions. A test does not prove a design is universally better. It does not explain every reason behind behavior. It does not erase seasonality, interference, novelty effects, missing data, or implementation bugs. Statistical significance is not practical importance, and practical importance is not ethical permission. If the team wants to understand why people struggle, add usability testing. Numbers can locate a difference. Real observation can reveal the friction behind it. The reusable A/B testing prompt formula Use this base instruction with any prompt below: “Act as an experiment-planning assistant. I am evaluating [decision, audience, surface, baseline, constraint, and deadline]. Use only the verified information I provide. Produce [artifact] with assumptions, exclusions, risks, open questions, required instrumentation, guardrail metrics, and human review points. Separate facts, hypotheses, calculations, observations, interpretations, and decisions. Do not invent data, sample sizes, statistical significance, causal explanations, customer reactions, or confidence intervals.” That separation matters. “Variant B had a 3.2% observed checkout rate” is a measurement. “The shorter form caused the change” is a causal interpretation that depends on valid assignment, clean instrumentation, and analysis. “Ship it” is a decision involving value, risk, and customer impact. Never paste customer names, emails, account identifiers, event-level behavior logs, health or financial information, credentials, confidential revenue data, unreleased strategy, proprietary experiment results, or raw analytics exports into an unapproved AI tool. Use aggregated or synthetic examples and approved systems. Confirm retention, access, training, residency, and deletion rules with the appropriate privacy and security owners. What to collect before prompting Do not type “make me an A/B test.” Give the model a compact, verified packet. InputWhy it mattersHuman check Decision to informPrevents testing for entertainmentProduct owner confirms Eligible audienceDefines who can enter the testAnalyst checks exclusions Baseline metricGrounds expected changeData owner validates window Proposed mechanismMakes the hypothesis falsifiableResearch/design reviews Primary metricLimits metric shoppingAnalyst approves definition Guardrail metricsExposes hidden harmProduct, support, risk review Assignment unitPrevents cross-group leakageExperiment owner verifies Instrumentation mapConnects behavior to eventsEngineering runs QA Sample and duration planControls uncertainty and cyclesQualified analyst reviews Decision ruleStops post-result improvisationAccountable owner signs off Fingerprint the exact build, feature flags, event definitions, audience query, experiment configuration, start time, exclusions, and analysis version. Otherwise the team may analyze a test that never existed in the form described. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI A/B testing prompts Replace bracketed text with sanitized, verified details. These prompts create planning and communication artifacts. They do not run the experiment or certify the statistics. 1. Turn a goal into a falsifiable hypothesis “Using this business goal, user problem, current experience, baseline evidence, proposed change, intended audience, and constraints: [paste], draft three falsifiable hypotheses. Use the format: If we [change], then [specific audience] will [observable behavior], causing [primary metric] to change because [mechanism]. For each, list assumptions, contradictory evidence, guardrails, and what result would fail to support the hypothesis. Do not invent a minimum detectable effect.” “Improve engagement” is not a hypothesis. It is a wish wearing office clothes. A useful hypothesis links one intervention to one observable behavior and explains why the primary metric should move. Humans should challenge the mechanism before testing. If nobody believes the causal story, a positive result may still be hard to interpret. If the change bundles five ideas, split it or admit that the test can only evaluate the bundle. 2. Define the primary metric and guardrails “Given this decision, hypothesis, user journey, event dictionary, baseline window, known risks, and business constraints: [paste], propose one primary metric and up to five guardrails. For every metric, specify numerator, denominator, eligible population, attribution window, exclusions, direction of improvement, known failure modes, and data owner. Flag proxy metrics and metrics that could reward harmful behavior. Require analyst approval.” A primary metric prevents the team from opening twenty charts and promoting whichever one turned green. Guardrails make sure a conversion win did not also increase refunds, errors, support contacts, unsubscribe rates, latency, or accessibility barriers. Do not let AI invent definitions from event names. checkout_complete might fire twice, exclude mobile, or mean the page loaded rather than payment settled. The instrumentation owner must verify semantics with real logs and code. 3. Document the baseline without cherry-picking “Using these approved aggregate metrics, dates, audience definitions, release history, traffic patterns, and known incidents: [paste], create a baseline brief. Show the exact period, central tendency, variability, sample counts, weekday or seasonal patterns, segment differences that were specified in advance, missing-data notes, and events that make the period unrepresentative. Do not select a friendlier window or calculate values absent from the input.” A baseline taken during an outage, holiday, campaign spike, or previous experiment can distort the plan. Compare multiple relevant windows when appropriate, but document why one will govern the test. If baseline behavior is unstable, fix the measurement or delay the experiment. A sophisticated prompt cannot stabilize a metric that changes whenever marketing sends an email. 4. Brainstorm variants that test the mechanism “Given this verified hypothesis, current experience, research evidence, brand rules, accessibility requirements, technical constraints, and prohibited patterns: [paste], propose six materially different variant directions. For each, explain which part of the mechanism it tests, what must remain constant, implementation risk, accessibility considerations, and possible unintended behavior. Do not write deceptive scarcity, hidden defaults, obstructive cancellation, or other dark patterns.” Variant generation is where AI is fast and occasionally useful. It can create breadth before humans choose a coherent treatment. The useful question is not “Which button color wins?” It is “Which change best tests the proposed mechanism without smuggling in unrelated differences?” Designers, researchers, legal reviewers, and engineers still decide what is acceptable and buildable. A manipulative variant can lift a short-term metric while damaging trust. That is not optimization. That is borrowing against the customer relationship. 5. Check eligibility and assignment rules “Review this experiment audience, eligibility query, assignment unit, exposure event, exclusion rules, mutual-exclusion groups, device/account behavior, and rollout configuration: [paste]. Produce a preflight checklist for sample-ratio mismatch, repeat assignment, cross-device contamination, household or team interference, bot traffic, employee traffic, late enrollment, and users exposed before assignment. Mark every item as verified, unverified, or not applicable based only on supplied evidence.” Randomization is not magic if one person can enter both groups, accounts influence coworkers, or the exposure event fires after the outcome. Decide whether the assignment unit should be user, account, household, session, organization, location, or something else. A checklist helps humans inspect the setup; it does not prove the setup is valid. Run an A/A test or platform-specific validation where appropriate. Ask the experiment platform owner and analyst to review the actual configuration. 6. Identify confounders and novelty effects “Using this test plan, product calendar, campaign schedule, release calendar, audience behavior, seasonality notes, and dependency map: [paste], create a threat-to-validity register. Cover concurrent launches, novelty and learning effects, spillover, carryover, selection bias, instrumentation drift, outages, sample-ratio mismatch, missing data, and external events. For each, state the detection signal, mitigation, owner, and whether it could invalidate the result.” Not every odd result needs a clever story. Sometimes a payment provider failed. Sometimes Variant B launched only on a newer app version. Sometimes users clicked because the new thing was new. Use change impact analysis prompts to map dependencies before launch, then have owners verify them. AI can list familiar risks. The team knows which campaign, migration, pricing change, or executive demo is about to collide with the test. 7. Build an instrumentation QA checklist “From this experiment spec, event dictionary, analytics schema, assignment logic, expected user paths, error states, platforms, and environments: [paste], draft an instrumentation QA checklist. Include control and variant exposure, event deduplication, property values, timestamps, identity stitching, consent states, exclusions, failed actions, retries, latency, and dashboard reconciliation. Add expected results and evidence links for each check. Do not mark anything passed.” Run the checklist in real environments with known test accounts. Compare client events, server events, warehouse rows, and dashboard totals where applicable. Confirm the control is unchanged and the variant is actually visible to assigned users. This is a good companion to an AI QA checklist, but a generated checklist is only a starting point. Engineers and analysts must inspect actual payloads, logs, queries, and experiment-platform output. 8. Write the decision rule before seeing results “Using this approved hypothesis, primary metric, guardrails, practical-effect threshold, statistical method selected by the analyst, planned sample, duration, stopping policy, and risk tolerance: [paste], draft a pre-analysis decision table for ship, iterate, continue, stop, and investigate. Include conditions for guardrail harm, inconclusive results, sample-ratio mismatch, instrumentation failure, conflicting metrics, and practically trivial effects. Do not invent thresholds or recommend optional stopping.” Writing the rule first makes motivated reasoning harder. It prevents “We always cared most about seven-day retention” from appearing immediately after the click-through result disappoints everyone. Do not ask AI to choose the statistical method or sample size unless a qualified analyst is using it as a drafting aid and verifies every assumption. Fixed-horizon, sequential, Bayesian, cluster-randomized, and other designs have different rules. Mixing them casually is how dashboards become fan fiction. 9. Summarize results without overclaiming “Using only these approved aggregate results, pre-analysis plan, quality checks, metric definitions, confidence or posterior outputs calculated by the analyst, segment plan, anomalies, and experiment dates: [paste], draft a results summary. Separate observed effects, uncertainty, guardrail outcomes, data-quality issues, preplanned segment findings, exploratory findings, limitations, and unanswered questions. Do not claim causality if assignment or instrumentation failed. Do not call an inconclusive result ‘no effect.’” A clear summary should let a skeptical reader trace each statement to a metric, query, notebook, or review. Include absolute values as well as relative change. “Up 20%” can mean moving from five users to six. Avoid segment fishing. If the overall result is flat but one tiny audience slice looks miraculous, label it exploratory unless that segment was predeclared and adequately powered. Use the finding to form a future hypothesis, not to manufacture a victory lap. 10. Draft a human-reviewed recommendation “Given this experiment decision, approved results summary, customer research, guardrail outcomes, implementation cost, reversibility, operational risk, strategic context, and unresolved questions: [paste], draft ship, iterate, stop, and follow-up options. For each, show supporting evidence, counterevidence, affected users, risks, monitoring needs, rollback trigger, and accountable approver. Recommend no option unless the supplied decision rule clearly supports it.” The winning variant is not automatically the right product choice. A tiny lift may not justify permanent complexity. A larger lift may depend on a pattern the company should not use. A flat result can still expose broken assumptions or save months of development. Record the final rationale with an AI decision log prompt and monitor the released experience with post-launch monitoring prompts. Experiments estimate behavior under test conditions. Production keeps moving. Common ways teams fool themselves Peeking and stopping on green Repeatedly checking a conventional fixed-horizon test and stopping when it crosses a threshold can inflate false positives. Follow the analysis and stopping plan chosen before launch. If the organization needs continuous monitoring, use a method designed for it and have a qualified analyst configure it. Treating “not significant” as “no difference” An inconclusive test may be underpowered, noisy, short, or compatible with a range of meaningful effects. Report the uncertainty. Do not turn “we could not distinguish the effect” into “the experiences are identical.” Optimizing a proxy into the ground Clicks, opens, time on page, and task starts can be useful proxies. They can also reward confusion, accidental taps, or coercion. Pair proxy metrics with downstream outcomes and guardrails that represent user value. Changing the test midstream Editing traffic allocation, eligibility, variants, metrics, or event definitions can alter interpretation. Sometimes changes are necessary. Log them, involve the analyst, and decide whether the experiment must restart. Letting the model write the story first If AI receives only the winning chart, it will produce a persuasive explanation because that is what text models do. Give it the preregistered hypothesis, complete approved results, anomalies, quality checks, and limitations. Then make a human defend every causal sentence. For a broader practical approach to working with these tools, read the no-BS guide to using AI at work. The short version: delegate structure, not accountability. Frequently asked questions Can ChatGPT calculate the sample size for an A/B test? It can explain inputs or format a calculation reviewed by an analyst, but it should not be the authority. Sample planning depends on the metric distribution, baseline, minimum effect worth detecting, assignment unit, design, error rates, power, expected attrition, and analysis method. Wrong assumptions can produce a precise-looking bad answer. Use approved experimentation software and qualified statistical review. Can AI analyze raw customer experiment data? Only if company policy, consent, contracts, privacy controls, security review, and the approved tool explicitly permit it. Prefer aggregate, de-identified data. Raw event logs often contain identifiers, behavior trails, URLs, free text, or sensitive attributes. Keep analysis in governed systems and provide the model only the minimum safe context. How many variants should an experiment include? There is no universal number. More variants divide traffic and increase analysis complexity. Use the smallest set needed to answer the decision. Every treatment should represent a meaningful hypothesis, not a designer’s favorite shade of blue. Have the analyst account for multiple comparisons and the traffic available. How long should an A/B test run? Long enough to satisfy the preapproved sample and duration plan and capture relevant behavior cycles, without violating the chosen stopping method. A week is not automatically enough; a month is not automatically better. Subscription renewal, repeat use, weekday patterns, novelty, and delayed outcomes may require different windows. What if the primary metric wins but a guardrail gets worse? Follow the decision rule written before launch. Investigate severity, data quality, affected users, reversibility, and whether the harm is acceptable at all. Do not hide the guardrail in an appendix. A faster checkout that increases duplicate charges is not a winner. Can AI choose the winning variant? No. It can map verified evidence to a predeclared decision table. Humans must decide whether the result is trustworthy, practically valuable, ethical, affordable, reversible, and aligned with customer interests. The accountable owner signs the decision. Is A/B testing always better than usability research? No. A/B tests estimate differences in measured behavior. Usability research can reveal confusion, expectations, accessibility barriers, and reasons behind behavior. Early concepts, low-traffic products, safety-critical flows, and questions about meaning may need interviews, observation, expert review, simulations, or other methods instead. What should we do with an inconclusive result? Report it honestly. Check data quality, uncertainty, achieved sample, treatment strength, and whether the tested change actually represented the hypothesis. Then stop, iterate, or run a better-designed follow-up based on the original decision value. Do not rename inconclusive as “control won.” The rule that keeps experiments honest Use AI to improve the brief, expose missing fields, draft checklists, and make results easier to review. Do not use it to manufacture certainty. The boundary is simple: AI structures the hypothesis; humans decide whether it matters. AI drafts the metric table; analysts validate definitions and methods. AI proposes variants; designers and product owners protect users. AI lists confounders; engineers and analysts inspect reality. AI summarizes approved results; humans own interpretation and action. If you want the larger field guide for staying useful while AI handles more drafting and organizing, Don’t Replace Me is the low-drama version: use the machine for speed, keep judgment and accountability human. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance AI Account Planning Prompts: Build a Better Plan 10 AI account planning prompts to organize buyer evidence, map stakeholders, expose risks, test strategy, and plan next actions without inventing intel. AI promptsaccount planningsales strategy --- ## AI Acceptance Criteria Prompts: 10 Templates URL: https://dontreplace.me/articles/ai-acceptance-criteria-prompts-stop-shipping-vague-work acceptance criteriaAI promptsproduct managementQArequirements AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. July 10, 2026 Updated July 10, 2026 Acceptance criteria are one of those things that looks like work but often isn't. Someone writes "user can log in," everyone nods, the ticket moves to dev, and three weeks later you're arguing about whether "log in" includes SSO, expired sessions, locked accounts, or a mobile browser on iOS 16. It doesn't. Nobody wrote that down. AI acceptance criteria prompts can fix a lot of that friction. Not by thinking for you. By forcing you to write things down before the conversation gets expensive. Here's the honest version of how this works: AI is fast at structure and embarrassingly bad at judgment. It'll hand you a neat Given/When/Then block with the confidence of someone who has definitely shipped software before. Some of it will be right. Some of it will be invented. You're the one who knows which is which. That's the whole job. Why acceptance criteria are broken before AI even enters the room Most acceptance criteria fail at the source. Someone had a vague idea, typed a vague ticket, and called it a requirement. By the time it hits QA, the criteria have three interpretations and none of them are what the stakeholder meant. The usual causes: tickets written in a meeting when everyone was already thinking about lunch, scope captured verbally and never reconciled, "done" defined by whoever felt like closing the ticket, edge cases discovered in production by angry users. AI doesn't fix any of that. It just processes whatever you give it. And as Rule #13 in Don't Replace Me makes clear: garbage in, garbage out. Give AI a vague request and it returns confident fake precision. The output looks like real criteria. It has headers and bullet points and test scenarios. It's structured nonsense dressed up in a table. The prompts below treat AI as a drafting assistant, not a product owner. You bring the judgment. It brings the structure, the edge-case questions, and the speed. If you're new to using AI for structured work tasks, the no-BS starter guide to using AI at work covers the mindset before you get into specific prompts. What to never paste into AI tools before you start Before the prompts: a real warning, not a legal disclaimer someone copy-pasted. Do not put any of the following into unapproved AI tools: Customer PII or employee records Credentials, API keys, or access tokens Confidential strategy, unreleased product details, or roadmap specifics that aren't public Security vulnerabilities or known exploits Legal disputes, contracts, or settlement terms Financial forecasts, acquisition plans, or regulatory filings Medical or safety-critical information Sensitive client information of any kind Check what your organization has approved for AI use. If you're not sure, ask before you paste. "I used the free tier and it seemed fine" is not a data handling policy. The question of whether to tell your boss you use AI is related here, especially if your company doesn't have a clear AI policy yet. AI acceptance criteria prompts: the reusable formula Every prompt in this list follows a pattern you can adapt: Role: You are a product requirements analyst helping a team write testable acceptance criteria. Context: [Paste the vague request, ticket, or notes here. Remove anything sensitive.] Task: [Specific thing you want: criteria, edge cases, test scenarios, stakeholder questions, etc.] Constraints: [Any known limits: platform, audience, regulatory environment, done definition.] Output format: [Given/When/Then, numbered list, table, plain prose, etc.] The formula matters because AI responds to specificity. "Write acceptance criteria for this" gets you generic output. "Write Given/When/Then acceptance criteria for a logged-out user attempting to access a premium feature on mobile web, flagging any scope you're assuming" gets you something you can actually review. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → The 10 prompts Prompt 1: Turn a vague request into testable acceptance criteria Use this when a ticket says something like "make the checkout flow better." You are a product analyst. I have a vague feature request: [paste request]. Convert it into 5-8 testable acceptance criteria in Given/When/Then format. For each criterion, flag any assumption you're making about scope, user type, or system behavior. Do not invent business rules. Where you're uncertain, write [ASSUMPTION: describe it] so I can verify with stakeholders. Review everything flagged as an assumption. Those are your open questions. Answer them before the ticket moves. Prompt 2: Convert notes into Given/When/Then examples Use this after a meeting where someone captured requirements in a doc but they're messy. Here are my raw notes from a requirements conversation: [paste notes, no sensitive data]. Convert these into Given/When/Then acceptance criteria. Preserve the original intent. Where the notes are ambiguous or contradictory, flag the specific conflict instead of resolving it yourself. Output as a numbered list, one scenario per item. The goal isn't clean output. It's surfacing the conflicts before they become bugs. Prompt 3: Find ambiguity in an existing ticket Use this when you suspect a ticket is vaguer than it looks. Here is a draft acceptance criterion or ticket description: [paste it]. Identify every term or phrase that could be interpreted more than one way. For each ambiguity, explain the different possible interpretations and suggest a clarifying question to ask the stakeholder. Do not resolve the ambiguities yourself. This is one of the most useful things AI can do: play dumb in exactly the right places. If you use the AI workflow audit prompts framework to map your ticket-writing process, this prompt fits naturally into the definition-of-ready gate. Prompt 4: Map happy paths and edge cases Use this when you've got the main flow covered but you know there's something lurking. Here is a user flow for [feature name]: [describe the happy path]. Generate a list of edge cases and error states that a QA team should consider. Organize by: input edge cases, system state edge cases, permission/role edge cases, timing/concurrency edge cases, and failure/error states. Flag any that might have accessibility or security implications. Do not treat this as a complete test plan. Treat it as a starting checklist that your QA team then owns. Prompt 5: Define done vs. not done Use this to create a clear scope boundary before something goes into a sprint. I'm writing acceptance criteria for [feature]. Help me define what is explicitly IN scope (done) and what is explicitly OUT of scope (not done) for this ticket. Use the following context: [paste relevant details]. For each out-of-scope item, note whether it needs a separate ticket or a stakeholder decision. Output as two lists: Done When and Not Done (Deferred or Descoped). The "Not Done" list is often worth more than the "Done" list. It stops scope creep before it starts. Prompt 6: Check accessibility, privacy, security, and data requirements Use this before a ticket goes to dev, not after. Here are the acceptance criteria for [feature]: [paste criteria]. Review them for gaps in four areas: accessibility (WCAG 2.1 AA), privacy (data collection, retention, consent), security (authentication, authorization, input validation), and data handling (PII exposure, logging, third-party data sharing). For each area, flag what's missing or unclear. Note that your output should prompt human review by qualified specialists, not replace it. This prompt does not replace your security team, your privacy counsel, or your accessibility reviewer. It surfaces the questions so you bring the right people into the conversation earlier. For anything touching real risk, the AI risk assessment prompts cover how to structure that conversation with the right stakeholders. Prompt 7: Create QA test scenarios Use this to give a QA team something concrete to start from. Here are acceptance criteria for [feature]: [paste criteria]. Generate a QA test scenario list including: positive test cases (expected behavior), negative test cases (invalid inputs, unauthorized access, system errors), boundary tests, and regression considerations. Format as a table with columns: Test ID, Scenario, Precondition, Steps, Expected Result. Mark any scenario where the expected result is unclear given the current criteria. Every "expected result unclear" row is a ticket that will get reopened if you ship it. Prompt 8: Prepare stakeholder clarification questions Use this before a refinement session so you arrive with specific questions, not general confusion. I'm preparing for a requirements conversation about [feature]. Here are the draft criteria so far: [paste draft]. Generate a list of clarifying questions to ask stakeholders, organized by: scope questions, business rule questions, user/persona questions, integration and dependency questions, and launch or rollout questions. Phrase each question plainly. I'll use these to run the conversation, not replace it. Arriving with sharp questions is how you run a good requirements session. Arriving with AI-generated criteria as if they're final is how you run a bad one. Prompt 9: Rewrite criteria for non-technical stakeholders Use this when you need to get sign-off from someone who doesn't read tickets. Here are acceptance criteria written for engineers: [paste criteria]. Rewrite them in plain language for a non-technical stakeholder review. Focus on: what the user will be able to do, what they won't be able to do, what happens if something goes wrong, and what the business impact is. Avoid technical jargon. Flag any criterion that might require a business decision before the stakeholder can approve it. This works well alongside the AI briefing document prompts if you need to package the whole thing into a formal sign-off document for leadership. Prompt 10: Run a pre-ship honesty check Use this the day before you ship, not the day after. Here are the final acceptance criteria for [feature] and the current implementation status: [paste both]. Run a pre-ship honesty check. For each criterion: mark it as Verified (evidence exists), Assumed (likely but untested), Open (still unclear), or Risk (known gap). List any open questions that need an answer before launch. List any stakeholder, team, or function that should be looped in based on risk level. Be direct. Do not soften gaps. The output of this prompt is not a green light. It's a risk register. Someone with authority over the launch needs to see it and make a call. What AI can't do here, and who should own it AI will draft criteria. It will not make product decisions. It has no idea what your customers actually need, what your legal team has approved, what your SLAs say, or what "done" means to your executive sponsor. Every prompt above produces a draft. Drafts need owners. The owner is a named human with accountability for the decision. When the work affects real risk, escalate to the right people: Legal: contracts, compliance, terms of service, IP Security: authentication, authorization, data exposure, vulnerabilities Privacy: PII, consent, data retention, GDPR/CCPA obligations Accessibility: WCAG conformance, assistive technology compatibility Finance: cost thresholds, billing logic, financial reporting Customer leadership: promises made to specific accounts, SLA implications QA: test coverage, regression risk, known system fragility Engineering: architecture constraints, technical debt, integration dependencies AI can prompt you to ask the right questions. It can't answer them. That's not a limitation to work around. That's the job staying human. If you want to understand where that line sits more broadly, what AI can and can't do is worth reading before you start delegating anything consequential. For a broader framework on using AI for structured work without outsourcing the judgment, the AI project management prompts collection covers briefs, risk logs, and stakeholder updates with the same approach. Frequently asked questions What are AI acceptance criteria prompts? AI acceptance criteria prompts are structured instructions you give to tools like ChatGPT or Claude to help convert vague requests, notes, or tickets into testable acceptance criteria. They work best as a drafting starting point. A human with product knowledge needs to verify every output before it's treated as final. Can AI write acceptance criteria by itself? AI can draft acceptance criteria quickly, but it will invent business rules, miss context, and present assumptions as facts unless you explicitly tell it not to. The prompts here are designed to surface those gaps rather than hide them. A qualified owner needs to review, correct, and sign off on any AI-generated criteria. What's the difference between AI-drafted criteria and real acceptance criteria? Real acceptance criteria have named owners, verified business rules, confirmed edge cases, and explicit scope boundaries. AI-drafted criteria have structure. The difference is accountability. Use AI to get to structure faster, then do the work of verification before the ticket moves. Should I paste ticket details into ChatGPT or Claude? Only if your organization has approved that tool for the data you're pasting. Never paste customer PII, security vulnerabilities, unreleased product details, legal content, financial forecasts, or confidential strategy into unapproved AI tools. When in doubt, describe the scenario in abstract terms instead of pasting the actual data. What format should acceptance criteria be in? Given/When/Then (Gherkin format) is the most common for testable criteria because it separates precondition, action, and expected outcome clearly. Plain numbered lists work for simpler requirements. The format matters less than the specificity. Vague Given/When/Then is still vague. When should I loop in legal, security, or privacy review? Any time your acceptance criteria touch data collection, user authentication, financial transactions, health information, content moderation, third-party integrations, or anything regulated in your industry or jurisdiction. If you used Prompt 6 and it flagged gaps, those flags are your escalation trigger. Don't ship past them without a sign-off from the right specialist. If you want the full framework for staying useful while the tools get faster, Don't Replace Me is the field guide. It won't write your acceptance criteria. Neither should anything else you haven't reviewed. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Change Impact Analysis Prompts: Spot Damage AI change impact analysis prompts for mapping affected systems, teams, customers, QA checks, support gaps, and rollout risks. AI promptschange impactrisk AI Defect Triage Prompts: Sort Bugs Faster AI defect triage prompts for sorting bug reports, severity, duplicates, owners, workarounds, retests, and customer impact after launch. AI promptsdefect triagebug reports AI QA Checklist Prompts: Catch Mistakes Before They Ship 10 copy-paste AI QA checklist prompts for catching bugs, edge cases, and gaps before launch. Includes safety rules and a reusable prompt formula. AI promptsQAquality assurance --- ## AI Accessibility Testing Prompts: Catch Barriers URL: https://dontreplace.me/articles/ai-accessibility-testing-prompts-catch-barriers-before-users-do AI promptsaccessibility testingquality assuranceinclusive designsoftware testing AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. August 15, 2026 Updated August 15, 2026 Accessibility bugs are often ordinary product bugs wearing an invisibility cloak. A keyboard user cannot reach the checkout button. A screen reader announces six controls as “button.” An error appears in red but is never announced. At 200% zoom, the navigation eats half the page and the close control wanders off-screen. AI accessibility testing prompts can help turn verified requirements, design-system rules, and known risks into structured checks. They can propose scenarios, organize coverage, find unanswered questions, and turn sanitized observations into cleaner defect reports. They cannot experience your product as a disabled user, prove conformance, determine legal compliance, or approve a release. AI can draft a testing map. Humans still have to navigate the product, use real assistive technology, listen to disabled users, and own the result. Start with an AI test plan when the release strategy is unclear. Use AI test case prompts when requirements need traceable checks. Accessibility testing adds another requirement: successful task completion matters more than a suspiciously green checklist. Why accessibility testing needs human evidence Automated scanners are useful. They can catch missing labels, some contrast failures, invalid markup, and other machine-detectable problems. They cannot tell you whether focus order makes sense, whether a label is understandable, whether a workflow is exhausting, or whether a screen-reader announcement arrives at the moment someone needs it. AI has the same boundary. It can reason over the context you provide, but generated text is not test evidence. If the model says “the modal should trap focus,” that is a hypothesis or requirement—not proof that your modal does it. Useful accessibility coverage can include: keyboard access, visible focus, logical focus order, and escape behavior; accessible names, roles, states, values, and screen-reader announcements; headings, landmarks, page titles, link purpose, and navigation consistency; labels, instructions, validation, errors, status messages, and recovery; contrast, non-color cues, forced colors, and meaningful visual states; text resizing, browser zoom, reflow, spacing changes, and responsive layouts; captions, transcripts, audio description, autoplay, flashing, and motion controls; plain language, predictable interaction, time limits, and cognitive load; touch target size, orientation, voice input, switch access, and magnification; compatibility across the supported browser, operating system, and assistive-technology matrix. The right coverage depends on the product and its users. A payroll form, streaming player, internal dashboard, and public emergency service do not have the same risk. The model should structure the known context, not invent a universal accessibility program from vibes. Pair this guide with AI UI testing prompts for broader interface coverage or mobile app testing prompts when platform gestures, native controls, and mobile assistive technologies matter. The reusable accessibility testing prompt formula Use this base prompt, then add one specialized template below: “Act as an accessibility testing planning assistant. I am testing [named user flow] in [product/build] for [user groups] on [approved environment]. Use only the supplied requirements, adopted accessibility standard, supported platform and assistive-technology matrix, component specifications, known risks, and sanitized evidence. Produce [artifact] with setup, actions, expected behavior traced to a source, observable evidence, user impact, cleanup, and unanswered questions. Separate automated checks, manual checks, and disabled-user research. Do not invent execution results, screen-reader output, conformance, legal conclusions, severity, or release decisions.” Traceability is the anti-theater mechanism. If an expected behavior cannot be tied to an adopted requirement, approved design, component contract, platform convention, or confirmed product decision, label it as a question. Do not let the model quietly upgrade “common practice” into “our product passed.” Never paste customer PII, credentials, private recordings, production screenshots, support transcripts, health or disability data, unreleased designs, vulnerability details, regulated information, or confidential client material into an unapproved AI tool. Use synthetic accounts, redacted evidence, approved environments, placeholder secrets, and human privacy and security review. What to collect before prompting “Check accessibility” produces a generic checklist large enough to wallpaper the office. Give the model a narrow packet of verified, safe context instead. InputWhy it mattersHuman check Named flow and user goalDefines successful task completionProduct owner confirms scope Build, page, and environmentAnchors evidence to a versionTester records exact version Adopted requirementsDefines expected behaviorAccessibility owner confirms sources Supported platform matrixBounds compatibility workConfirm browser, OS, and AT versions Component and content specsSupplies intended semanticsDesign and engineering resolve conflicts User states and test dataExposes validation and recovery pathsUse synthetic, approved accounts Known risks and prior findingsPrioritizes consequential checksConfirm findings are current Evidence and severity rulesMakes defects reproducibleHumans own impact judgment Keep unknowns visible. If nobody has decided how a timeout warning should work with a screen reader, the output should contain a decision request—not an invented requirement in a handsome table. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI accessibility testing prompts Replace bracketed text with verified, non-sensitive information. These prompts create planning artifacts. They do not execute tests or prove compliance. 1. Turn requirements into a risk-based test matrix “Using these adopted accessibility requirements, named flows, supported platforms, components, user needs, known defects, and release changes: [paste], create a risk-based accessibility test matrix. Separate automated checks, keyboard checks, screen-reader checks, visual checks, cognitive checks, and research needs. Trace every expected result to a supplied source. Rank by user impact and change risk, and flag missing requirements.” A useful matrix connects requirements to tasks and evidence. It does not merely repeat guideline numbers. Include who tests, with what setup, on which build, and what observable result counts as evidence. Prioritize critical tasks such as authentication, purchase, consent, account recovery, document submission, and core navigation. A decorative icon with a weak label and an unreachable payment control are not equivalent just because both appear in the same audit report. 2. Test keyboard navigation and focus behavior “For this flow, interactive component list, overlays, routes, shortcuts, and focus-management rules: [paste], draft manual keyboard checks. Cover forward and reverse tab order, visible focus, activation, skip links, menus, dialogs, popovers, drag alternatives, escape behavior, route changes, focus restoration, and keyboard traps. Include expected behavior, observed-evidence fields, and stop conditions.” Run the flow without a mouse or touch input. Confirm every required action is available, focus is visible, and the sequence follows the task rather than the DOM’s archaeological layers. For dialogs and temporary surfaces, check initial focus, containment where required, escape or close behavior, and return focus. Do not mark the test passed because pressing Tab eventually finds something clickable three browser screens away. 3. Plan screen-reader names, roles, states, and announcements “Using this semantic component specification, content, state model, and supported screen-reader matrix: [paste], create manual checks for accessible names, roles, values, expanded or selected states, reading order, dynamic updates, loading, errors, confirmations, and route changes. Provide expected meaning, not fabricated spoken output. Flag controls whose purpose depends only on visual context.” Test with the actual supported screen reader and browser or operating-system pairing. Record versions and interaction mode. Different combinations can expose different behavior, and one successful announcement does not prove universal compatibility. Names should be concise and useful in context. Ten “Learn more” links or twelve unlabeled icon buttons make a page technically interactive and practically ridiculous. Check that state changes and async updates are announced without hijacking the user’s current position. 4. Check headings, landmarks, links, and page structure “Given this page outline, navigation model, content hierarchy, and reusable layout: [paste], draft checks for page title, language, headings, landmarks, bypass links, lists, tables, link purpose, current-page indication, and repeated navigation. Compare visual grouping with semantic structure and identify questions where the source is ambiguous.” Screen-reader users often navigate by headings and landmarks rather than reading every node. A visually polished page with random heading levels can feel like a book whose table of contents was assembled by a raccoon. Inspect the accessibility tree and then navigate with the assistive technology. Confirm headings describe sections, landmarks are distinct, links make sense in nearby context, and data tables expose meaningful headers. Visual size alone does not create structure. 5. Test forms, instructions, errors, and recovery “For this form, field inventory, validation rules, required states, help text, submission outcomes, and recovery paths: [paste], create accessibility tests for labels, instructions, autocomplete, required indicators, grouping, input formats, inline validation, error summaries, focus movement, announcements, preserved data, and successful correction. Cite the source for each expected result.” Test empty submission, malformed values, server errors, expired sessions, and multiple simultaneous errors. Confirm the user can find the problem, understand it, correct it, and submit without re-entering unrelated data. Color and placeholder text are not labels. A red border is not an explanation. An error summary that moves focus but does not link to fields is only halfway helpful. Human testers should judge whether recovery is understandable, not merely whether an error node exists. 6. Check contrast and non-color communication “Using these approved colors, component states, themes, backgrounds, charts, icons, and adopted contrast requirements: [paste], create checks for text, large text, controls, focus indicators, disabled and selected states, links, errors, charts, hover, high-contrast or forced-colors mode, and images of text. Separate measurable contrast from human interpretation and flag missing tokens.” Use approved measurement tools on actual rendered colors. Design-token values may blend with overlays, opacity, antialiasing, images, or state changes. Record the foreground, background, component state, location, and tool result. Then check meaning. A chart can meet contrast thresholds and still require color vision to distinguish series. An error can have an icon and still be incomprehensible. Non-color cues need to survive the real interaction, not just the static mockup. 7. Test zoom, text resizing, reflow, and spacing “For this responsive flow, supported viewport range, text components, overlays, sticky regions, and zoom or reflow requirements: [paste], draft checks for browser zoom, text-only resizing, narrow viewport reflow, user text-spacing overrides, orientation, magnification, clipping, overlap, horizontal scrolling, hidden controls, and preserved task completion. Include exact setup and evidence fields.” Large text and zoom expose fragile assumptions quickly. Check navigation, dialogs, tables, sticky headers, tooltips, toasts, date pickers, and multi-step forms. Content should not vanish under fixed panels or require pixel-perfect scrolling to find the submit button. Do not shrink the browser and call it equivalent to zoom. Use the specified method and record it. Test the whole task, including errors and confirmation—not only the photogenic default state. 8. Review motion, timing, audio, and video alternatives “Using these animations, autoplay behavior, timed interactions, flashing content, audio, video, and media requirements: [paste], create checks for pause and stop controls, reduced-motion preference, keyboard access, captions, transcripts, audio description needs, volume control, flashing thresholds, time-limit warnings, extensions, and interruption recovery. Flag decisions requiring specialist or legal review.” Confirm reduced-motion settings change meaningful animation without hiding information. Check that carousels, moving banners, and auto-updating regions can be paused where required. A tiny pause button that disappears on keyboard focus is performance art, not access. For media, verify captions and transcripts against the actual content. AI-generated captions can be a starting draft, but names, technical terms, speaker changes, timing, and consequential statements require human review. Never claim a media alternative is accurate because the model produced it confidently. 9. Check cognitive accessibility and plain language “For this task, audience, content, decision points, memory demands, timeout rules, authentication steps, and support options: [paste], draft a cognitive accessibility review. Check clarity, consistency, chunking, predictable controls, instructions at the point of need, error prevention, confirmation, progress, interruption recovery, and avoidance of unnecessary memory or calculation. Distinguish content suggestions from verified user evidence.” AI can identify long sentences, unexplained jargon, inconsistent labels, and complicated branching. It cannot decide whether the experience is understandable to every user. Test with representative people, especially when the task is high stakes or the audience has known access needs. Look for needless memory tests: codes shown on one screen and requested on another, instructions that disappear during entry, arbitrary timeouts, and forms that erase work. Plain language is not childish language. It is accurate language that does not charge users a comprehension tax. 10. Turn observed barriers into actionable bug reports “Using only these sanitized observations from a completed accessibility test: [paste], draft defect reports with build, environment, platform, assistive technology and version, preconditions, exact steps, observed behavior, verified expected source, user impact, evidence reference, reproducibility, and open questions. Separate observation from interpretation. Do not invent severity, compliance claims, spoken output, or remediation.” A strong report describes the barrier and the blocked task, not only the markup symptom. “Button has no accessible name” is useful. “Screen-reader user encounters three unlabeled buttons and cannot identify which one submits payment” explains the consequence. Humans should confirm reproduction, severity, ownership, and the fix. Suggested remediation may affect component architecture, content, design, security, or workflow. Treat generated code as a candidate for review, then rerun the original task with real assistive technology. For general defect structure, use these AI bug report prompts. For release-level coverage, add an AI QA checklist without pretending the checklist replaces testing. A practical workflow that avoids compliance theater Name the task. “Accessibility test the site” is not a task. “Create an account, recover a password, and confirm the dashboard with keyboard and supported screen reader” is. Confirm the sources. Gather adopted requirements, component contracts, supported platforms, and current designs. Remove sensitive material. Use synthetic users and redact evidence before prompting. Generate a draft matrix. Require traceability, unknowns, and separate automated from manual checks. Review before execution. Accessibility, design, engineering, QA, content, privacy, and security owners correct the plan. Execute on the named build. Record actual tools, versions, steps, observations, and evidence. Report barriers by impact. Explain the blocked or degraded user task without inventing legal conclusions. Fix and retest. Verify the original path and nearby component instances. A patched selector is not the same as a usable experience. Include disabled users. Expert testing and user research answer different questions; mature teams need both. This workflow matches the broader rule in the no-BS guide to using AI at work: give AI bounded work, verify the output, and keep accountability attached to a human. Common ways this goes wrong Treating generated checks as evidence A test case is a plan. A model’s prediction is not an observation. Only executed checks on the identified build produce evidence. Claiming conformance from an automated scan Automation catches a valuable subset of issues. A zero-error report does not prove keyboard usability, useful semantics, understandable content, or successful task completion. Testing one assistive-technology combination Support policies should define the matrix. Record versions and results. Do not generalize one browser and screen-reader pairing to every platform. Inventing severity from a guideline reference Severity depends on blocked tasks, affected users, frequency, workarounds, risk, and product context. Let the model organize evidence; let accountable humans judge impact. Pasting sensitive user evidence into public tools Accessibility research can include disability information, voices, faces, account details, and deeply personal context. Follow consent, retention, access, privacy, security, and vendor rules. “We removed the email address” may not be enough to de-identify a story. Replacing disabled-user research with personas Generated personas cannot report fatigue, confusion, trust, dignity, or coping strategies. Use them, at most, to brainstorm questions—not as fake participants. If your team is still fuzzy on the boundary, read what AI can and cannot do. Fast pattern generation is useful. It is not lived experience or accountability. Frequently asked questions Can AI perform accessibility testing automatically? It can support automated tools, organize findings, suggest checks, and draft scripts. Unless it is explicitly connected to approved tools, it has not inspected your product at all. Even connected automation covers only part of accessibility. Manual task testing and disabled-user research remain necessary. Can these prompts prove WCAG conformance? No. They can help map adopted requirements to planned checks and organize verified evidence. Conformance evaluation requires appropriate expertise, scope, methods, representative pages or flows, manual testing, and accountable review. Legal obligations vary, so involve qualified accessibility and legal specialists where needed. Which screen reader should we test? Use the combinations in your documented support policy, informed by product platforms and users. Record the screen reader, browser or operating system, versions, settings, and interaction mode. Do not choose one pairing because it happens to be installed on a developer’s laptop and declare victory. Should AI write accessibility fixes? It can draft candidate markup, tests, labels, or remediation notes. Humans must review semantics, component behavior, content meaning, compatibility, security, and regressions. Retest the complete user task with the relevant assistive technology after the change. How should we prioritize accessibility defects? Consider whether the barrier blocks or degrades a critical task, who is affected, how often it occurs, whether a reasonable workaround exists, and the product’s risk. Guideline mapping helps trace requirements but should not replace user-impact judgment. Is an automated accessibility score useful? As a trend signal, sometimes. As proof of usability or conformance, no. Scores can hide severe manual issues and can encourage teams to optimize what the scanner sees instead of what users need. Keep the underlying findings and task evidence visible. Can AI replace accessibility specialists or disabled-user research? No. AI is useful for formatting, variation, traceability, and gap spotting. Specialists contribute testing judgment and remediation expertise. Disabled users contribute lived experience and reveal barriers, strategies, and priorities that synthetic personas cannot supply. What information is safe to paste into an AI tool? Use only data approved by your organization and vendor policy. Prefer synthetic accounts, placeholder content, redacted screenshots, sanitized markup, and generalized observations. Exclude credentials, customer PII, disability or health data, private recordings, unreleased product details, vulnerabilities, and regulated or confidential information unless an explicitly approved environment and process permits it. The point is better testing, not fewer humans AI accessibility testing prompts are valuable because accessibility work contains plenty of structure: matrices, requirement mapping, state combinations, evidence fields, and repeatable report formats. Let the machine accelerate that structure. Keep the human work human: operating real interfaces, using assistive technology, listening to disabled people, interpreting impact, deciding remediation, protecting sensitive evidence, and owning the release. That is the practical thesis behind Don’t Replace Me by Dmitry Kargaev: use AI for speed without confusing speed for judgment. The goal is not a prettier compliance document. It is a product more people can actually use. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI API Testing Prompts: Catch Broken Integrations AI API testing prompts for contracts, edge cases, authorization, retries, error responses, and human-verified integration evidence. AI promptsAPI testingquality assurance AI Compatibility Testing Prompts: Catch Breakage 10 AI compatibility testing prompts for browser, device, OS, responsive, locale, network, defect triage, and evidence-based release decisions. AI promptscompatibility testingcross-browser testing AI Exploratory Testing Prompts: Find Hidden Bugs AI exploratory testing prompts for focused charters, unusual user journeys, sharper observations, reproducible bugs, and human-owned release decisions. AI promptsexploratory testingquality assurance --- ## AI Account Planning Prompts: Build a Better Plan URL: https://dontreplace.me/articles/ai-account-planning-prompts-build-a-better-plan AI promptsaccount planningsales strategycustomer successsales enablement AI Account Planning Prompts: Build a Better Plan 10 AI account planning prompts to organize buyer evidence, map stakeholders, expose risks, test strategy, and plan next actions without inventing intel. September 04, 2026 Updated September 04, 2026 An account plan should describe what you know, what you do not know, and what your team will do next. It should not be a fan-fiction dossier in which every stakeholder loves you, the budget is definitely coming, and the renewal closes itself. AI account planning prompts can organize verified CRM records, meeting notes, product usage signals, goals, risks, and approved commercial data. They can reveal gaps and help your team prepare better questions. They cannot read buyer politics, validate a relationship, predict a purchase, or transform an optimistic seller note into evidence. AI can structure account evidence and draft hypotheses. Humans still own discovery, relationships, commercial judgment, privacy, approvals, and the account. These ten templates help account executives, customer success leaders, founders, and consultants build useful plans without appointing autocomplete as vice president of imaginary revenue. What an account plan should contain A practical account plan connects reliable evidence to deliberate action. It helps a team understand the customer, coordinate relationships, protect current value, identify legitimate opportunities, and decide what to learn next. Keep these categories separate: Verified fact: supported by a dated, approved source. Customer statement: what a named role actually said, without creative interpretation. Seller interpretation: your reading of the situation, still subject to confirmation. Hypothesis: a testable possibility, not a CRM fact wearing a tie. Unknown: information the team does not currently have. Goal: an outcome the customer has confirmed matters. Signal: observable behavior that may deserve investigation. Risk: a condition that could damage adoption, value, renewal, or trust. Opportunity: a qualified problem your organization may be able to solve. Next action: a specific step with an owner and date. If those labels disappear, account planning becomes optimism with formatting. “The VP is an executive sponsor” may mean the VP attended one call nine months ago. “Expansion opportunity” may mean someone clicked a webinar link. A language model will happily smooth those distinctions unless you force it to preserve them. Start with clean evidence. These AI sales prompts for discovery and follow-up can help prepare questions, while sales pipeline review prompts can expose deals that are moving mainly because the stage field says so. The evidence-first account planning formula Add this instruction to any prompt below: “Act as an account planning assistant. Use only the sanitized, approved records I provide for [account type, planning horizon, business goal, and artifact]. Separate verified facts, customer statements, seller interpretations, hypotheses, signals, risks, unknowns, and proposed actions. Cite a supplied source label and date for every material factual claim. Mark unsupported items [VERIFY]. Do not invent stakeholders, reporting lines, influence, sentiment, priorities, budgets, competitors, product usage, intent, renewal likelihood, legal terms, or commitments. Do not infer sensitive traits or private information. Preserve conflicts and missing data instead of resolving them silently.” That final sentence matters. Models complete patterns. Give one executive name and a vague meeting note, and the machine may produce a tidy political map that nobody actually verified. Tidy fiction is still fiction. Never paste customer names, personal contact details, contracts, credentials, private CRM commentary, pricing exceptions, health or financial information, confidential strategy, regulated data, or legally sensitive material into an unapproved AI tool. Use approved systems, minimum necessary data, de-identification, access controls, retention limits, and human privacy, legal, security, and commercial review. What to collect before prompting Create a sanitized planning pack. Keep authoritative records in approved systems, not in a wandering chat transcript. InputWhy it mattersHuman owner Account goals and planning horizonDefines what the plan is forAccount owner Dated customer statementsGrounds priorities in evidenceRelationship owner Stakeholder interactionsShows actual contact, not assumed influenceAccount team Current products and usageEstablishes the present relationshipCustomer success or product Outcomes and success measuresConnects work to customer valueCustomer and success lead Open opportunitiesSeparates qualified work from wishful whitespaceSales lead Renewal and contract milestonesAnchors timingCommercial owner Risks, blockers, and support issuesPrevents surprise optimismSuccess and support leads Procurement, security, and legal constraintsRoutes specialist reviewRelevant specialists Competitor information with sourcesAvoids rumor becoming strategyAccount owner Source labels and datesMakes every claim traceablePlan editor Unknowns and next questionsTurns gaps into discoveryNamed action owner Stale input creates stale strategy. A stakeholder map from last year is not “mostly fine” after a reorganization. A product usage export without context cannot tell you whether activity represents value, testing, seasonality, or one enthusiastic intern. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI account planning prompts Replace brackets with sanitized, approved information. Every output is a draft for accountable human review. 1. Clean and label the account evidence “Review these sanitized records with source labels and dates: [paste]. Create sections for verified facts, customer statements, seller interpretations, hypotheses, signals, risks, conflicts, stale items, and unknowns. Cite the source for every item. Mark anything without adequate support [VERIFY]. Do not infer motives, sentiment, authority, budget, urgency, or purchase intent.” This prompt creates the evidence layer for everything else. It is deliberately boring. Boring is excellent when the alternative is presenting a hallucinated org chart to your sales director. Have each record owner check the result against the source. Preserve disagreement. If one contact says adoption is strong and product data shows a decline, the plan should show that conflict and assign investigation. It should not average them into “adoption is moderate.” Also set a freshness rule. A six-month-old objective may be historical context rather than a current goal. Label dates clearly enough that reviewers can judge whether evidence still deserves weight. 2. Build a sourced account snapshot “Using only this verified account evidence: [paste], draft a one-page account snapshot. Include current relationship, confirmed goals, products or services in use, documented outcomes, active work, key milestones, known constraints, material risks, open questions, and the next review date. Add a source label beside every factual claim. Put hypotheses in a separate section. Do not create a health score or forecast unless I provide an approved method and inputs.” A useful snapshot helps a new teammate understand the account without reading every meeting note since civilization began. It should compress evidence, not manufacture certainty. Keep the snapshot short enough to review before a call. If it grows into a seventeen-page history of every email, ask the model to prioritize information that changes a decision during the stated planning horizon. Humans should approve that priority. Avoid decorative scores unless your organization has a real scoring model. A machine-generated “82% healthy” number is not analysis. It is a progress bar attached to vibes. 3. Map stakeholders without inventing influence “Create a stakeholder evidence table from these dated interactions: [paste]. Use columns for anonymized role, confirmed responsibilities, stated goals, documented concerns, interaction date, relationship owner, observed participation, influence evidence, unknowns, and next respectful action. Use ‘unknown’ when authority, sentiment, or reporting lines are not verified. Do not infer influence from title alone or infer personal traits.” Stakeholder maps often rot into organizational astrology. A senior title does not prove decision authority. Frequent replies do not prove internal influence. Silence does not prove opposition. Use neutral language. “Raised a security concern in the May review” is evidence. “The security lead is blocking the deal” is an interpretation unless the customer confirmed it. That difference affects how your team behaves and whether it damages trust. Relationship owners should validate entries directly through normal account work. Do not use AI to profile private lives, infer protected characteristics, or engineer manipulative pressure. The goal is coordinated service, not surveillance with colorful boxes. 4. Separate facts from strategic hypotheses “Review this proposed account strategy and supporting evidence: [paste]. Produce a three-column table: claim, classification, and test. Classify each claim as verified fact, customer statement, seller interpretation, hypothesis, or unknown. For each hypothesis, propose a respectful discovery question or observable evidence that could confirm or reject it. Do not rewrite hypotheses as facts.” Strong account teams use hypotheses. Weak plans hide them. For example, “The customer may need better reporting before renewal” can be a useful hypothesis if support tickets and meeting notes point that way. The next move is to ask and observe, not to generate a proposal for an analytics add-on before anyone confirms the problem. Give each hypothesis an owner and expiration date. If nobody tests it, it should not linger in the CRM until repetition turns it into folklore. 5. Connect customer goals to verified success measures “Using these customer-confirmed goals, current activities, and available measurements: [paste], create a goal-to-evidence map. For each goal, show the customer’s wording, source and date, current baseline if supplied, approved success measure, evidence available, evidence missing, risks, and next validation step. Do not invent baselines, targets, attribution, ROI, or guarantees.” Goals such as “improve efficiency” or “drive adoption” are too vague to steer an account. AI can help expose the missing parts, but only the customer and accountable team can define what success actually means. Be especially careful with attribution. Increased revenue after implementation does not prove your product caused it. Reduced tickets may reflect seasonality, staffing, or a process change. State what the evidence shows and what it cannot establish. If success measures are absent, generate questions for the next review. Do not let the model fill the blank with a benchmark from somewhere else and quietly pretend it belongs to this customer. 6. Review current value and legitimate whitespace “Review these verified products, usage records, outcomes, customer goals, unresolved problems, and approved capabilities: [paste]. Create two separate sections: current value evidence and possible whitespace hypotheses. For each whitespace hypothesis, show the supporting evidence, missing qualification, customer benefit to validate, potential conflict, and next discovery question. Do not call a possibility an opportunity, recommend an unapproved product, or invent demand.” Whitespace is where disciplined planning goes to die. A blank square in a product matrix does not mean the customer should buy the thing. It only means they have not bought it. Begin with value already promised. If current adoption, support, or outcomes are weak, expansion pressure can feel absurd to the customer. These customer retention prompts help inspect signals without pretending the model knows why someone might leave. A qualified opportunity requires a real problem, relevance, stakeholder engagement, a plausible process, and human commercial judgment. The model can prepare the questions. It cannot qualify the buyer by itself. 7. Build a risk and blocker review “Analyze these sanitized account records: [paste]. Create a risk register with evidence, source date, affected customer goal, possible impact, likelihood method if supplied, current control, owner, trigger, next action, and escalation path. Separate observed risks from hypothetical scenarios. Flag stale or contradictory evidence. Do not assign probability, severity, blame, or legal conclusions without an approved method.” Risks include more than renewal anxiety. Look for adoption gaps, unresolved support issues, stakeholder turnover, missing executive alignment, unclear success measures, security reviews, procurement timing, product dependencies, implementation capacity, and promises the team may not be able to keep. Use an approved scoring method if one exists. Otherwise, keep likelihood and impact qualitative and explain the evidence. Precision without a method is costume jewelry for spreadsheets. For broader consequence checks, use these AI risk assessment prompts. Specialists still own security, legal, financial, accessibility, and privacy decisions. 8. Draft a relationship action plan “Using this verified stakeholder map, customer goals, open questions, milestones, and risks: [paste], draft a 30/60/90-day relationship action plan. For each action include purpose, customer value, internal owner, appropriate customer role, evidence behind the action, preparation needed, due date, dependency, and completion signal. Mark unverified assumptions. Avoid manipulative tactics, manufactured urgency, excessive contact, or actions based on inferred private information.” The plan should answer why each interaction helps the customer, not merely why your forecast wants another meeting. Good actions include validating a success measure, resolving an adoption obstacle, confirming a decision process, preparing a useful review, or connecting the right specialists. “Touch base” is not an action unless touching bases somehow changes the evidence. Balance activity with respect. More messages do not automatically create a stronger relationship. Let the human owner judge channel, timing, cultural context, and whether the customer has actually invited the conversation. 9. Stress-test the account strategy “Act as a skeptical account review panel. Using only this evidence pack and draft strategy: [paste], identify unsupported claims, stale sources, circular reasoning, missing stakeholders, untested assumptions, weak customer value, hidden dependencies, conflicting dates, privacy concerns, forecast leakage, and actions without owners. Quote the relevant claim and source. Ask up to 12 review questions. Do not invent counterfacts or declare the strategy correct.” This is where AI is genuinely handy: relentless comparison without social discomfort. It can notice that your strategy depends on a champion whose last documented interaction was eleven months ago. Run the stress test before a formal account review, then have teammates challenge it. A model may miss organizational nuance, misunderstand a product signal, or flag a harmless conflict. Treat findings as review prompts, not verdicts. Compare commercial assumptions with the approved forecast process. These sales forecasting prompts can help document evidence without pretending uncertainty has left the building. 10. Prepare the human review and next-action log “Turn this draft account plan and reviewer comments into a human review checklist. Include factual verification, source freshness, customer goal confirmation, stakeholder accuracy, product and usage accuracy, opportunity qualification, risk ownership, privacy and legal review, commercial approval, accessibility, unresolved conflicts, next actions, owners, and due dates. Add a decision log showing approved, rejected, deferred, and [VERIFY] items. Do not approve anything.” The final prompt is a handoff, not a robotic blessing. Name the people who must verify their domains. The account owner cannot silently approve security conclusions; the model cannot approve anything at all. Store the reviewed plan where access, revision history, and retention are controlled. Record meaningful changes. If a hypothesis is rejected, keep enough history to prevent it from reappearing next quarter as a “new insight.” Use stakeholder update prompts to communicate approved actions internally without dumping private customer context into a giant distribution list. A simple account planning workflow Use the prompts as a sequence rather than throwing the entire CRM into one giant context window: Define the decision. State the planning horizon and what this plan needs to help the team decide. Minimize the data. Export only approved, relevant, sanitized records. Label every source. Include owner, date, and record type. Clean the evidence. Separate facts, interpretations, hypotheses, and unknowns. Build focused artifacts. Create the snapshot, stakeholder map, goal map, and risk review separately. Validate with owners. Ask relationship, product, success, support, finance, legal, privacy, and security owners where relevant. Test hypotheses. Turn uncertainty into respectful discovery, not confident prose. Stress-test the strategy. Look for stale evidence and missing dependencies. Approve actions. Assign human owners, dates, and completion signals. Review on a cadence. Update evidence and archive stale assumptions. This is the same basic principle behind using AI at work responsibly: automate structure and comparison, not accountability. The no-BS guide to using AI at work is a useful reset when your process starts treating generated text as truth. Common ways AI account plans go wrong The CRM dump A team exports everything and asks for “insights.” The result is broad, confident, and difficult to audit. Start with a decision and a bounded evidence pack instead. The fictional org chart The model fills missing reporting lines and influence based on titles. Require unknown labels and sourced relationship evidence. The whitespace fantasy Every unpurchased product becomes an expansion recommendation. Separate possible relevance from a qualified customer problem. The sentiment horoscope Polite language is classified as enthusiasm; slow replies become resistance. Do not infer emotion or intent from thin communication signals. The zombie objective An old customer goal survives every plan because nobody checks its date. Add freshness labels and reconfirm goals. The decimal-point health score The plan reports a precise score without an approved model. Use transparent evidence and a real scoring method, or skip the number. The privacy bonfire Private CRM notes, contracts, or personal information are pasted into an unapproved tool. Minimize, sanitize, control access, and follow policy before prompting. The plan nobody owns The document contains twenty excellent suggestions and zero accountable owners. Every action needs a human, due date, and completion signal. Understanding what AI can and cannot do helps prevent most of these errors. Fast synthesis is valuable. Hidden certainty is not. Frequently asked questions Can ChatGPT create an account plan from CRM data? It can draft a plan from a carefully selected, sanitized, approved export. Do not connect or upload customer records unless your organization has approved the tool, data handling, access, retention, and purpose. Verify every material claim against the authoritative source. What are the best AI account planning prompts? The best prompts define the decision, restrict the model to supplied sources, separate facts from hypotheses, preserve unknowns, prohibit sensitive inference, and request a specific artifact. A shorter prompt with strong evidence controls beats a theatrical mega-prompt built on stale data. Can AI identify decision-makers in an account? AI can organize documented roles, statements, and interactions. It cannot reliably determine real influence, authority, or internal politics from titles and meeting attendance. Treat those as unknown until humans verify them through respectful discovery. Can AI find expansion opportunities? It can compare verified customer goals and problems with approved capabilities to propose hypotheses. That is not qualification. A human team must confirm the problem, value, stakeholders, timing, fit, and commercial process before calling it an opportunity. How often should an account plan be updated? Update it when material evidence changes and on a cadence appropriate to the account. Stakeholder turnover, major support issues, goal changes, product adoption shifts, procurement events, and renewal milestones are obvious triggers. Date every source so stale assumptions are visible. What customer data should never go into a public AI tool? Do not paste personal contact details, credentials, contracts, private CRM notes, pricing exceptions, confidential strategy, payment information, regulated records, or legally sensitive material into an unapproved tool. Follow your organization’s policy and use minimum necessary, de-identified information even in approved systems. Should AI calculate account health or renewal probability? Only if your organization has an approved, documented method, valid inputs, monitoring, and accountable human review. A language model should not invent probability from prose. Explain the evidence, limitations, and uncertainty behind any score. Does AI replace account managers or sales judgment? No. It speeds up organization, comparison, drafting, and gap detection. Relationship context, trust, discovery, judgment, negotiation, privacy, commercial approval, and accountability remain human work. That is the whole point: use the machine for speed, then spend human attention where consequences live. The useful boundary A good account plan is not the longest document or the prettiest dashboard. It is a shared, evidence-based view of the customer that leads to better questions and responsible action. Use AI to clean notes, expose missing sources, structure maps, compare a strategy with its evidence, and prepare reviews. Do not use it to invent buyer intelligence, assign motives, profile people, qualify imaginary opportunities, or approve commercial decisions. The durable advantage is not generating plans faster than everyone else. It is knowing which claims deserve trust, which questions deserve a conversation, and which promises your team can responsibly make. That human judgment is the moat. If you want the broader field guide for keeping that boundary intact, Don’t Replace Me by Dmitry Kargaev carries the same principle through the rest of modern work: let AI handle speed, and keep truth, taste, trust, and accountability attached to humans. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Churn Analysis Prompts: Find Why Customers Leave 10 AI churn analysis prompts to investigate why customers leave, test assumptions, protect customer data, and turn evidence into accountable action. AI promptschurn analysiscustomer retention AI Customer Renewal Prompts: Prepare With Evidence 10 AI customer renewal prompts to verify outcomes, expose account risks, prepare options, and protect trust without inventing value or customer sentiment. AI promptscustomer renewalcustomer success AI Customer Retention Prompts: Keep Customers 10 AI customer retention prompts to diagnose churn signals, plan useful follow-up, protect customer data, and keep human judgment in charge. AI promptscustomer retentioncustomer success --- ## AI Action Item Prompts: 10 Templates That Work URL: https://dontreplace.me/articles/ai-action-item-prompts-turn-meetings-into-work ai action itemsmeeting promptsai at workchatgpt promptsproductivity AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. July 14, 2026 Updated July 14, 2026 The meeting ended 40 minutes ago. You have three pages of notes that look like a crime scene. Everyone nodded at the end. Nobody wrote anything down. By tomorrow, two people will remember different decisions, one person will claim they never agreed to anything, and the action items will live and die in someone's notebook. AI action item prompts can help with this. Not by replacing your judgment, but by doing the extraction work you'd otherwise skip because you're already late for the next meeting. Here's exactly how to use them, what to paste in, and where to stop before AI starts inventing things that never happened. Why meetings produce noise instead of work Most meetings generate three things: decisions, suggestions, and vibes. The problem is that all three sound identical in the notes. "We should probably look into the vendor pricing" is not a task. It's a thought someone had out loud. AI can't tell the difference unless you tell it to look. The other issue is that clean formatting isn't the same as real agreement. A neat numbered list with names and dates looks like accountability. It isn't, unless those names and dates came from what was actually said and were confirmed by the actual people. This is Rule #5 from Don't Replace Me: It's Not Smart. It's Fast. AI will give you a beautiful action list in ten seconds. Whether the tasks are real, whether the owners agreed, whether the deadlines are possible: that's still on you. One more thing worth naming. Most meetings have a graveyard of action items from previous meetings that nobody followed up on. The new list doesn't fix that. If you're going to use these prompts, pair them with a habit of actually checking last week's list before the next meeting starts. AI can help you extract. It can't force closure. What to paste in, and what to leave out Before we get to the prompts, the input matters as much as the template. Paste in your rough notes, a transcript, or a bullet summary. Be specific about what was said. What you should never paste into an unapproved AI tool: Customer PII or private client conversations Employee records, HR issues, or performance data Financial forecasts, legal disputes, or unreleased product details Security vulnerabilities, credentials, or access information Medical, safety, or sensitive personal information Confidential strategy or acquisition planning If your meeting touched any of that, get internal approval for the tool you're using before anything goes in. When real risk is involved, escalate to legal, security, privacy, HR, finance, or whoever owns that area. AI action item prompts are useful. They're not a compliance function. Also worth noting: the quality of AI output here is directly proportional to the quality of your notes. If your notes say "discussed the thing with marketing" you'll get useless output. If they say "Sarah and Tom disagreed on timeline for the campaign launch, Tom said he needs two more weeks, Sarah pushed back, no resolution" you'll get something you can actually work with. The prompts below are designed to get the most out of messy notes, but they can't perform miracles on three words and a doodle. The reusable AI action item prompt formula Every prompt in this article follows the same structure. You can use it as a template for anything not covered below. You are an action item assistant. I'll give you [notes / transcript / summary]. Your job is to [specific task]. Rules: - Only use what's in the source material. Do not invent owners, deadlines, or decisions. - Flag anything unclear or missing as an open question. - Mark confidence level: HIGH (explicitly stated), MEDIUM (implied), LOW (unclear). - Do not guess. Ask. [Paste notes here] Keep that formula. It's the difference between useful output and confident fiction. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → AI action item prompts: 10 copy-paste templates Prompt 1: Extract action items from rough notes Use this when your notes are a mess and you need a starting point. You are an action item assistant. I'll give you rough meeting notes. Extract every action item mentioned. For each one, note: - The task (what needs to happen) - Who was mentioned as responsible (or flag as "owner unassigned") - Any deadline mentioned (or flag as "no deadline stated") - Confidence level: HIGH (explicitly agreed), MEDIUM (implied), LOW (unclear) Do not invent owners or deadlines. Flag gaps as open questions. [Paste notes here] Prompt 2: Turn a transcript into owners and due dates Use this after an auto-transcribed meeting. Expect gaps. That's fine. You are an action item assistant. I'll give you a meeting transcript. For each action item you find: - State the task clearly - Name the owner (use exact name from transcript, or flag "owner not named") - State the deadline (exact date if mentioned, or "no deadline given") - Quote the relevant line from the transcript as your source Do not summarize. Do not infer. Quote and flag. [Paste transcript here] Prompt 3: Separate decisions from suggestions This one is genuinely useful and underused. Most action item lists mix the two, and then people argue about it for three weeks. You are an action item assistant. I'll give you meeting notes. Sort the content into two lists: DECISIONS: things that were agreed, confirmed, or signed off SUGGESTIONS: things that were raised, proposed, or "we should probably..." For each decision, note who made it or confirmed it. For each suggestion, note who raised it and whether next steps were agreed. Flag anything where you can't tell which category it belongs to. [Paste notes here] Prompt 4: Find tasks with no owner Nothing dies faster than an action item with no name attached. You are an action item assistant. I'll give you a meeting action list. Identify every task that has no clearly named owner. Group them into: - Tasks where the owner can probably be inferred (explain your reasoning) - Tasks where the owner is genuinely unclear (flag for follow-up) Do not assign owners. Flag and explain. I will assign them. [Paste action list here] Prompt 5: Identify dependencies and blockers For anything with moving parts, this saves you a week of surprises. You are an action item assistant. I'll give you a meeting action list. For each task, identify: - Whether it depends on another task being completed first - Whether it requires input, approval, or resources from someone else - Any blockers mentioned or implied in the notes Flag dependencies as: SEQUENTIAL (can't start until X is done) or PARALLEL (can run at the same time). Flag anything uncertain as an open question. [Paste action list here] For more on managing outputs like this into a real plan, the AI project management prompts article has templates for turning action lists into proper briefs. Prompt 6: Draft a follow-up email This is the prompt most people want first. Use it last. Get the action list right, then draft the email around it. You are a meeting follow-up assistant. I'll give you a verified action list from a recent meeting. Draft a follow-up email that: - Opens with a one-sentence summary of what was decided - Lists each action item with owner, deadline, and any noted dependency - Ends with a clear next meeting date or check-in point if mentioned - Uses plain, professional language (no jargon, no filler) Tone: clear and direct, not corporate. [Paste verified action list here] See also the AI meeting follow-up prompts article for more templates in this area. Prompt 7: Create a client-safe action list Not everything in your internal notes belongs in a client email. This prompt helps you filter. You are an action item assistant. I'll give you an internal meeting action list. Create a client-safe version that: - Includes only tasks relevant to the client - Removes any internal comments, budget discussions, or team-only notes - Uses clear language a client would understand without internal context - Flags anything you're unsure whether to include Do not edit the substance of what was agreed. Only filter and clarify. [Paste internal action list here] If the meeting involved sensitive client information, get sign-off before using any AI tool. When client trust is on the line, a human review isn't optional. Prompt 8: Write an internal team task recap Different from the client version. This one is for your team and should include context, dependencies, and risk flags. You are a team update assistant. I'll give you a meeting action list. Write an internal recap that includes: - What was decided and by whom - Each task with owner, deadline, and dependency notes - Any open questions that still need resolution - Any risks or blockers flagged during the meeting - A brief note on what happens if any task slips Keep it short. Teams don't read long updates. [Paste action list here] This pairs well with AI status report prompts when you need to push the recap up to stakeholders. Prompt 9: Check for unrealistic commitments Someone in your meeting said "I'll have that done by Friday." It's Wednesday. The thing takes two weeks. This prompt helps you catch it before it becomes an incident. You are a task review assistant. I'll give you an action list with owners and deadlines. Review each task and flag: - Deadlines that seem tight given the scope of the task - Tasks where one person owns multiple heavy deliverables in the same window - Dependencies that make the deadline impossible (X can't finish until Y is done) - Anything where the task description is too vague to estimate effort Do not rewrite the tasks. Only flag concerns and explain your reasoning. [Paste action list with owners and deadlines here] Prompt 10: Audit whether the action list is safe to send Use this before anything goes out. External, internal, doesn't matter. You are an action list reviewer. I'll give you a draft action list I'm about to send. Check for: - Tasks with no named owner - Deadlines with no date (just "soon" or "ASAP") - Decisions that appear in the list but weren't explicitly agreed in the notes - Anything that could be misread as a commitment the team didn't make - Sensitive content that shouldn't be in a written record without review (flag the category, don't repeat it) - Open questions that haven't been addressed Return a clean pass/fail and a list of issues to fix before sending. [Paste action list here] What AI gets wrong about action items, and when it matters AI is good at extraction. It's bad at politics. It doesn't know that when Sarah said "that sounds fine," she meant she'd think about it. It doesn't know that the deadline Dave mentioned was aspirational. It doesn't know that one of the "decisions" on the list was a suggestion someone floated that nobody actually confirmed. This is the core thing to understand: AI processes text. It doesn't process subtext. The meeting transcript might say "Marcus will handle the budget review." What the transcript can't tell you is that Marcus has a conflict with the finance lead, has three other deliverables due that week, and said "sure" in the tone of someone who absolutely didn't mean it. You know that. The model doesn't. That gap matters most in two situations. One is when the action list will be used as a record, something people refer back to in disputes or reviews. The other is when the stakes are high enough that a misattributed task or a fictitious deadline creates real problems downstream. In those cases, the AI does the first draft and you do the verification. Not the other way around. The AI decision log prompts article is worth reading before you turn meeting outputs into a record people will refer back to. A quick word on recurring meetings One-off meetings are easy. You run the notes, you send the list, done. Recurring meetings are where action items go to die. Every week the same items show up, slightly reworded, with new deadlines. Nobody calls it out because calling it out is awkward. The list looks productive. The work doesn't move. If you're using these prompts for a recurring meeting, add one step: before running the new notes, paste in last week's action list and ask the AI to flag which items are appearing for the second or third time. It won't know why they're repeating. But it can surface the pattern, which is usually enough to start a real conversation about what's actually blocking those tasks. That's not a prompt people think to write. But it's probably the most useful one in this list for anyone running a team that meets regularly. Frequently asked questions What's the best way to use AI for action items after a meeting? Start with raw notes or a transcript and use a prompt that extracts tasks with source quotes rather than summaries. Always verify owner names, deadlines, and decision status against what was actually said. AI is good for structure, but the accountability has to be human. Can AI assign owners to action items automatically? It can suggest owners based on what's in the notes, but it shouldn't assign them without confirmation. If the meeting notes don't name someone explicitly, the right move is to flag it as unassigned, not guess. Unconfirmed owners cause more confusion than no owner at all. What if my meeting transcript is messy or incomplete? Use it anyway, but tell the AI what's missing. A prompt like "flag anything where you're uncertain" gives you a list of gaps to fill rather than a confident fabrication. Garbage in produces garbage out, but flagged garbage is easier to fix than confident fiction. See what AI can and can't do for more on this. Should I paste confidential meeting notes into ChatGPT? Not if they contain client PII, employee data, legal disputes, financial forecasts, unreleased product information, or anything subject to regulatory or contractual restrictions. Check your company's approved tools policy first. When in doubt, escalate to legal, security, or whoever owns data governance at your organization. How do I handle action items that weren't explicitly agreed in the meeting? Keep them off the list, or flag them separately as "suggested but not confirmed." Sending out action items that weren't agreed creates false consensus. Someone will come back to that email later and dispute it. Better to send a shorter, accurate list than a complete one that includes things nobody signed up for. What's the difference between an action item and a decision? A decision is something agreed or confirmed in the meeting. An action item is a task someone agreed to complete. Decisions create context. Action items create work. Mixing them up is how people end up confused about what was resolved versus what still needs doing. Use Prompt 3 above to separate them before you send anything out. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Meeting Agenda Prompts: 10 Templates That Work 10 copy-paste AI meeting agenda prompts for decision meetings, kickoffs, 1:1s, retros, and more. Includes the reusable formula behind every good prompt. ai meeting agendameeting promptsai at work AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts AI Meeting Follow-Up Prompts (10 Templates) 10 copy-paste AI meeting follow-up prompts for ChatGPT and Claude. Turn messy notes into recaps, action items, and exec summaries without inventing decisions. ai promptsmeeting follow-upaction items --- ## AI Anxiety at Work: How to Stop Spiraling URL: https://dontreplace.me/articles/ai-anxiety-is-real-here-s-how-to-stop-spiraling ai anxietyai job fearwill ai replace my jobhow to use ai at workcareer survival AI Anxiety at Work: How to Stop Spiraling 65% of employees feel AI anxiety at work. Here's how to convert that fear into something useful without buying a $997 course or losing more sleep. May 07, 2026 Updated 12 May 2026 You're lying awake at 2am, running the math on how many of your job tasks a chatbot could do. The number keeps getting bigger. This is AI anxiety at work, and according to EY research, 65% of employees feel it. So at minimum, you've got company. That's not nothing. Shared dread is still dread, but at least you're not broken. The problem isn't that you're anxious. The problem is that the anxiety isn't converting into anything useful. It's just sitting there, a low hum in the background of every meeting where someone mentions "AI integration." So let's do something about it. AI anxiety at work is rational, not a sign you're falling apart Here's the thing nobody says in the productivity content and the LinkedIn reassurance posts: your fear makes sense. AI tools are getting genuinely better at tasks that used to require expensive humans. The tools are real. The improvement is real. The companies looking at headcount through a cost-reduction lens are real. You're not catastrophizing. You're paying attention. Therapists are noticing. CNBC has reported that therapists are seeing a rise in AI-related anxiety sessions, with clients specifically worried about job loss, identity, and what it means to be good at something if a machine can approximate it. This is a clinical trend now. It has a name. So the first thing to do is stop treating your fear like a character flaw. It isn't. The fear is information. The question is what you do with it. Why your anxiety is working exactly as designed Rule #1 in Don't Replace Me is: Fear Is a Subscription Model. The headlines make money by keeping you scared. The guy with the $997 "AI Survival Masterclass" makes money by keeping you scared. The podcast that releases three episodes a week about AI replacing everything makes money by keeping you scared. Your career gets better when you stop renting your nervous system to them. That's not dismissing the real risk. It's pointing out that the loudest voices in this conversation have a financial interest in your panic. Calm, well-adapted people don't click as hard. They don't buy as many courses. They're bad for engagement. The average AI doom headline is not journalism. It's a business model. Recognizing that doesn't mean you ignore the underlying reality. It means you stop letting the scariest possible framing of that reality set up camp in your head at midnight. This isn't your first apocalypse In 1811, a group of workers in Nottingham started breaking machinery. They were called Luddites, and history has turned them into shorthand for people who are afraid of progress. But they weren't stupid. They were scared, and they had good reason to be. Machines were genuinely changing their work. The part that gets left out: most people adapted. The machines stayed. The people who learned the new rules kept working. The same pattern showed up with the printing press, the telephone, the calculator, the personal computer, the internet, the smartphone. Every single time, the new thing killed some jobs and created others, and the people who treated the new thing as something to understand rather than something to outrun came out fine. The data on AI job replacement is messier and more optimistic than the headlines suggest. Goldman Sachs estimates AI could affect 300 million jobs globally. The same report notes that throughout history, automation has consistently created more jobs than it destroyed. The World Economic Forum's 2025 Future of Jobs report projects 170 million new roles emerging while 92 million are displaced. That's a net positive, buried under the scarier number. None of that makes the transition frictionless. But it does mean you're not facing something historically unprecedented. You're facing the latest version of a pattern humans have been navigating for centuries, and mostly winning. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → The generational split nobody talks about Gen Z is the most anxious cohort about AI, and also the most adaptable. That's not a contradiction. Younger workers came up in a world where every skill they learned in school felt slightly outdated by graduation. They're used to the ground shifting. The anxiety is real, but so is the muscle memory for dealing with change. They've had to build it. Older workers tend to feel the anxiety differently. Not panic, exactly, but a quieter dread about relevance. About whether the expertise they've spent 20 years accumulating still means something when a tool can approximate parts of it in seconds. Both fears are valid. Both are also somewhat misdirected. The real risk isn't that AI knows what you know. It's that someone else starts using AI for the parts of their job they hate, gets faster, and you're still doing everything manually because the learning curve felt inconvenient. That's the actual threat. Not a robot. A colleague with a better workflow. This is worth sitting with for a second. Most people picture the AI replacement scenario as a dramatic announcement: "We're switching to AI, please pack your desk." In reality it's quieter. Someone on your team starts producing reports twice as fast. Someone else stops asking for help with the stuff that used to take them all afternoon. The comparison happens before the conversation does. What actually helps: the first 72 hours Anxiety without an outlet becomes paralysis. Here's a concrete way to spend 72 hours that converts the dread into something more useful. First 24 hours: audit your tasks, not your job title. Your job title isn't what AI can replace. Specific tasks are. Sit down and write out what you actually do in a typical week. Every recurring task. Then split them into two lists: tasks that require judgment, relationships, context, and physical presence, and tasks that are mostly information processing, formatting, summarizing, or pattern-matching. The second list is where AI can help you. Not replace you. Help you. The goal is to get faster at the stuff that doesn't need you, so you have more time for the stuff that does. Next 24 hours: touch one tool. Not a course. Not a YouTube rabbit hole. Open ChatGPT or Claude, pick the most tedious thing on your second list, and try offloading it. Write a prompt. See what comes back. It'll probably be mediocre the first time. That's fine. You're building intuition, not installing a replacement for yourself. The practical guide to using AI at work walks through exactly this if you want a framework. The core idea is simple: start with the tasks you hate. Not the ones that define you. The ones that drain you. Final 24 hours: name one skill that's yours. What do you do that requires someone to actually trust you? Where does your judgment matter? What context do you carry that isn't written down anywhere? Client relationships. Institutional knowledge. The ability to read a room. The taste that comes from doing something for years and knowing when something is right versus merely correct. That's your moat. It doesn't eliminate the need to adapt, but it's the foundation. If you know what's yours, you can build from there. How AI anxiety at work shows up differently by industry The fear has the same shape everywhere but very different textures depending on what you actually do for a living. In marketing and content, the anxiety tends to be acute and immediate. People who write for a living watched GPT-4 show up and produce passable copy in seconds. The dread is visceral because the comparison is obvious. But the thing those writers usually miss is that "passable" is doing a lot of work in that sentence. Clients who need passable got cheap. Clients who need good still need a human who has taste, who understands the brand, who knows when something lands versus when it sounds like an AI trying to sound human. That gap is real and it's not closing as fast as the panic suggests. In finance and accounting, the anxiety is about different things: that AI can now draft reports, summarize filings, flag anomalies in data sets, and produce analysis that used to take a junior analyst a full day. That's true. It's also true that the judgment call, the client relationship, the regulatory interpretation that requires actual experience, none of that is going anywhere. The junior roles are changing. The senior roles are changing more slowly. In HR and management, the fear is often quieter but cuts deeper. These are roles built on human judgment, on reading people, on making calls that affect other people's lives. The idea that an AI could screen candidates or flag performance issues feels like a category violation, not just a workflow change. And honestly? The instinct there is correct. The tools that do those things are often doing them badly, with bias baked in at the training stage. In trades and physical work, the anxiety is almost zero, and that's mostly justified. If your job requires showing up somewhere with your body and doing something that requires manual skill, context, and real-time judgment, you're not the person losing sleep about this. The robot plumber is not imminent. The point isn't to map your job onto a risk chart. It's to get specific about what the anxiety is actually about, because vague dread is harder to work with than a specific worry. When the anxiety stays anyway Here's the honest part. Some of you will do all of the above and still feel it. The author of this site's companion book said it plainly in an interview: "Every day I see new tools and think I'm going to be replaced and die poor. But I still wake up every day and try something new." The guy who wrote the survival guide still feels the fear. He just stopped letting it be the whole story. That's the realistic version of managing AI anxiety. Not eliminating it. Not replacing it with toxic positivity about how AI is actually amazing. Just refusing to let it make decisions for you. Fear is information. Panic is just fear with the volume too high. If the anxiety is genuinely disrupting your sleep, your work, your ability to function, that's worth talking to someone about. Not because you're broken, but because you're carrying something heavy and you don't have to carry it alone. What therapists are seeing in these sessions is that naming the fear specifically, instead of letting it be a formless dread, usually helps. What's the actual thing you're afraid of? Losing income? Losing identity? Being seen as obsolete? Being overtaken by someone younger and faster? Get specific. Formless dread is the hardest kind to work with. The one question worth asking yourself Here's a way to cut through the noise. In five years, which version of you is more at risk: the one who spent this period anxious and waiting, or the one who spent it learning to use the tools well enough that they're part of your workflow? The skills that protect your career from AI aren't exotic. They're domain expertise combined with basic AI fluency combined with the human relationships that don't scale. You probably already have the first one. The second takes 72 hours to start building. The third you've been building your whole career. The Luddites who adapted kept working. The ones who waited for the machines to go away didn't have that option. The machines don't care either way. You still do. That's actually the advantage. Frequently asked questions Is AI anxiety at work a real thing, or am I just being dramatic? It's real. EY research found that 65% of employees report feeling anxious about AI's impact on their work. Therapists are reporting a measurable increase in sessions focused on AI-related job fears. You're not being dramatic. You're responding to a genuine change in the environment. Will AI actually take my job? Probably not all of it, and probably not soon. AI replaces tasks, not jobs. The realistic risk is that someone else uses AI to do the repetitive parts of their job faster, which makes them more productive and you comparatively slower. Check out what's actually happening with AI job replacement for the longer answer. What should I actually do if I'm scared about AI at work? Start small. Open ChatGPT or Claude and use it for one task you genuinely find tedious. Don't take a course. Don't buy anything. Just touch the tool. Then do it again. The anxiety shrinks once you have hands-on experience with what these tools can and can't actually do. Why does the AI news feel so scary all the time? Because scared people click more. Fear is a subscription model. Headlines that say "most people will adapt just fine but the transition will take a few years" don't get shared. Headlines that say "AI will replace 80% of jobs by 2030" do. The incentive structure of media rewards the scariest framing, not the most accurate one. How is this different from past technology scares? It's not, structurally. The printing press, the industrial revolution, the internet, and now AI all followed the same pattern: displacement of specific tasks, creation of new categories of work, a painful transition period, and then a new equilibrium. The Luddites were scared and right that things were changing. The part they got wrong was thinking the machines could be stopped. Should I be more worried if I'm older in my career? The anxiety shows up differently by generation, but the underlying risk is the same for everyone: people who adapt stay relevant, people who wait get left behind. Older workers have something Gen Z doesn't, which is deep domain expertise and established relationships. Those matter enormously. The thing to add is basic AI fluency, not a career change. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Job Replacement Statistics: What the Data Says Goldman Sachs, WEF, McKinsey: what AI job replacement statistics actually say vs. what gets quoted in headlines. The real numbers, with context. AI job replacementAI statisticsfuture of work How to Use AI at Work: A No-BS Starter Guide Learn how to use AI at work without the hype, the courses, or the 50-tool stack. Start with the tasks you hate and go from there. how to use AI at workAI for beginnersChatGPT at work AI Skills for Non-Technical People (2026) The AI skills non-technical people actually need in 2026. No coding, no bootcamp, no $997 course. Four real skills and a minimum viable stack by profession. ai skillsnon-technicalprompt engineering --- ## AI API Testing Prompts: Catch Broken Integrations URL: https://dontreplace.me/articles/ai-api-testing-prompts-catch-broken-integrations-faster AI promptsAPI testingquality assurancesoftware testingintegration testing AI API Testing Prompts: Catch Broken Integrations AI API testing prompts for contracts, edge cases, authorization, retries, error responses, and human-verified integration evidence. August 12, 2026 Updated August 12, 2026 An API can return 200 OK while doing the wrong thing, reject a valid request with a useless error, accept a duplicate payment, or expose data to the wrong role. The status code is only the beginning of the story. AI API testing prompts can help turn a verified contract into a focused test matrix. They can organize fields, propose boundary cases, identify unanswered questions, draft sanitized request shapes, and summarize observations. They cannot connect to your API unless explicitly equipped to do so, know undocumented behavior, verify a live response, safely infer authorization rules, or decide that an integration is ready for production. AI can draft the test map. A human still has to run the requests, inspect the state changes, and prove what happened. If you need the broader strategy first, use AI test plan prompts. For requirement-level checks, see AI test case prompts. API testing needs both structure and skepticism. A beautiful collection of generated requests is not evidence that anything works. What good API testing actually checks API testing is not merely sending a valid JSON body and confirming that a response arrived. A useful test examines the contract, the business rule, the resulting state, and the behavior around failure. Depending on the endpoint, that may include: the correct method, path, headers, and content type; required, optional, nullable, and read-only fields; accepted formats, ranges, lengths, and enumerations; authentication and authorization by role, tenant, and resource owner; response status, body, headers, and schema; persistence and downstream side effects; duplicate requests, retries, and idempotency; pagination, filters, sorting, and stable ordering; timeouts, dependency failures, and partial completion; backward compatibility across supported versions; logging, correlation IDs, and approved evidence; cleanup and a named human decision owner. The exact list comes from the real contract and the risk of the operation. A read-only lookup is different from a money movement. A public catalog endpoint is different from an employee-record endpoint. The model should help arrange verified facts, not flatten every API into the same generic checklist. A passing test also needs an oracle: the source used to decide whether behavior is correct. That might be an OpenAPI document, acceptance criteria, a business rule, a data model, a security policy, or a confirmed statement from the service owner. “The model thought this response looked reasonable” is not an oracle. It is autocomplete in a tiny lab coat. The reusable AI API testing prompt formula Start with this foundation, then add one of the specialized prompts below: “Act as an API testing planning assistant. I am testing [endpoint and method] in [approved environment and version]. Use only the supplied API contract, business rules, authentication model, supported clients, known risks, and sanitized examples. Produce [requested artifact] with setup, request variations, expected results traced to a source, state or side effects to inspect, evidence to capture, cleanup, and unanswered questions. Separate confirmed requirements from hypotheses. Do not invent endpoints, credentials, responses, limits, security controls, test results, or release decisions.” That last sentence matters. Models are excellent at completing patterns. If an API has a GET /orders/{id} route, a model may confidently assume there is a DELETE /orders/{id} route because the shape looks familiar. Familiar is not documented. Never paste API keys, bearer tokens, passwords, customer PII, production payloads, raw private logs, payment details, health data, HR records, vulnerability details, regulated data, or confidential client material into an unapproved AI tool. Use placeholder secrets, synthetic records, redacted evidence, approved test environments, and human privacy and security review. What to collect before prompting The quality of the output depends on the quality of the input. “Test our user API” produces generic sludge. Gather the smallest safe set of verified context first. InputWhy it mattersHuman check Endpoint, method, and API versionDefines the actual surfaceConfirm current deployed version Contract or schemaSupplies fields and response shapesResolve stale or conflicting docs Business rulesExplains meaning beyond syntaxName the authoritative source Authentication modelDefines identity setupUse placeholder credentials only Authorization matrixDefines who may do whatSecurity owner confirms boundaries Environment and dependenciesMakes results reproducibleConfirm safe test scope Synthetic fixturesEnables realistic requestsPrivacy review and cleanup plan Error conventionsMakes negative checks preciseConfirm status and body requirements Retry and idempotency rulesPrevents duplicate side effectsConfirm keys and retention window Evidence and escalation rulesTurns failures into useful defectsName owners and storage location Keep unknowns visible. If nobody has confirmed the maximum page size, ask the model to create a clarification question. Do not let it manufacture 100 because that number has strong API-documentation vibes. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI API testing prompts Replace bracketed text with verified, non-sensitive information. Every output is a draft test artifact, not proof of execution. 1. Turn an API contract into a test matrix “Review this verified contract, endpoint purpose, business rules, and risk notes: [paste]. Create a test matrix covering valid requests, required and optional fields, response schema, business outcomes, important negative cases, state changes, dependencies, and cleanup. For every expected result, cite the supplied source or label it ‘needs human confirmation.’ Do not invent routes or behavior.” This is the useful first pass when documentation is spread across an OpenAPI file, a ticket, and three messages nobody wants to admit are part of the specification. Ask for traceability. Each row should connect to a contract field, requirement, or risk. Rows without a source are questions, not assertions. Then have the API owner review the matrix before anyone mistakes generated breadth for approved scope. 2. Test required, optional, nullable, and read-only fields “Using this request schema and field rules: [paste], draft tests for omitted required fields, omitted optional fields, explicit nulls, empty values, defaults, unknown fields, read-only fields, deprecated fields, and conflicting combinations. Include a sanitized request shape, expected status and body only where documented, state inspection, and cleanup.” APIs often blur “missing,” null, empty string, empty array, and zero. The differences matter. A PATCH endpoint may treat omission as “leave unchanged” and null as “clear this value.” If that distinction is undocumented, the test should expose the ambiguity instead of guessing. Also inspect what persists. A response may correctly reject a read-only field while a background process quietly writes part of the request. Human testers verify both the visible response and the resulting state. 3. Probe boundaries and malformed inputs “From these verified formats, lengths, ranges, enumerations, and cross-field rules: [paste], propose boundary and malformed-input tests. Cover values at, below, and above documented limits; invalid types; malformed dates; unsupported enum values; Unicode where relevant; duplicate elements; and conflicting fields. Do not generate exploit payloads. Mark undocumented limits as questions.” Boundary testing should follow the contract. If an amount permits two decimal places, test the documented edges and rounding behavior. If the contract says nothing about rounding, ask who owns that rule. Do not turn a general AI tool into an improvised security scanner. Injection testing, vulnerability validation, and adversarial payloads need authorized scope, qualified owners, controlled tools, and careful evidence handling. 4. Challenge authentication and authorization assumptions “Using this verified authentication flow, role-permission matrix, tenant model, and resource-ownership rules: [paste], draft an authorized test plan for missing identity, expired identity, wrong scope, wrong role, cross-owner access, cross-tenant access, stale sessions, and role changes. Use placeholder tokens only. Include setup, expected source, response checks, state checks, audit evidence, and stop conditions.” Authentication asks who you are. Authorization asks what you may do. A valid token should not automatically grant access to every object whose identifier can be guessed. Test through approved accounts and environments. Never paste real tokens into the prompt. Do not accept “the UI hides the button” as evidence that the API blocks the action. The service response and resulting state are the things that matter. 5. Verify status codes and error bodies “Review these documented response and error conventions: [paste]. Build tests for successful outcomes, validation failures, missing resources, conflicts, authentication failures, authorization failures, rate limits, dependency failures, and unexpected server errors. Check status, content type, response schema, machine-readable error code, safe message, correlation ID, and absence of sensitive internals.” A useful error tells an approved client what category of problem occurred without dumping stack traces, queries, secrets, or private records. It should also be stable enough that clients do not have to parse a witty sentence. Verify business truth, too. 201 Created is wrong if no resource exists. 400 Bad Request may be wrong if the operation partially changed state. The body and database effect can contradict the status line. 6. Test pagination, filtering, and sorting “Using these verified collection rules and a synthetic dataset with known records: [paste], draft tests for default page size, minimum and maximum page size, empty results, final page, invalid cursor, repeated cursor, filters, combined filters, ascending and descending sort, ties, newly inserted records, and stable traversal. Include expected record IDs from the fixture rather than invented data.” Collection endpoints can look healthy while quietly skipping or duplicating records. Offset pagination may shift when new records arrive. Cursor pagination may reuse or invalidate cursors. Tied sort values can create unstable ordering unless a secondary key is defined. Build a known synthetic fixture, then compare actual IDs and counts. “The page looked about right” is how missing records acquire permanent residency. 7. Check idempotency, duplicates, and retries “From this documented mutation, idempotency policy, retry behavior, and side-effect map: [paste], create tests for initial submission, exact retry with the same key, retry with changed content, duplicate request without a key, concurrent duplicate submissions, expired keys, timeout followed by retry, and downstream delivery. Include state, response, audit, and cleanup checks. Do not assume undocumented guarantees.” This matters most when an operation charges money, creates orders, sends messages, provisions resources, or triggers external work. A client may time out after the server succeeds and retry because it never saw the response. The human tester must inspect every relevant side effect. One database row does not prove one email, one payment, or one downstream event. Coordinate failure simulation with environment owners so a test does not become tomorrow’s incident review. 8. Explore rate limits, timeouts, and dependency failure “Using the approved non-production environment, documented limits, timeout policy, dependency map, and safe simulation methods: [paste], draft tests for approaching a rate limit, exceeding it, retry guidance, client timeout, upstream timeout, unavailable dependency, slow dependency, partial response, recovery, and observability. Include stop conditions and prohibit uncontrolled load.” Do not ask AI to invent load levels, then blast a shared environment. Rate and resilience tests require approved thresholds, monitoring, isolation, and owners who know the blast radius. Useful checks include whether limits apply by user, token, tenant, or address; whether retry headers match reality; whether failures expose private internals; and whether recovery creates duplicate side effects. AI risk assessment prompts can help structure the human review before testing begins. 9. Review contract and backward compatibility “Compare these two verified API contract versions and the supported-client policy: [paste]. Produce a compatibility review covering removed or renamed fields, changed requiredness, type changes, enum changes, default changes, status or error changes, path changes, pagination changes, and semantic behavior. Separate additive changes from potentially breaking changes and draft verification tests for each flagged item.” A schema can remain technically valid while semantics break a client. Changing the meaning of a status, default sort order, timestamp zone, or empty collection may cause more damage than adding a field. Use the output to focus regression testing, then run the checks against supported clients and versions. The model compares supplied documents; it does not know which undocumented behaviors clients depend on. 10. Convert observations into a reproducible defect “Turn these sanitized request and response observations into a draft API bug report: [paste]. Include environment and build, endpoint and method, safe setup, exact steps, sanitized request, actual response, expected result traced to a requirement, resulting state, reproducibility, timestamps or correlation IDs, impact questions, cleanup, and missing evidence. Do not assign severity or claim a root cause.” A raw response dump is not a defect report. The reader needs the starting state, operation, expectation, observed behavior, and evidence that makes the result repeatable. Use AI bug report prompts for more reporting patterns. A human still confirms the defect, removes sensitive data, judges impact, routes it, and decides whether it blocks release. A practical workflow that avoids test theater Use these prompts inside a human-owned process: Confirm the source of truth. Resolve contract versions and conflicting requirements. Define risk and scope. Name the endpoint, environment, operation, and potential impact. Sanitize the input. Replace secrets and personal data with safe fixtures. Generate a draft artifact. Ask for traceability and explicit unknowns. Review before execution. API, QA, security, and product owners correct assumptions. Run real tests. Use approved clients, automation, environments, and accounts. Inspect side effects. Check persistence, queues, events, downstream systems, and audit records. Capture evidence. Store sanitized requests, responses, versions, timestamps, and correlation IDs. Triage with humans. Confirm defects, severity, ownership, and release impact. Turn stable checks into regression coverage. Keep exploratory questions separate from automated assertions. For realistic fixtures, AI test data prompts can help design synthetic records. For a final release-oriented pass, use an AI QA checklist. In both cases, AI organizes the work. It does not create evidence by describing what evidence might look like. Common mistakes Treating an OpenAPI file as the whole truth A contract may describe shape without explaining business meaning, permissions, side effects, retry guarantees, or operational limits. Pair it with verified requirements and owner review. Letting AI invent expected responses A plausible JSON body is still fiction. Expected results need a source. Unknown behavior should become a question. Testing only response codes Inspect the body, headers, persisted state, downstream events, audit trail, and user-visible outcome where relevant. A neat status can hide a messy failure. Pasting live secrets into a chatbot Do not. Use placeholders and approved secret injection at execution time. Keep production data and private logs out of general AI tools. Confusing generated tests with executed tests A model can write “verify that no duplicate order is created.” It has not verified anything. Execution records and observed evidence belong to humans and test systems. Automating unstable assumptions Review generated checks before placing them in CI. An invented expectation becomes a reliable false alarm, which is a very efficient way to train a team to ignore testing. Frequently asked questions Can ChatGPT test an API for me? It can help design tests and draft code or request examples from supplied documentation. Unless it has an explicitly configured, authorized connection, it cannot call the API or observe results. Even with tools, humans must control credentials, environment scope, evidence, and release decisions. Can I paste an OpenAPI specification into an AI tool? Only if the specification is approved for that tool and contains no secrets or confidential details that policy prohibits sharing. Internal endpoint names, schemas, and security models may themselves be sensitive. Follow your organization’s data rules. Should AI generate my API test cases or test code? It can produce a useful first draft. Review every endpoint, assertion, fixture, dependency, and cleanup step. Run the code in a controlled environment and keep human ownership of coverage and risk. How do I stop AI from inventing endpoints or fields? Supply the exact current contract, require every test to cite a supplied source, instruct the model to label unknowns, and reject any route or field that cannot be traced. Then perform a human review. Are AI-generated security tests safe? Not automatically. Security testing requires authorization, qualified ownership, controlled environments, approved tools, and careful handling of vulnerability evidence. Do not use a general prompt as permission to attack a system. What is the best first API testing prompt? Start by turning one verified endpoint contract into a traceable matrix of valid requests, negative cases, expected outcomes, state checks, and unknowns. Narrow, reviewed scope beats a giant generic checklist. Can AI decide whether an API bug blocks release? No. It can summarize supplied impact evidence and missing questions. Product, engineering, QA, security, and business owners judge severity and release risk. How should I use API prompts in an existing QA process? Use them to accelerate preparation, gap review, documentation, and reporting. Keep real execution, evidence collection, defect confirmation, and sign-off in the existing human-owned process. If you need a broader reality check on those boundaries, read what AI can and cannot do. The bottom line AI API testing prompts are useful when they turn verified context into sharper questions and executable drafts. They are dangerous when polished output is mistaken for a contract, a test run, or a security judgment. Give the model safe facts. Demand traceability. Keep unknowns unknown. Then run the tests, inspect the side effects, and make the decision with humans who understand the system and its consequences. That is the wider point of Don’t Replace Me: use AI for speed and structure without handing it facts, access, judgment, or accountability it does not possess. A fast assistant can improve API testing. It cannot own whether your integration works. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance AI Compatibility Testing Prompts: Catch Breakage 10 AI compatibility testing prompts for browser, device, OS, responsive, locale, network, defect triage, and evidence-based release decisions. AI promptscompatibility testingcross-browser testing AI Exploratory Testing Prompts: Find Hidden Bugs AI exploratory testing prompts for focused charters, unusual user journeys, sharper observations, reproducible bugs, and human-owned release decisions. AI promptsexploratory testingquality assurance --- ## AI Assumption Log Prompts: Stop Planning on Guesswork URL: https://dontreplace.me/articles/ai-assumption-log-prompts-stop-planning-on-guesswork AI promptsproject managementassumption logsplanning AI Assumption Log Prompts: Stop Planning on Guesswork Use 10 AI assumption log prompts to expose project guesswork, design evidence checks, assign review owners, and stop unverified beliefs becoming commitments. September 16, 2026 Updated September 16, 2026 Your project plan says the vendor can handle the volume. Someone remembers hearing that on a call. Nobody can find the call notes. Somehow, this has become a launch commitment. That is what an assumption log is for: catching the moment a reasonable guess puts on a suit and starts issuing instructions. These AI assumption log prompts help you extract unverified beliefs, define evidence checks, and keep decisions connected to what the team actually knows. AI can draft the table and ask useful questions. It cannot confirm a vendor limit, approve a security exception, or make an unavailable colleague agree to a deadline. Use the templates below with your approved workplace AI tool. Start with one consequential assumption rather than feeding the model your entire company and hoping it becomes a project manager. What belongs in an assumption log? An assumption is something you are treating as true so planning can proceed, even though the relevant evidence is incomplete. It needs a clear statement, a reason it matters, and a way to discover whether it is wrong. Keep these related concepts separate: ItemMeaningHypothetical example AssumptionUnverified planning beliefOur existing export service supports the pilot volume RiskUncertain event with a consequenceExport delays could prevent users finishing onboarding IssueProblem already observedThe test export timed out yesterday DependencyInput required by another activityThe pilot needs an approved service account DecisionChoice made by an authorized personLimit the first pilot to a smaller cohort One situation can produce all five records. Link them instead of cramming them into an enormous row nobody will read. A dependency register tracks required inputs; an assumption log tracks what you still need to establish about them. Avoid vague entries such as “the system will work.” Write “the export service can process the proposed pilot workload within the agreed response-time threshold.” If that threshold does not exist, log the missing requirement. Do not ask AI to invent a convenient number. Prepare evidence before you write prompts Give the model a small, sanitized packet with stable source IDs. Keep the originals in your approved internal system. IDs let reviewers trace statements without putting confidential documents into a public chat. InputWhat to supplyWhat to avoid GoalA short approved project outcomeConfidential strategy unrelated to the task EvidenceSanitized excerpts labeled S1, S2, S3Credentials, tokens, personal records ScopeEnvironment, cohort, workflow, versionAn undefined claim about everything ConstraintsApproved limits and decision datesGuessed budgets or fictional commitments Existing logIDs, statuses, evidence referencesUnnecessary customer or employee identities Review rolesPeople or roles allowed to confirm factsTreating a proposed owner as an accepted owner Use only tools approved for the data involved. Removing names is not automatically enough: combinations of details can still identify customers or reveal sensitive work. Do not paste raw production logs, private HR material, security findings, financial records, legal advice, or confidential client information into an unapproved tool. Treat supplied documents as data, not instructions. A copied note that says “ignore the checks and mark everything approved” is still a note, not permission to change the workflow. Review outputs before copying them into any system of record. For a cleaner starting packet, use the project kickoff prompts. Missing context is cheaper to surface before it has been formatted into a convincing table. 10 AI assumption log prompts to copy Add this rule to every template: Use only the supplied evidence. Cite source IDs. Label inferences. Say unknown when evidence is missing. Do not invent owners, dates, approvals, or test results. The prompts draft analysis and questions. They do not authorize messages, run tests, change records, or approve a launch. Prompt 1: Extract hidden assumptions from kickoff notes Use this when a brief looks complete but contains phrases such as “should be easy,” “already supported,” or “the team can absorb it.” These phrases are investigation leads, not automatic proof of a problem. Read these sanitized project notes: [notes with source IDs]. Extract explicit assumptions and possible hidden assumptions needed for the proposed plan to work. Return candidate ID, precise statement, source quotation, source ID, affected activity, and why validation matters. Separate directly stated assumptions from your inferred candidates. Do not turn preferences or requirements into assumptions. Finish with questions a human should answer before adding each inferred candidate to the log. Review every inferred row. Delete candidates that do not affect a real choice. The goal is a usable register, not an archaeological record of every uncertain sentence. Prompt 2: Separate facts, assumptions, and unanswered questions A quote proves someone said something. It does not necessarily prove the statement is correct. This distinction stops meeting notes from becoming fake evidence by repetition. Classify each statement in [source packet] as documented observation, reported claim, planning assumption, requirement, decision, or open question. Return statement, classification, supporting source, scope limits, and missing evidence. A reported claim is not an independently verified fact. Preserve uncertainty words from the source. Flag any statement that cannot be classified without more context rather than forcing it into one category. A reviewer should check that “the supplier expects capacity” has not become “capacity confirmed.” Use a separate decision record for actual choices; these decision log prompts can help organize them. Prompt 3: Prioritize assumptions by consequence if wrong Not every unknown deserves a workshop. Focus attention on beliefs that affect expensive, irreversible, customer-impacting, or time-sensitive choices. Review [assumption log] against [project goals and known constraints]. For each assumption, describe a plausible consequence if false, the affected decision, reversibility, and when the team would need to know. Use qualitative priority with an explanation, not fabricated probabilities. Separate consequences supported by the packet from scenarios you propose for review. Return the smallest set of assumptions that need attention before [decision gate]. Have accountable owners confirm the consequences and priority. A model cannot estimate operational impact from missing architecture. For broader consequence mapping, pair this with change impact analysis. Prompt 4: Design a cheap, falsifiable evidence check “Ask whether it seems fine” is not much of a test. A useful check states what would count against the belief and what the result would actually cover. For assumption [ID and statement], propose the smallest authorized evidence check that could challenge it. Use these constraints: [budget, environment, access, timing]. Return method, required inputs, proposed reviewer, success criterion drawn from supplied requirements, failure signal, limitations, and next decision. If a threshold is absent, mark it as requiring agreement. Prefer existing documentation or a safe scoped test where appropriate. Do not suggest production changes, sensitive data access, or security testing without explicit authorization. A subject-matter expert must approve the method and interpret the result. One passing sample does not validate every configuration. Good acceptance criteria make this review much less theatrical. Prompt 5: Ask an owner for confirmation without implying approval Ownership is an agreement, not a name that appeared in an AI table. Draft a narrow request the recipient can answer accurately. Draft a short confirmation request for [proposed reviewer role] about assumption [ID]. Include the exact claim, relevant source references, scope, why the upcoming decision depends on it, and the evidence requested. Offer response options: confirmed within stated limits, contradicted, needs more investigation, or wrong reviewer. Describe [date] only as a requested response date unless the packet shows an accepted commitment. Do not send the message. Check the recipient and wording yourself. Silence is not confirmation. If no reviewer accepts ownership, leave that gap visible instead of letting the formatting solve it on paper. Prompt 6: Set review triggers instead of forgetting the log Evidence ages. A statement can be well supported for one vendor version and irrelevant after an upgrade. Review should follow meaningful changes, not only calendar guilt. For each row in [log], propose conditions that should trigger revalidation: version changes, workload changes, scope expansion, supplier changes, expired evidence, or approaching decisions. Base triggers on the actual statement and source scope. Return ID, trigger, rationale, proposed review role, and requested review point. Separate event-based triggers from optional calendar reviews. Do not assign dates that have not been agreed. Keep triggers observable. “Review if circumstances change” is too vague to help anyone. “Review before doubling the pilot cohort” gives someone a recognizable moment to act. Prompt 7: Find contradictions without choosing the convenient source Two documents can disagree because one is stale, because they describe different environments, or because someone is wrong. AI should surface the collision, not quietly vote. Compare [source packet] with [assumption log]. Identify claims that conflict or appear inconsistent. For each pair, show exact source IDs, dates if supplied, scope differences, and the decision affected. Do not declare the newest source correct merely because it is newest. Return possible explanations as hypotheses and a specific question for the appropriate reviewer. Keep unrelated claims separate. Check the quotations against their sources. Ask whether the apparent contradiction disappears when environment, cohort, or terminology is clarified. A confident synthesis that erases disagreement is worse than an untidy but honest log. Prompt 8: Update a row when new evidence arrives Do not overwrite history. A reviewer needs to understand what changed and why the status changed with it. Propose a patch to [existing log] using [new evidence with IDs]. Return ID, old value, proposed value, evidence reference, reason, and required reviewer. Preserve previous evidence and uncertainty. Distinguish supported within scope, contradicted, partly tested, and unresolved. Leave status unchanged when the evidence does not address the original claim. Do not apply the patch; produce a reviewable change list. Approve changes individually. A failed test might contradict an assumption, or the test setup might be invalid. The log should reflect the human interpretation and its limits, not simply the model's first explanation. Prompt 9: Prepare a decision-gate summary An executive summary should make uncertainty legible. It should not launder it into a reassuring shade of green. Prepare a decision brief from [log] for [specific gate]. Include supported assumptions and their limits, contradicted assumptions, unresolved assumptions that affect the gate, missing reviewers, and available options from the supplied plan. Cite IDs for every statement. Separate evidence from recommendations. List what the accountable decision-maker must accept, resolve, or defer. Do not announce approval or infer risk acceptance from the planned launch date. Reviewers should be able to click through to each underlying record. Use the go/no-go decision prompts when you need a fuller structure for the actual decision. Prompt 10: Close or escalate an unresolved assumption Some beliefs are tested. Some become irrelevant. Some remain uncertain and require an explicit choice. “Closed” should explain which of those happened. Review assumption [ID] using [latest evidence and decision record]. Draft one proposed disposition: supported within documented scope, contradicted and replanning required, superseded, no longer relevant, or unresolved and escalated. Explain why, cite evidence, retain limitations, and identify affected linked records. If uncertainty is being accepted, require an explicit decision-maker and acceptance record; do not invent them. Draft the next question and suggested recipient role if disposition is blocked. Use escalation plan prompts to frame the request. Escalation is not punishment. It is moving an unresolved decision to someone who can actually make it. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Worked example: a pilot export assumption The following example is fictional. It illustrates the workflow, not a real test result or a recommended performance threshold. A team plans a customer export pilot. Source S1, the kickoff note, says the existing export service should be adequate. Source S2, a test note, records success with a small sample. Neither source defines the proposed pilot volume or acceptable completion time. A poor AI summary says: “Export capacity confirmed; proceed.” A useful draft says: “Existing evidence covers only the sample in S2. Pilot capacity remains unverified, and the target workload is missing.” IDPlanning beliefEvidenceStatusNext action A-01Existing export service supports the intended pilot workloadS1 proposes it; S2 covers a smaller sampleUnresolvedAgree workload and completion criterion, then approve a scoped test A-02Support can handle the expected pilot questionsS1 expects existing coverage; no staffing confirmationUnresolvedRequest coverage confirmation from the appropriate support reviewer Notice what is absent: fictional owners, fabricated due dates, numerical confidence scores, and a magical approval stamp. Suppose a later approved test, recorded as S3, satisfies the agreed criterion for the pilot cohort. A reviewer can mark A-01 supported for that cohort and configuration, preserving the test reference. A larger rollout still triggers review. If the test fails, link the resulting issue and reconsider scope rather than editing the original assumption until it appears true. This example also shows why one log cannot replace every project artifact. Capacity validation, support coverage, and the launch decision are connected, but they have different evidence and authority requirements. A review cadence people can actually maintain At kickoff, capture assumptions that materially affect scope, effort, staffing, or delivery. Before each consequential decision, review the beliefs that decision depends on. Revisit a row when its trigger fires. Close irrelevant entries so important ones do not disappear under administrative compost. In a short review, ask: Which claim matters to our next decision? What evidence changed since the last review? What could show the claim is false? Who has agreed to review it? What happens if we still do not know in time? Use stable IDs and an existing approved project tool. You do not need a new subscription to store twelve rows and three uncomfortable questions. Include unresolved assumptions in stakeholder updates when they change the interpretation of progress, rather than reporting certainty the team does not possess. Frequently asked questions What is an AI assumption log? It is an ordinary assumption log drafted or maintained with AI assistance. The model can extract candidate beliefs, compare evidence, and propose review questions. People still validate the claims, accept review responsibilities, and make decisions. Calling it AI-powered does not make the underlying evidence stronger. Can ChatGPT verify project assumptions? Not merely by reading and restating them. It can help identify what evidence would be needed and analyze material you provide, but analysis can be mistaken or incomplete. Verification depends on appropriate source records, authorized tests, and qualified review. A confident answer is not an independent observation of your project. How is an assumption log different from a risk register? The assumption log records beliefs used for planning. A risk register records uncertain events and their potential consequences. If an assumption might be false and that would threaten delivery, link it to the relevant risk. Keep the claim and its consequence distinct so reviewers know which evidence addresses which question. What columns should an assumption log template include? Start with ID, precise statement, source references, scope, consequence if false, evidence check, review role, status, and review trigger. Add an agreed deadline or decision link when relevant. Keep proposed owners visibly separate from accepted owners. More columns are useful only if somebody uses them to make or review a decision. Should AI assign confidence scores to assumptions? Avoid unsupported numerical scores. A model-generated percentage can look scientific while reflecting no calibrated measurement. Prefer a description of evidence coverage and gaps: tested for one cohort, reported by a supplier, contradicted by a log, or not yet investigated. If your organization has a defined scoring method, supply it and have reviewers check its application. When can an assumption be closed? When a responsible reviewer records why it no longer needs active tracking: adequately supported within scope, contradicted with replanning handled, superseded, or irrelevant. An unresolved assumption may instead need explicit risk acceptance or escalation. Neither a passed deadline nor a lack of objections is a substitute for that record. Make the guess visible before it makes the decision Pick one sentence in your current plan that begins with “we assume.” Give it a source, a consequence, and a check. Use AI to shorten the clerical work, then have a human review the result. That is the practical approach behind Don't Replace Me: use the tool to expose gaps and save effort without outsourcing the judgment that makes the work trustworthy. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Contingency Plan Prompts: 10 Templates 10 copy-paste AI contingency plan prompts for ChatGPT and Claude. Map failures, assign owners, draft comms, and verify assumptions before a crisis hits. contingency planningAI promptsbackup plans AI Dependency Mapping Prompts: Find What Breaks AI dependency mapping prompts for finding upstream work, downstream risks, hidden handoffs, owners, schedule conflicts, and launch blockers. AI promptsdependency mappingproject management AI Dependency Register Prompts: Keep Projects From Surprising You Use AI dependency register prompts to map owners, blockers, assumptions, and downstream impact before one small change becomes a three-week incident. AI dependency mappingproject managementAI prompts --- ## AI Automation Prompts: Automate the Boring Stuff URL: https://dontreplace.me/articles/ai-automation-prompts-automate-the-boring-stuff ai automation promptsworkflow automationSOP templatesAI at workprocess mapping AI Automation Prompts: Automate the Boring Stuff 10 copy-paste AI automation prompts for mapping workflows, writing SOPs, adding approval gates, and QA testing. Practical templates with safety rules included. June 13, 2026 Updated June 13, 2026 There's a version of you that spends three hours a week copying data from one spreadsheet into another. There's another version that delegates that to an AI-assisted workflow, takes 20 minutes to set it up, and uses the other two hours and forty minutes for something that actually requires a brain. AI automation prompts can get you from the first version to the second. But only if you know what to ask, what to hand off, and where to stay in the loop yourself. This is the practical guide. Ten copy-paste prompts, a reusable formula, and a few honest warnings about what happens when people let AI design workflows without any human in the room. What makes an AI automation prompt actually work? The short version: specificity and scope. Most people write prompts like "automate my onboarding process" and wonder why the output looks like a consulting deck nobody asked for. A useful automation prompt names the task, the context, the tools involved, the expected input and output, and any constraints. That's the formula: Task + Context + Tools + Input → Output + Constraints So instead of "automate my onboarding process," you get: "I run a five-person marketing team. Our new hire onboarding takes three hours of my time per person. The steps are [X, Y, Z]. We use Notion, Slack, and Google Drive. Draft a workflow that reduces my manual involvement to just the first meeting and final sign-off." That's something an AI can actually work with. The vague version gets you a confident-looking list of suggestions that could apply to literally any company on earth. Before you start: what not to paste into an AI tool Stop here for two minutes before you throw your company's process docs into ChatGPT. Do not paste any of the following into an unapproved AI tool: API keys, login credentials, or access tokens Real customer data, names, or contact information Employee performance records or HR files Unreleased product strategy or pricing Financial data, legal documents, or security incident details Proprietary process docs that live under an NDA Replace real names with placeholders ("Customer A," "Employee 1"). Swap real figures for sample numbers. Run your company's actual sensitive workflows only in tools your IT or security team has approved and reviewed. If you're not sure what's approved, ask before you start, not after you've pasted six months of customer records into a free tool. This isn't bureaucratic theater. It's just not making a preventable mistake. The prompts below work just as well with fake sample data. AI automation prompts: 10 templates you can use today These are designed for normal work situations. Copy them, fill in the brackets, and adjust for your actual context. Prompt 1: Should this task even be automated? Use this before you build anything. "I'm considering automating [describe the task]. It happens [frequency] and currently takes [time]. The inputs are [what comes in] and the output is [what goes out]. Who currently handles it: [role]. What could go wrong if it runs incorrectly: [consequences]. Given this, help me evaluate whether automation makes sense, what the risks are, and what I'd need to have in place first." This is the prompt most people skip. They jump straight to building the Zapier flow and discover six months later that the "automated" process has been sending the wrong data to the wrong place. Spend ten minutes here first. Prompt 2: Map a messy process before you automate it "Here's a rough description of how we currently handle [process]: [paste your notes]. It's messy and inconsistently followed. Based on this, help me map the actual steps in order, identify where the process usually breaks down, flag any steps that require human judgment, and note any steps that seem redundant or unclear. Use a numbered list." Rule #13 from Don't Replace Me puts it plainly: garbage in, garbage out. If your source notes are vague, the AI will invent plausible-sounding steps. Your job is to review the output and flag anything it invented. The AI is mapping. You're verifying. Prompt 3: Turn rough notes into a working SOP "I've mapped out the following steps for [process name]: [paste step list]. Convert this into a formal Standard Operating Procedure. Include: purpose, scope (who this applies to), step-by-step instructions in plain language, decision points where human judgment is required, and a 'things that can go wrong' section at the end. Do not invent policy, approvals, or compliance requirements I haven't specified." That last line matters. AI will happily write "Step 7: Ensure GDPR compliance" without any idea whether your process actually satisfies GDPR. You need to catch and verify every claim like that. Prompt 4: Design a Zapier or Make-style workflow "I want to automate the following using [tool name, e.g., Zapier, Make, n8n]: [describe trigger and desired outcome]. The systems involved are [list apps]. The input data looks like [describe fields or paste a sample with fake data]. The desired output or action is [what should happen]. Draft a step-by-step workflow including trigger, filters, actions, and error handling. Flag any steps that will need a human check before they run." Do not let this output go live without someone who actually knows the tools reviewing it. AI can describe a logical workflow. It can't tell you whether your specific Zapier plan supports that many steps, or whether the API field mapping is actually correct. Prompt 5: Create intake form questions for a new workflow "I'm building an intake process for [what you're collecting, e.g., client project requests, IT support tickets, content briefs]. The person submitting the form is [describe them]. The person reviewing it needs to be able to [what they need to do with it]. Draft 8-12 intake questions that would capture everything needed without being overwhelming. Flag any questions where the answer will affect routing or priority." Good intake design is half the battle. Most automated workflows fail because the input data is incomplete or inconsistent. This prompt helps you fix that before you build. Prompt 6: Write QA checks for an automation "I've built an automation that does the following: [describe what it does, step by step, using fake/sample data]. Draft a QA checklist I can run before going live. Include: what inputs to test, what outputs to verify, edge cases to check, failure modes to simulate, and what a 'pass' looks like for each check." Test with fake data. Always. Run the live version first in a sandbox or with a small pilot group. An automation that quietly fails at scale is worse than one that never launched. Prompt 7: Identify edge cases you haven't thought of "Here's my planned automation for [process]: [describe it]. I've tested the happy path. Help me think through edge cases that could break it or produce incorrect results. Consider: unusual input formats, missing data, timing issues, duplicate entries, permission problems, and scenarios where the automation should stop and escalate to a human instead of continuing." This one earns its keep. AI is genuinely useful for "what could go wrong" brainstorming because it has seen a lot of broken systems in its training data. You still have to decide which edge cases are worth building for, but it'll surface things you hadn't considered. Prompt 8: Add human approval gates to a workflow "Here's my automation workflow: [describe it]. I want to add human approval gates at the right places. Suggest where a human should review before the automation proceeds, what information they'd need to see to make that decision, what happens if they approve vs. reject, and how long the workflow should wait before escalating or timing out." This is where most automation plans break down. Someone builds a fully automated flow, removes every human checkpoint, and is then surprised when the automated system approves a client deliverable that's completely wrong. Knowing where to keep humans in the loop is a skill. AI can help you map the gates. You decide which ones stay. Prompt 9: Draft a handoff message for an automated step "My automation will send a notification/message to [recipient role] when [trigger]. The context they'll need: [what happened, what they need to do, what the deadline is]. Draft a clear, brief message that explains the situation and next step without requiring them to look anything up. Tone: [professional/casual]. Include a link placeholder and a clear action item." Automated messages that make people hunt for context get ignored. This prompt helps you write the version they actually read. Prompt 10: Review an automation after it's been running "My automation has been running for [time period]. Here's what it does: [describe]. Here's what I've noticed: [any issues, complaints, workarounds people are using]. Help me conduct a post-launch review. Suggest: metrics to check, user feedback questions to ask, common failure patterns to investigate, whether the original goals are being met, and what I'd change in the next version." Automation is not "set it and forget it." You need a rollback plan when something breaks, an audit log to see what ran and when, and a review cycle to catch drift. Build those in before launch, not after the first incident. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → The safety rules, restated plainly You're going to want to skip this section. Don't. AI can draft a workflow, suggest edge cases, write SOPs, and map a process in minutes. It cannot: Verify that your workflow complies with your company's actual security policy Confirm that the integration you've described actually works the way you think it does Take accountability when an automated process sends the wrong data to the wrong person Own the outcome when a client commitment gets made by a bot without human review Do not let AI invent policies, approvals, integrations, data mappings, compliance claims, budgets, timelines, or customer commitments. Review every output before it touches real systems. Keep an audit log. Know how to turn it off. Test with fake data first. The human in the loop isn't a bottleneck. It's the point. For a broader look at where that line is, what AI can and can't do is worth reading before you hand off anything with real consequences. What this connects to in your broader ops stack These automation prompts don't exist in isolation. If you're building out workflows, you probably also need solid project tracking (the AI project management prompts are useful here), clean documentation (AI documentation prompts cover turning messy notes into structured docs), and eventually a policy layer (AI policy prompts help you write the rules people will actually follow). The mistake most teams make is automating before they've documented. You end up with a Zapier flow that nobody understands, that breaks every three months, and that only one person knows how to fix. Map it first. Document it. Then automate. In that order. Frequently asked questions What are AI automation prompts? AI automation prompts are structured instructions you give to a tool like ChatGPT or Claude to help you design, map, or document a workflow. They're not code. They produce text: SOPs, checklists, workflow descriptions, QA steps, intake questions. A human still builds and runs the actual automation. Can I use ChatGPT to build a Zapier or Make workflow? You can use it to describe and plan the workflow, which is genuinely useful. You can't paste the output directly into Zapier and have it run. You'll still need to build the actual automation in the tool, verify the field mappings, and test it before going live. Think of the AI output as a blueprint, not a finished build. What data is safe to paste into an AI tool when designing automations? Fake or anonymized data only, unless your company has approved a specific tool for sensitive use. Replace real customer names, IDs, and contact details with placeholders. Never paste API keys, credentials, financial records, HR data, or anything under an NDA. Check your company's AI usage policy first. If there isn't one, that's a gap worth flagging. The AI policy prompts article has templates to help draft one. Should every step in an automation have a human approval gate? Not every step, but more than most people include. The right gates are where: the output affects a customer directly, the data is sensitive, the action is irreversible, or the decision requires context the automation doesn't have. AI is helpful for identifying where those gates should be. You decide which ones to keep. How do I know if a task is worth automating? Use Prompt 1 above before you build anything. The basic test: does it happen frequently, follow consistent rules, use predictable inputs, and not require judgment calls? If yes, it's a candidate. If the answer is "sometimes, it depends," that's a signal to document the decision logic before automating, not after. A good starting point is the how to use AI at work guide for thinking about which tasks to hand off first. What happens when an automation breaks? You need a rollback plan before launch, not after the first incident. That means: knowing how to disable the workflow immediately, having an audit log of what ran and when, and knowing who to notify. AI can help you draft a rollback plan as part of your QA checklist. The accountability for fixing it is yours. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Workflow Audit Prompts: 10 Templates 10 copy-paste AI workflow audit prompts for ChatGPT and Claude. Map processes, find bottlenecks, and sort tasks by automation potential safely. ai workflow auditworkflow automationAI prompts AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts --- ## AI Brainstorming Prompts That Actually Work URL: https://dontreplace.me/articles/ai-brainstorming-prompts-better-ideas-less-outsourcing ai brainstorming promptsai at workcreative promptschatgpt promptsbrainstorming AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. June 09, 2026 Updated June 09, 2026 Most brainstorming sessions produce the same dozen ideas, reshuffled. You know the meeting. Someone says "what if we made it more interactive?" Someone else says "what about a newsletter?" The whiteboard fills up. Nothing gets chosen. Everyone leaves feeling vaguely like they did something. AI brainstorming prompts can break that loop, but only if you use them right. The trap is treating the tool like a creative oracle. It isn't one. It's fast, it's tireless, and it has read more marketing copy than any human alive. That also means it trends toward the mean. The safe angle. The LinkedIn-ready insight. The idea everyone else already had. Your job isn't to let AI brainstorm for you. Your job is to use it to generate raw material, then apply the thing AI genuinely can't fake: your judgment about what fits your audience, your brand, and this specific moment. Here's how to do that without outsourcing your brain. Why most AI brainstorming produces garbage The ideas are generic because the prompts are generic. "Give me ideas for a marketing campaign" produces exactly the content you'd expect: a list of ten things that sound like they were written by a committee in 2019. The fix isn't a better AI model. It's a better prompt. Specificity is the whole game. The more context you give, the fewer generic answers you get back. Tell it the audience, the constraint, the goal, the tone, the things you've already tried, and what "good" looks like. That's the formula. Think of it this way: you're not asking AI to be creative. You're giving it a tight brief and asking it to generate variants. Volume first, taste second. The creative judgment stays with you. One more thing before the templates: don't paste confidential information into unapproved AI tools. No unreleased product roadmaps, no real customer data, no employee issues, no legal matters, no financial projections, no security incidents. Anonymize whatever context you need to include. That's not paranoia, it's just common sense. The reusable brainstorming prompt formula Every good AI brainstorming prompt has the same skeleton. You can use this as a checklist before you hit send. Goal + Audience + Constraints + Existing context + What "good" looks like + Format request In practice it looks like this: "I'm trying to [achieve X goal] for [specific audience]. The constraint is [time/budget/channel/tone]. We've already tried [what hasn't worked]. A good idea would [criteria]. Give me [number] options and explain the angle behind each one." That structure will outperform "give me ideas for X" every single time. It's not magic. It's just a better brief. A useful way to think about why this works: AI is a pattern-matching system. Give it a vague pattern and it returns the most statistically average match. Give it a specific, detailed pattern and it narrows toward something closer to your actual situation. It's still not thinking. But the material it produces becomes more useful to someone who is. 10 copy-paste AI brainstorming prompts These are organized by where you are in the process: starting from scratch, sharpening what you have, pressure-testing, and moving to action. Use them in Claude, ChatGPT, or whatever tool your company has approved. 1. Generate rough ideas from a goal and audience I need to [achieve this goal: e.g., increase sign-ups for a free trial, re-engage lapsed customers, build awareness in a new segment]. My audience is [describe: role, context, what they care about, what they're skeptical of]. The channel and format are [e.g., email series, social content, in-app prompt, sales pitch]. Give me 10 rough ideas. For each one, give me the core angle in one sentence. Don't polish them. I want quantity and variety, including some weird ones. 2. Create contrarian or unexpected angles Here's the obvious angle for [topic/campaign/product/initiative]: [paste the obvious version] Give me 8 contrarian or counterintuitive takes on this. The goal is to find an angle that would make the target audience stop scrolling. Each idea should feel like it could start an argument or challenge a common assumption in this space. 3. Turn a boring idea into sharper versions Here's an idea that feels flat: [paste the flat idea] The target audience is [describe]. Give me 6 sharper versions. Each should have more tension, more specificity, or a stronger point of view. Explain what makes each version different from the original. 4. Pressure-test an idea against objections Here's an idea I want to stress-test: [paste the idea] Play devil's advocate. Give me: - 5 reasons a skeptical [target audience member] would dismiss this immediately - 3 ways this could backfire - 2 things we'd need to verify before committing to this Don't be polite. I want the hardest objections. This one is useful before you present anything upward. Better to hear the objections from a chatbot than from your VP. 5. Find constraints and edge cases We're considering [idea or plan]. Help me think through constraints and edge cases: - What assumptions does this idea depend on being true? - What resources (time, money, skills, tools) does it actually require? - What could go wrong in the first 30 days? - What's the edge case where this works perfectly? Where it fails completely? I'm not looking for reassurance. I'm looking for the places where this plan could fall apart. 6. Cluster messy ideas into themes Here's a brain dump of ideas I've collected: [paste your list, however messy] Organize these into 4-6 themes. For each theme: - Name it in 3-5 words - Summarize the core logic in 1-2 sentences - List which ideas from my dump belong in it - Identify any gaps or underexplored angles in that theme Don't invent new ideas yet. Just organize what's here. This one saves a lot of whiteboard time. 7. Create campaign or content hooks The campaign/content is about: [describe the topic, product, or message]. The audience is: [describe]. The goal is: [what action or feeling you want]. The tone is: [formal/casual/bold/educational/etc.]. Give me 10 hooks. Each should be a single sentence or headline that would make someone stop and want to know more. Include a variety of formats: a question, a counterintuitive claim, a specific number, a bold statement, a relatable frustration. Do not use "Discover," "Unlock," "Transform," or "Game-changer." That last line matters. You'd be surprised how often you have to say it. 8. Adapt one idea for different channels Here's a core idea: [paste the idea] Adapt this for the following channels. Keep the core angle but adjust the format, tone, and hook for each: - LinkedIn post (professional, medium length) - Short email subject line + preview text - Slack message to an internal team - 30-second verbal pitch - One-sentence prompt for a sales call Keep the same message. Change how it lands. 9. Build an idea shortlist with scoring criteria Here are [number] ideas I'm considering: [paste ideas] Help me build a scoring framework to choose between them. First, ask me 3 clarifying questions to understand what "good" means in this context. Then, once I answer, score each idea on: - Alignment with goal (1-5) - Feasibility with our constraints (1-5) - Differentiation from what competitors or alternatives already do (1-5) - Risk level (1 = low risk, 5 = high risk) Show the scores in a table. Flag any ideas where the scores conflict in interesting ways. 10. Create a next-step experiment plan I want to move forward with this idea: [paste the idea] Give me a lightweight experiment plan to test it before committing fully. Include: - The one key assumption we need to validate first - The smallest possible version we could run in [timeframe: e.g., one week, one sprint] - What success looks like (what we'd measure) - What failure looks like - What we'd decide based on each outcome Keep it practical. I don't need a project plan. I need a way to find out if this is worth doing. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What humans still have to do after running these prompts The prompts give you raw material. They don't replace the decisions. Before you act on anything AI generates, someone with actual knowledge of your business needs to verify: does this match your brand? Is it feasible? Has legal seen it? Does it reflect what your customers actually think, or is it just what sounds plausible to a language model that's never met your customers? AI will invent market data if you ask it to. It'll produce fake-sounding quotes, plausible-but-wrong statistics, and confident claims about your competitors that aren't true. It doesn't know the difference between "a thing that sounds right" and "a thing that is right." This is what Don't Replace Me calls Rule #5: it's not smart, it's fast. The value is in the volume and the angles. The judgment about what's actually good, honest, feasible, and right for your audience still lives with you. The book covers why taste is the specific skill that doesn't get automated, and it's the thing most people skip thinking about when they're impressed by how quickly AI can generate a list. If you want to see how AI fits into broader creative workflows, the AI marketing prompts guide covers campaign-level thinking, and what AI can and can't do is worth reading before you start trusting AI-generated ideas too much. How to run a better AI brainstorming session in practice Most people open a chat window, type something vague, get mediocre output, and conclude that AI isn't useful for creative work. The problem is almost never the tool. Here's a sequence that actually works, from blank page to shortlist. Start by writing out your actual goal in one sentence. Not "improve brand awareness." Something like: "Get 200 supply chain managers to sign up for a webinar about our new compliance feature, by the end of Q3." That sentence contains the audience, the goal, the format, and a timeline. It's already a brief. Run prompt 1 with that context. Don't edit the output. Just read it and mark the three ideas that feel interesting, even vaguely. Then run prompt 2 on the obvious angle you were already considering. The contrarian takes are often where the better ideas are hiding. Take your marked ideas from step one, combine them with anything interesting from step two, and run prompt 6 to cluster them. You'll usually find two or three real themes in what felt like a messy pile. Pick the strongest theme, run prompt 4 to stress-test it, and then run prompt 10 to figure out how you'd actually test it before committing. You've just gone from blank page to a testable hypothesis in about 20 minutes. That's the workflow. It's not complicated. It just requires you to resist the urge to use AI as a magic idea machine and treat it as a structured thinking partner instead. The "garbage in" problem in brainstorming Vague prompts produce vague ideas. You've probably seen this. "Give me campaign ideas for our software product" returns a list that could work for literally any software company. It's not useless. It's just not good. The fix is specificity about the thing that makes your situation different. What's the constraint? What's the real audience skepticism? What have you already tried? What would make your CEO say "that's interesting" instead of "yeah we've done this"? The more of that context you load in, the more the output actually resembles your situation. Not because the AI understands your company, but because it's pattern-matching against something closer to your actual brief. It's still pattern-matching. You still have to judge the output. But the patterns it finds become more relevant. The AI research prompts guide has a related point worth reading: AI is useful for structuring what you know, not for inventing what you don't. A short note on what not to paste This keeps coming up because people keep ignoring it. Don't paste into AI tools: actual customer names or data, unreleased product features or pricing, employee performance issues, legal disputes, security vulnerabilities, proprietary financial information, or anything you wouldn't want in a public data breach. Anonymize your context. Use "a B2B SaaS company in logistics" instead of your company name. Use "a 35-year-old supply chain manager" instead of a real customer quote. You get almost the same quality of output, and you haven't handed your company's confidential data to a third-party system. Also: never let AI invent facts for you. If a prompt response includes statistics, research findings, customer insights, or competitor data, verify every single one before you repeat it. The model sounds confident whether it's right or wrong. Frequently asked questions What are the best AI prompts for brainstorming ideas? The best AI brainstorming prompts are specific. Include your goal, audience, constraints, and what "good" looks like before asking for ideas. Generic prompts produce generic ideas. Templates like the ones above (contrarian angles, pressure-testing, clustering) give you more useful raw material than "give me ideas for X." Can AI replace brainstorming sessions with my team? No, and you probably don't want it to. AI is useful for generating volume and angles quickly, especially before a session. But the judgment about what fits your audience, your brand, and your moment still needs humans in the room. Use AI to show up to the meeting with better raw material, not to skip the meeting. Is it safe to use AI for work brainstorming? It depends on what you paste in. Don't include real customer data, unreleased strategies, financial projections, legal matters, or anything proprietary in unapproved AI tools. Anonymize your context and check your company's AI use policy first. The output quality is roughly the same with anonymized inputs. Why does AI brainstorming keep producing generic ideas? Because your prompts are too vague. "Ideas for a campaign" is a brief that could apply to any company. Add your audience's specific skepticism, the constraint you're working under, what you've already tried, and what a good idea would accomplish. That context is what forces AI out of the generic middle. Should I tell my manager I used AI to brainstorm? This depends on your workplace. If you're curious about the broader question of disclosure, the guide on whether to tell your boss you use AI covers it in detail. Short version: the work is still yours. The tool is just a faster way to generate raw material. Do I need to verify ideas that AI generates? Yes, especially anything that includes claimed facts, statistics, customer insights, or competitor information. AI models produce confident-sounding output whether or not it's accurate. Treat every AI brainstorm output as a draft, not a source. A human with actual knowledge of your business needs to check feasibility, brand fit, accuracy, and ethics before anything gets acted on. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Meeting Agenda Prompts: 10 Templates That Work 10 copy-paste AI meeting agenda prompts for decision meetings, kickoffs, 1:1s, retros, and more. Includes the reusable formula behind every good prompt. ai meeting agendameeting promptsai at work AI Project Management Prompts That Actually Work 10 copy-paste AI project management prompts for briefs, risk logs, stakeholder updates, and recovery plans. Real templates, no hype. ai project managementproject management promptsai at work --- ## AI Briefing Document Prompts: 10 Templates That Work URL: https://dontreplace.me/articles/ai-briefing-document-prompts-10-templates-that-work ai promptsbriefing documentmeeting preparationexecutive briefingworkplace AI AI Briefing Document Prompts: 10 Templates That Work 10 copy-paste AI briefing document prompts for meetings, exec decisions, client calls, and launch readiness. With a reusable formula and safety rules. July 03, 2026 Updated July 03, 2026 A good briefing document is the difference between a meeting that makes a decision and a meeting that schedules another meeting. Most people write them badly: too long, wrong audience, half the context missing, no clear ask at the end. Then everyone shows up underprepared and the first 20 minutes are recap. AI briefing document prompts won't fix bad judgment, but they'll help you structure what you know, surface what you're missing, and produce something people actually read before they walk in the room. That's the job. Not magic. Useful. Here are 10 copy-paste prompts, a reusable formula, and some honest caveats about when to keep the AI out of it entirely. The reusable AI briefing document prompt formula Before the 10 templates, here's the skeleton. Every good briefing prompt follows this structure: Tell AI what it is: "Write a [type of brief] for [audience]." Tell AI the context: Paste in your notes, bullet points, or rough background. The more specific you are, the less it invents. Tell AI what the reader needs to do: "The reader needs to [decide X / approve Y / show up prepared to Z]." Tell AI the constraints: Length, tone, what to include, what to leave out. Tell AI what you want flagged: Missing information, open questions, assumptions, risks. That's it. Everything below is a version of this formula applied to a real situation. One rule before you paste anything: do not put customer PII, employee records, credentials, confidential strategy, unreleased product details, financial forecasts, legal disputes, security incidents, contracts, board materials, or sensitive client information into AI tools your company hasn't approved. If the brief involves legal, security, HR, financial, safety, or customer-impacting risk, get a human expert to review the output before it goes anywhere. 10 AI briefing document prompts you can use today Prompt 1: Create a meeting pre-read Use this when you need people to arrive ready, not catching up. Write a meeting pre-read for a [type of meeting] with [audience/attendees]. Background context: [paste your rough notes here] The meeting goal is: [one sentence on what decision or outcome you need] Format: - One-paragraph context summary (4-6 sentences) - 3-5 key points they need to know before arriving - The specific decision or question they're being asked to engage with - Any known disagreements or open questions - What preparation they should do in advance (if any) Keep it under [X] words. Flag any gaps in my notes where you'd need more information. Prompt 2: Brief an executive before a decision Executives don't have time for background. They need the ask, the options, the risks, and the recommendation. Fast. Write an executive briefing for [name/role] ahead of a decision about [topic]. Context: [paste relevant background] They need to decide: [state the specific decision] Recommended option: [your recommendation, if you have one] Other options considered: [list them briefly] Key risks of each option: [what could go wrong] What happens if no decision is made by [date]: [consequence] Named owner of next steps: [person] Format: half a page maximum. Lead with the ask. Put context below. End with a clear "Decision needed by [date]." Flag any assumptions I'm making that should be checked before this goes to them. Prompt 3: Prepare for a client call Not a full account review. Just: what do they need to know, what do you need from them, what could go sideways. Write a briefing note to prepare me for a call with [client name, role, or type of client, no PII]. Call purpose: [what the call is for] Current status of the relationship or project: [brief summary] Open issues or concerns: [anything sensitive or unresolved] What I need from this call: [specific ask or outcome] What they're likely to raise: [anticipated questions or friction points] Format: - One-paragraph call context - Three things to know before dialing - Two to three questions I should ask - What success looks like at the end of this call - Any red flags or sensitivities to be aware of Keep it to one page. Note anything you think I should verify before the call. Prompt 4: Summarize a project backgrounder You've got notes, Slack threads, email chains, a half-finished doc. Turn it into something a new stakeholder can read in five minutes. Write a project backgrounder for someone joining [project name] who has no prior context. Raw input: [paste your notes, bullet points, or rough summary here] Cover: - What the project is and why it exists - Current status (as of [date]) - Key stakeholders and their roles - Decisions already made (and why) - Open questions and unresolved issues - Risks or constraints the reader should know about - What happens next and by when Format: clear headers, 400-600 words. Flag anything that seems inconsistent or missing from my notes. If you're building a proper decision trail for the project, the AI decision log prompts are useful to run alongside this. Prompt 5: Write a launch readiness brief Before a product launch, campaign go-live, or system release, someone needs to confirm that everything that needs to be true is actually true. Write a launch readiness brief for [what is launching] on [target date]. Background: [paste relevant context about the launch] Include: - What is launching and for whom - Go/no-go criteria and who owns each check - Risks or dependencies not yet resolved - Stakeholders who need to sign off and by when - Rollback or contingency plan if something goes wrong - Open items that need resolution before launch Format: one page, structured for a leadership or cross-functional review. Flag any criteria where I haven't given you enough information to assess readiness. Prompt 6: Prepare for a vendor demo You're about to watch a vendor show you their product. You need to have good questions, not just sit there impressed by the UI. Write a vendor demo briefing for [product/vendor name or type of tool]. Context: [what problem you're trying to solve, any prior conversations with this vendor, current tools you use] We need to evaluate: - [capability 1] - [capability 2] - [capability 3] Include: - One-paragraph context on why we're looking at this - Five to seven questions to ask during the demo - Three red flags that would disqualify this vendor - What a successful demo looks like for us - Next steps if it goes well Note any areas where I should do more research before the call. Prompt 7: Turn research notes into a briefing memo You've done the reading. Now you need to turn raw research into something a non-expert can use. This is also where the AI research prompts for desk research are worth bookmarking. Turn these research notes into a briefing memo for [audience, e.g., "a senior manager with no technical background"]. Notes: [paste your research notes here] The memo should cover: - The key finding or situation (2-3 sentences) - Why it matters for [our context] - What the evidence says (cite sources I've provided, don't invent new ones) - What's uncertain or contested - What questions remain unanswered - What, if anything, we should do with this information Format: 300-500 words, plain language. If any of my notes are unclear or contradictory, flag them. Do not ask AI to fill gaps in your research by generating facts. It will. They won't be real. Every claim in a briefing document should come from a source you can name. Prompt 8: Create an interview briefing Hiring managers walking into interviews without preparation make bad decisions. This helps. Write an interview briefing for a [role title] interview. Candidate information: [use only information appropriate to share, no protected characteristics, no information your company policy restricts, keep it to their stated experience and application context] Role context: [what the role is, what matters most, what gap it fills] Include: - One-paragraph candidate context based on what I've provided - Three to four areas to explore in depth based on the role requirements - Five specific questions tied to those areas - What a strong answer looks like for each question - Red flags to watch for Keep tone neutral and focused on job-relevant criteria only. Prompt 9: Identify missing context before sending a brief This one's a pre-flight check. Before you send anything to anyone, run this. Review this draft briefing document and identify what's missing. Draft: [paste your draft] Check for: - Claims made without a named source - Assumptions that aren't flagged as assumptions - Missing decision owner or named accountable person - No clear ask or next step - Risks or objections the reader might have that aren't addressed - Sections where more context would help the reader decide - Anything that could be misinterpreted without additional explanation Give me a prioritized list of what to fix before this goes out. This is the prompt most people skip. It's the most useful one. Prompt 10: Tailor one brief for different audiences The same project update lands differently with a CFO, a project team, and a client. Same facts. Different framing. I have a briefing on [topic]. Rewrite it for [audience] who cares most about [their key concern]. Original brief: [paste] For this audience: - Lead with [what matters to them] - Cut or summarize [what doesn't] - Adjust the level of technical detail to [their level] - Change the tone to [more formal / less formal / more direct / more reassuring] - Make the ask explicit for their specific role Do not change any facts. Only change what you emphasize and how you frame it. The full breakdown of when to adjust framing versus when adjusting framing is just spin is covered in Don't Replace Me, specifically in the chapters on taste and judgment. Short version: framing for your audience is good communication. Framing to obscure bad news is a different problem, and AI will happily help you do it if you let it. What AI briefing prompts can't do for you AI is fast at structure. It's bad at knowing what's true, what's missing, and what the room needs to hear versus what it wants to hear. A briefing document written with AI still needs: Verified facts. If a number matters, check it. AI will invent plausible-sounding figures. Named owners. AI puts "[Owner TBD]" everywhere. You need actual names. Real confidence levels. "We believe X" and "We know X" are different sentences. AI flattens them. Your judgment on what to include. AI doesn't know your organizational politics, your stakeholder's sensitivities, or what happened in the meeting three weeks ago that everyone is still annoyed about. The starter guide on using AI at work covers this more broadly, but the short version: treat AI like a fast first drafter. You're still the editor, the fact-checker, and the person whose name is on it. For anything involving legal risk, security incidents, HR matters, financial decisions, or customer impact: have a qualified human review the output before it goes anywhere. A well-structured brief with wrong information is worse than a rough brief with correct information. The goal isn't a beautiful document. The goal is a room full of people who are actually ready to make a decision. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Frequently asked questions What is an AI briefing document prompt? An AI briefing document prompt is a structured instruction you give to a tool like ChatGPT or Claude to help you write a briefing: a short document that prepares someone for a meeting, decision, call, or review. You provide the context and the ask, and the AI helps with structure, language, and surfacing gaps. The human still owns the facts and the judgment. Can AI write a briefing document from scratch? It can produce a template, but not a good brief. A real briefing requires context that only you have: what's actually true, what the reader already knows, what they'll resist, and what decision they actually need to make. AI can help you organize and articulate that context once you've given it the raw material. It can't supply the material itself. What should I never put into an AI tool when writing a briefing? Customer PII, employee records, financial forecasts, legal dispute details, security incidents, unreleased product information, confidential contracts, board materials, and sensitive client information should not go into AI tools your company hasn't explicitly approved for that data. When in doubt, anonymize or summarize before pasting. How do I make sure my AI-generated brief doesn't contain made-up facts? Tell the AI explicitly: "Do not add facts I haven't provided. If something is unclear or missing, flag it instead of filling the gap." Then check every specific claim, number, or statistic against your source materials before sending. The AI research prompts article has more on this. What's the difference between a briefing document and an executive summary? A briefing prepares someone for an upcoming event or decision. An executive summary recaps what happened or what a longer document says. They overlap, but briefings are forward-facing and usually include an explicit ask. If you need the exec summary version, the AI executive summary prompts cover that separately. How long should an AI-generated briefing document be? One page for most purposes. Half a page for executives. Two pages maximum if the topic genuinely requires it. If your briefing is longer than two pages, you're either including too much background, haven't clarified the ask, or need a different format entirely. Use the Prompt 9 pre-flight check to cut it down before it goes out. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Executive Summary Prompts: 10 Templates 10 copy-paste AI executive summary prompts for project updates, board briefs, risk summaries, and more. Includes the 5-part prompt formula. ai promptsexecutive summaryworkplace ai AI Policy Prompts: Write Rules People Follow 10 copy-paste AI policy prompts for HR and ops leads. Map AI use, set data rules, write disclosure policies, and build a rollout memo people actually read. ai policyworkplace aiai prompts AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation --- ## AI Budget Planning Prompts: Build a Better Plan URL: https://dontreplace.me/articles/ai-budget-planning-prompts-build-a-better-plan ai budget planningbudget promptsai at workfinancial planningchatgpt for work AI Budget Planning Prompts: Build a Better Plan 10 copy-paste AI budget planning prompts for scenarios, headcount requests, cost-cutting, and stakeholder narratives. With safety rules and a reusable formula. June 19, 2026 Updated June 19, 2026 Budget season arrives like a bad smell. Everyone's suddenly an expert on "strategic allocation" and "ROI frameworks," and somehow you're supposed to produce a coherent plan from a folder of last year's actuals, a spreadsheet someone built in 2019, and three competing priorities from three different managers. AI budget planning prompts won't fix the folder or the managers. But they will help you organize your thinking, structure your scenarios, and draft the narrative that gets a plan out of your head and onto paper before the deadline hits. Here's what they won't do: invent real numbers. That's on you. This article gives you 10 copy-paste prompts, a reusable formula, and the safety rules that keep you from turning a fast AI output into a confidently wrong financial document. The right way to think about AI for budget planning AI is a fast planning assistant. Not a finance department. Not an accountant. Not a CFO who's reviewed your contracts. What it's genuinely good at: organizing inputs you give it, generating scenario structures, flagging missing categories, drafting stakeholder narratives from bullet points, and building checklists. It'll do those things faster than you will, and it won't complain about the meeting that got canceled three times. What it will confidently get wrong: vendor pricing it doesn't know, ROI calculations based on assumptions it invented, tax treatment, accounting classifications, legal obligations, and anything that requires knowing your actual business. As what AI can and can't do explains plainly, it's fast pattern completion, not judgment. Rule #13 from Don't Replace Me applies here harder than almost anywhere else: garbage in, garbage out. Vague prompts produce confident budget fan fiction. Real numbers, real constraints, and real decisions stay yours. What never goes into an AI tool Before the prompts, the safety list. This is non-negotiable. Do not paste any of the following into an AI tool that hasn't been approved by your IT or security team: Payroll data or individual employee salaries Employee records or HR files Customer names, contact details, or any PII Bank account details or payment credentials Confidential vendor contracts or undisclosed pricing Board materials or pre-board drafts Unreleased financial results Acquisition or divestiture plans Sensitive competitive strategy documents Work with structures, categories, and anonymized examples. "Headcount cost for a team of four at average loaded cost of $X" is fine. Pasting the payroll register is not. The reusable AI budget planning prompt formula Every prompt in this article follows the same structure. Learn it once, adapt it forever. Role + Context you're providing + Specific output you need + Constraints or format requirements Example: "You are a planning assistant helping a marketing manager. I'll give you a list of current software subscriptions and estimated annual costs. Organize them into categories (essential, nice-to-have, redundant), flag any missing categories, and output a table. Do not invent pricing. Use only what I provide." That's it. Role, context, output, constraints. Every prompt below follows this pattern. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 copy-paste AI budget planning prompts Prompt 1: Clarify your budget goals before you start Use this before you open a spreadsheet. It forces clarity on what the plan actually needs to accomplish. You are a planning assistant. I'm building a budget for [department/team/project] for [time period]. My primary goals are: [list 2-4 goals]. My key constraints are: [headcount, spend ceiling, timeline, approvals needed]. Ask me 5 clarifying questions that would help me build a better plan. Focus on what's ambiguous, what I might be missing, and what tradeoffs I haven't addressed yet. Do not give me a budget. Just ask the questions. Prompt 2: Organize your current spend Paste in your categories and numbers (no employee names, no vendor contract details). Get a clean structure back. You are a planning assistant. Here is my current spend by category for [period]: [paste your list of categories and amounts]. Reorganize this into: fixed costs, variable costs, one-time costs, and optional costs. Flag any categories that look like they might be missing based on a typical [department type] budget. Output a table. Use only the numbers I've provided. Prompt 3: Find missing cost categories The budget you forgot to include is the one that comes back to haunt you in Q3. You are a planning assistant helping a [role] at a [company type/size] plan a [department] budget. Here are the categories I've included so far: [list categories]. Based on common [department] budget structures, what categories am I likely missing? List them with a one-sentence explanation of why each tends to get overlooked. Do not invent costs. Just identify categories I should verify with my team. Prompt 4: Build base, conservative, and aggressive scenarios Scenarios are the part of budget planning most people skip because they're tedious. This speeds it up. You are a planning assistant. I have a base budget of [total] for [purpose]. My key cost drivers are: [list 3-5 drivers with rough amounts]. Build me a three-scenario framework: - Base: as planned - Conservative: 15% cut, focusing on [priority area] - Aggressive: 20% increase, focused on [growth area] For each scenario, list: what changes, what gets cut or added, and what the main risk is. Use only the numbers and drivers I've given you. Flag any assumption you'd need me to verify. Prompt 5: Draft a headcount request Headcount requests live or die on narrative. This prompt builds the structure; you fill in the real numbers. You are a planning assistant helping me draft a headcount request. Here is the context: [describe the role, the need, and the business problem it solves]. Current team size: [number]. Estimated cost: [range or placeholder]. Timeline: [when needed]. Draft a one-page headcount justification that includes: the business case, what's currently not getting done, the cost of not hiring, and what success looks like in 12 months. Use a professional but direct tone. Flag any section where I need to add real data. Prompt 6: Compare software renewal options Before you auto-renew anything, run it through this. Pair it with real vendor quotes from procurement. For more depth on evaluating tools, the AI vendor evaluation prompts article covers the full framework. You are a planning assistant helping me evaluate a software renewal decision. Here are my options: [list options with costs, features, and contract terms you know from memory or public information only]. Compare them on: total annual cost, functionality match to my needs [describe needs], contract flexibility, and switching cost if we change. Output a comparison table. Note: I will verify all pricing with real vendor quotes before deciding. Flag anything that needs verification. Prompt 7: Prepare a cost-cutting plan When the number has to go down, this helps you cut with logic instead of just slashing the travel line. You are a planning assistant. My budget needs to be reduced by [amount or %]. My current spend categories are: [list]. My non-negotiable costs are: [list]. My primary goal for this period is: [one sentence]. Suggest a cost-cutting approach that: protects the non-negotiables, minimizes impact on [goal], and identifies the cuts most likely to be reversible if conditions change. Rank cuts from lowest to highest impact. Flag any cuts that might have hidden costs (e.g., contract penalties, rehiring cost, lost productivity). Prompt 8: Pressure-test your ROI assumptions This is the one that saves you from presenting a business case that falls apart in the room. For decisions with higher stakes, combine it with the AI decision-making prompts framework. You are a planning assistant helping me stress-test a business case. Here is my ROI argument: [describe investment, expected return, and timeline]. My key assumptions are: [list them]. Challenge these assumptions. For each one, ask: What has to be true for this to hold? What's the most likely way this fails? What's a more conservative version of this estimate? Do not produce a revised ROI figure. Just help me see where my reasoning is weakest before I present it. Prompt 9: Write a stakeholder budget narrative The spreadsheet doesn't sell the plan. The narrative does. This turns your numbers into a readable document. You are a planning assistant helping me write a stakeholder summary for a budget plan. Audience: [describe: executive team, board, department head, etc.]. Key numbers: [total budget, major categories, year-over-year change]. Strategic context: [one paragraph on what this budget is trying to achieve]. Write a 3-paragraph executive summary that explains: what we're investing in and why, what we're not doing and why, and what we need from the audience (approval, input, decision). Tone: direct and professional. Flag anywhere I need to add specifics. Prompt 10: Create a finance review checklist Use this before you send anything to finance, procurement, or leadership. It won't replace their review. It'll make sure you're not the one who forgot something obvious. You are a planning assistant. I'm about to submit a [department] budget for [period] for finance review. Budget total: [amount]. Key components: [list]. Generate a pre-submission checklist covering: data completeness, documented assumptions, named cost owners, source-of-truth references, required approvals, compliance flags I should check with finance, and anything else that commonly causes budget revisions at this stage. Format as a checklist I can work through before submission. What AI can't decide for you The prompts above are for organizing thinking and drafting documents. They're not the plan. Your plan needs real things that no prompt produces. Real vendor quotes, not AI guesses. Cash timing, not just annual totals. Renewal dates in a calendar, not just a category list. Total cost of ownership including implementation, training, and the person who has to manage the thing. Rollback plans for decisions that might not work. Finance sign-off. Procurement review where contracts are involved. The AI risk assessment prompts article covers how to structure the risk side of any plan. And if you're trying to understand where your time and money is actually going before you plan, the AI workflow audit prompts are worth running first. AI is fast at structure and narrative. It's useless at knowing your vendor relationships, your internal politics, your cash position, or which number your CFO is actually going to push back on. The human value in budget planning is knowing which tradeoffs matter. A 15% headcount cut that saves money on paper but destroys delivery capacity is a fake saving. AI won't tell you that. You know it. That's the judgment that doesn't show up in a prompt. A word on accountability Whatever comes out of these prompts goes through you before it goes anywhere. You're the named owner. You verify the numbers. You get the real quotes. You check the contract terms with legal or procurement. You confirm the accounting treatment with finance. If someone asks "where did this number come from," the answer should be "I calculated it from our actuals" or "I got a quote from the vendor." Not "the AI suggested it." How to use AI at work covers the broader principle: AI as a fast assistant that handles the tedious structural work so you can spend your time on the judgment calls that actually matter. Budget planning has a lot of tedious structural work. These prompts handle that part. You handle the rest. Frequently asked questions Can AI actually help with budget planning? Yes, for the structural and narrative parts. AI is good at organizing categories, drafting scenarios, flagging missing items, and writing stakeholder summaries from your bullet points. It can't produce accurate vendor pricing, tax advice, accounting treatment, or any number that requires knowing your actual business situation. What information should I never put into an AI tool for budget planning? Never paste payroll data, employee records, customer PII, bank details, confidential vendor contracts, board materials, unreleased financials, acquisition plans, or sensitive competitive strategy. Work with anonymized categories and amounts. Check which tools your company has approved before using any AI for finance-adjacent work. How do I make AI budget prompts more accurate? Give it real context: your actual categories, real constraints, your actual goals. The more specific your input, the less the AI has to guess. Always flag outputs that need verification before treating them as real numbers. The rule is: you provide the facts, AI provides the structure. Will AI replace finance teams or budget analysts? No. What AI can do is handle the formatting, scenario-building, and narrative drafting that takes up time but doesn't require financial judgment. The judgment about which cuts are real savings, what the risk profile of an investment is, and what the cash timing means for the business stays with people who understand the actual business. See what jobs AI can't replace for more on where human judgment holds. What's the biggest mistake people make using AI for budgeting? Treating AI output as a finished plan. The prompts above produce drafts and structures. Every number needs a source. Every assumption needs an owner. Every approval still goes through the right channel. A polished AI-generated budget document that nobody verified is worse than a messy spreadsheet with real numbers, because it looks credible. Do I need special AI tools for budget planning? No. ChatGPT or Claude handle everything in this article. The prompts work in either. You don't need a specialized finance AI product. If your company has Microsoft Copilot integrated with your tools, that works too, and may have better data access controls for sensitive financial work. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Automation Prompts: Automate the Boring Stuff 10 copy-paste AI automation prompts for mapping workflows, writing SOPs, adding approval gates, and QA testing. Practical templates with safety rules included. ai automation promptsworkflow automationSOP templates AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts --- ## AI Bug Report Prompts: 10 Templates That Work URL: https://dontreplace.me/articles/ai-bug-report-prompts-fix-weird-glitches-fast ai bug report promptsbug reportingQAsoftware testingChatGPT at work AI Bug Report Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn vague bug complaints into structured tickets. Templates for triage, handoffs, regression checklists, and more. July 23, 2026 Updated July 23, 2026 Someone files a bug report that says "the thing doesn't work." You're a QA analyst, a product manager, a developer on call, or just the person who happened to be nearest when the wheel fell off. The ticket sits in the backlog, vague and untouchable, because nobody knows what "the thing" is or when it stopped working or whether it affects one user or ten thousand. AI bug report prompts can fix this. Not by thinking for you, but by forcing structure into the chaos before it lands on an engineer's desk. This is a copy-paste template article. Ten prompts, each doing one specific job. Use them with ChatGPT, Claude, or whatever tool your company hasn't blocked yet. But read the safety rules first, because there's a short list of things you must never paste into any of these. Before you touch a prompt: what not to paste into AI This is non-negotiable. Before you copy anything into an AI tool, strip out: Customer names, emails, or any personally identifiable information (PII) Passwords, API keys, access tokens, or session data Raw production logs that contain user data Security vulnerabilities or anything touching authentication Proprietary code or internal system architecture details Legal disputes, HR issues, or compliance-related correspondence Medical records, financial data, or any regulated information Private client conversations or confidential contracts Redact first. Paste second. If you're not sure whether your company has approved the AI tool you're using, check before you use it for anything work-related. The five-minute explainer on what AI can and can't do covers why this matters more than most people assume. One practical system: keep a text file called "redaction template" open whenever you're doing bug triage. Every time you copy something to paste into AI, you paste it into the template first, swap real data for placeholders like [USER_EMAIL] or [ACCOUNT_ID], then copy the clean version. Takes about 30 seconds. Saves a potential data incident. The reusable AI bug report formula Every prompt in this article follows the same structure. You give AI the raw mess, you tell it the format you need, and you verify everything it produces before it goes anywhere near a ticket. The formula: [Raw input] + [Desired output format] + [Constraint: verified evidence only, flag uncertainty] That last constraint is the one people skip. AI will confidently invent plausible-sounding reproduction steps, fill in missing environment details with guesses, and write severity rationale as if it has access to your error logs. It doesn't. It's pattern-matching against every bug report it's ever trained on. Useful for structure. Dangerous when mistaken for diagnosis. Every prompt below includes an instruction to flag uncertainty rather than fill it in. Keep that line in. Remove it and you're asking the tool to make things up. This also means the prompts are deliberately conservative. A ticket with five honest "UNKNOWN" fields is better than a ticket with five confidently wrong answers. Engineers have limited time. Sending them after a ghost because the AI guessed wrong about the environment costs more than the five minutes it takes to chase down the real data. AI bug report prompts: 10 templates you can use right now Prompt 1: Turn a vague complaint into a ticket Use this when someone files a report that's basically a cry for help with no useful details. Prompt: "I received this complaint: [paste complaint, redacted]. Turn it into a structured bug ticket with: title, summary, steps to reproduce (mark any as 'UNCONFIRMED' if not stated), expected behavior, actual behavior, environment (mark as 'UNKNOWN' if not stated), and severity (mark as 'TBD pending investigation'). Do not invent details. Flag every gap." What you get back is a skeleton ticket with honest gaps labeled. That's the point. An engineer looking at a ticket with "UNCONFIRMED" and "UNKNOWN" fields knows exactly what to investigate first. This prompt is especially useful when a complaint arrives through a non-technical channel: a customer email forwarded by sales, a Slack message from a client-facing team, or a voice note someone transcribed. The rawer the input, the more useful this is. Prompt 2: Extract reproduction steps from a messy conversation Support threads, Slack messages, and customer emails are not bug reports. They're stories. This prompt pulls the procedural steps out. Prompt: "Here is a conversation about a bug: [paste thread, fully redacted of PII]. Extract the reproduction steps in numbered order. Use only what is explicitly stated. For any steps that seem implied but aren't confirmed, write them in italics with '(INFERRED)'. Do not add steps from general knowledge." The "(INFERRED)" flag is doing real work here. If step three is inferred and the engineer can't reproduce the bug, they know exactly where to focus their follow-up questions instead of retrying the whole sequence from scratch. Long support threads are the ideal input for this prompt. A 40-message Slack thread that starts with "something's broken" and ends with "oh wait it only happens on Safari" contains reproduction steps. They're just buried in three days of back-and-forth. This prompt finds them. Prompt 3: Separate expected vs. actual behavior This is the single most important part of any bug report and the most often missing. Engineers need to know what should have happened, not just what did. Prompt: "Based on this description: [paste description, redacted], write two sections. Section 1: Expected Behavior, what the user or specification says should happen. Section 2: Actual Behavior, what was observed. Keep them strictly separate. If the expected behavior isn't stated, write 'Expected behavior not documented, requires product owner input.'" That last instruction saves hours of back-and-forth. If the expected behavior genuinely isn't documented anywhere, that's a product problem, not just a bug. It's worth noting how often this distinction exposes a disagreement rather than a defect. Sometimes what looks like a bug is actually a feature behaving as designed, but the design was wrong. The moment you write down "expected: X, actual: Y," it becomes a conversation about whether X was the right expectation in the first place. That conversation is much easier to have early, before engineering time is spent fixing the wrong thing. Prompt 4: Summarize customer impact Useful when you need to escalate something or prioritize a ticket against twenty others. You need a clear, honest statement of who is affected and how badly. Prompt: "Based on this information: [paste details, redacted], write a two-paragraph customer impact summary. Paragraph 1: who is affected and what they can't do. Paragraph 2: any available data on frequency or scale (if no data is available, state that explicitly). Do not estimate numbers you don't have." This feeds directly into severity decisions and stakeholder communication. If you've got related work in flight, the AI stakeholder update prompts can help you translate this impact summary into an exec-friendly update. The "do not estimate numbers" instruction matters more than it looks. A casually invented "roughly 15% of users" in a customer impact summary can take on a life of its own. It ends up in a Slack escalation, then a leadership meeting, then a post-incident review. If you don't have the number, write that you don't have the number. Prompt 5: Collect environment details Half the bug reports that die in triage die because nobody knows what version, what browser, what operating system, or what account type the bug appeared on. This prompt generates a checklist. Prompt: "Generate an environment details checklist for this bug: [paste brief description, redacted]. Include: OS and version, browser or app version, device type, account type or user role, geographic region if relevant, network conditions if relevant, third-party integrations if relevant. Mark each field as CONFIRMED, UNCONFIRMED, or NOT APPLICABLE. Leave blanks where data hasn't been provided." Use this to chase down missing information before the ticket moves to development. A half-empty checklist is still better than no checklist. Send the checklist back to whoever filed the report with a request to fill in the blanks. Most people will. They just didn't know what to include the first time. Prompt 6: Clean up screenshot and log notes You've got a screenshot with post-it notes, a pasted error message, and three lines of log output. None of it is formatted. This prompt turns it into something an engineer can read without squinting. Prompt: "Here are my raw notes from a bug investigation: [paste notes, fully redacted of any credentials, user data, or sensitive information]. Format them as: (1) Visual evidence summary, what screenshots show, (2) Error messages, exact text only, (3) Log snippets, formatted as code blocks, (4) What this evidence confirms, (5) What is still unverified. Do not interpret the logs technically. Just organize what I've given you." Emphasis on "do not interpret technically." AI will happily tell you what a 502 error "probably means" for your specific infrastructure. It doesn't know your infrastructure. Keep the interpretation human. This prompt is particularly useful for QA analysts who aren't developers. You know what you observed. You've got the screenshots. You have the error text. What you might not have is confidence that your write-up reads clearly to an engineer. This prompt handles the formatting so you can focus on the accuracy. Prompt 7: Draft severity and priority rationale This is where bug reports get political. Everyone thinks their bug is P1. This prompt helps you write a defensible rationale based on actual criteria rather than whoever shouted loudest. Prompt: "Using these severity definitions: [paste your team's severity scale], and this bug information: [paste ticket details, redacted], write a severity and priority rationale. State the severity level, the primary reason for that level, and any factors that could upgrade or downgrade severity. Flag if any information needed to confirm severity is missing." This only works if you use your team's actual severity definitions, not made-up ones. If your team doesn't have written severity definitions, that's a separate problem and fixing it is worth an afternoon. A documented rationale also helps when a stakeholder pushes back on priority. Instead of a subjective argument, you have a written analysis tied to agreed criteria. "This is P2 because it affects less than 1% of users and a workaround exists, but it would become P1 if we confirm it's happening on enterprise accounts" is a much more useful conversation than "trust me, it's not that bad." Prompt 8: Generate edge-case retest ideas Once a bug is "fixed," the obvious retest is obvious. The one that bites you is the edge case nobody thought to test. This prompt generates candidates. Prompt: "This bug was reported: [paste summary, redacted]. The proposed fix is: [paste fix description]. Generate 10 edge cases or boundary conditions to retest after the fix is applied. Focus on: different user roles, different data states, concurrent usage, partial data entry, session edge cases, and integration touchpoints. Label each as HIGH/MEDIUM/LOW priority to retest." This feeds directly into regression planning. If you're building a full QA checklist around the fix, the AI QA checklist prompts handle the broader version. One pattern worth knowing: the most dangerous edge cases after a bug fix are usually the ones adjacent to the original failure. If the bug happened on mobile Safari, test desktop Safari too. If it happened for admin users, check read-only users. AI is genuinely good at generating these adjacencies because it's seen enough bug reports to know where fixes tend to create new problems. Prompt 9: Write an engineer-friendly handoff The ticket is written. The severity is set. Now it needs to go to an engineer without a thirty-minute meeting to explain what you already know. This prompt drafts the handoff note. Prompt: "Write a brief engineer handoff note for this bug: [paste ticket summary, redacted]. Include: what we know for certain, what we suspect but haven't confirmed, what we tried in investigation, what information is still needed from engineering, and what the first suggested investigation path is. Keep it under 200 words. Do not suggest a root cause." "Do not suggest a root cause" is doing important work. AI will suggest one. It will sound plausible. Engineers will spend time ruling it out. Let them find the root cause. The "what we tried in investigation" section earns goodwill with engineers fast. Nothing is more frustrating than spending twenty minutes reproducing a bug that QA already confirmed is 100% reproducible on Chrome 124, not on Firefox. Document what you already ruled out. It's not wasted effort, it's a gift. Prompt 10: Create a regression checklist after the fix The fix is deployed. Now you need to make sure you didn't break anything else. This prompt generates a regression checklist tied to the specific bug. Prompt: "A bug was fixed: [paste bug summary, redacted]. The change affected: [paste affected components or code areas]. Generate a regression checklist covering: the original bug (verify it's fixed), directly related functionality, integration points with other features, and any edge cases that were identified during the bug report. Format as checkboxes. Mark any item as 'OPTIONAL' if it's lower risk given the change scope." Pair this with AI release notes prompts when the fix ships to users. Customers and support teams shouldn't find out about a fix from a cryptic changelog entry. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → How to build these prompts into your actual workflow The prompts are useful in isolation. They're more useful as a sequence tied to your bug lifecycle. Here's a simple mapping: Bug reported: Prompt 1 (vague to structured), Prompt 5 (environment checklist) Bug in triage: Prompt 3 (expected vs. actual), Prompt 4 (customer impact), Prompt 7 (severity rationale) Bug in investigation: Prompt 2 (extract reproduction steps), Prompt 6 (clean up evidence) Fix proposed: Prompt 8 (edge case retest ideas) Fix ready for handoff: Prompt 9 (engineer handoff note) Fix deployed: Prompt 10 (regression checklist) Not every bug needs all ten. A minor UI glitch with clear reproduction steps doesn't need a customer impact summary. A payment processing failure needs everything. Match the prompt to the stakes. The bigger discipline is consistency. When the same person writes bug reports differently every week, triage slows down. When a team uses the same structure on every ticket, engineers build pattern recognition and move faster. Prompts are how you get that consistency without making it someone's job to enforce a style guide. What these prompts don't do They don't diagnose the bug. They don't read your logs. They don't understand your architecture, your user base, your SLAs, or your release calendar. They produce structure, not answers. The human value here, as Rule #7 in Don't Replace Me puts it, is knowing which failures matter. A polished ticket with a wrong severity wastes engineering time. A perfectly formatted handoff with invented reproduction steps sends engineers on a ghost hunt. AI makes things look done. That's not the same as done. Final sign-off on every ticket is yours. Always. The practical guide to using AI at work covers this philosophy more broadly if you want the full picture. Frequently asked questions Can AI write bug reports without human review? No. AI can structure and format bug information you provide, but it can't verify reproduction steps, confirm environment details, or diagnose root causes. Every AI-generated ticket needs human review before it goes to development. Missing or invented details in a bug report waste engineering time and delay fixes. Is it safe to paste error logs into ChatGPT or Claude? Only if the logs are fully scrubbed of user data, credentials, and anything proprietary. Raw production logs almost always contain PII or sensitive system information. Strip everything personal and sensitive before pasting. If you're unsure whether your company allows use of the AI tool in question, check first. What's the most common mistake in AI bug reporting? Treating the AI output as verified information. AI will fill gaps with plausible-sounding details if you don't explicitly tell it not to. Every prompt in this article includes an instruction to flag uncertainty rather than fill it in. Keep that instruction in every time. How do I handle bugs that involve security vulnerabilities? Don't use AI tools for these at all. Security vulnerabilities need to go through your organization's established security disclosure process. Pasting anything touching authentication, access control, or data exposure into a third-party AI tool is a serious risk. How do these bug report prompts relate to acceptance criteria? They're complementary. Acceptance criteria define what "done" looks like before development. Bug reports capture what went wrong after something ships. If you're finding bugs that point back to unclear acceptance criteria, the AI acceptance criteria prompts help close that loop upstream. What if my team doesn't have standard severity definitions? Then your severity is whatever the most senior person in the room says it is on any given day, which is a great way to have permanent firefighting culture. Write a simple four-level scale (Critical/High/Medium/Low) with one-sentence definitions for each. It takes an hour. Prompt 7 works much better with it. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance AI API Testing Prompts: Catch Broken Integrations AI API testing prompts for contracts, edge cases, authorization, retries, error responses, and human-verified integration evidence. AI promptsAPI testingquality assurance --- ## AI Change Impact Analysis Prompts: Spot Damage URL: https://dontreplace.me/articles/ai-change-impact-analysis-prompts-spot-downstream-damage AI promptschange impactriskQAoperations AI Change Impact Analysis Prompts: Spot Damage AI change impact analysis prompts for mapping affected systems, teams, customers, QA checks, support gaps, and rollout risks. August 05, 2026 Updated August 05, 2026 Change impact analysis is the part of work where someone says, “This is a small change,” and then three teams, two dashboards, one support macro, and a customer contract quietly burst into flames. Small changes are often small only from the chair of the person requesting them. A field name changes. A pricing rule moves. A workflow gets “simplified.” A migration cuts over. A policy gets rewritten. A release removes a weird old behavior that only one customer used, naturally the customer worth the most money. Everyone agrees the change is reasonable. Nobody agrees who checked the downstream damage. That is where AI change impact analysis prompts help. Not because AI can inspect your production systems, know your customers, approve risk, read the minds of support agents, or magically understand the weird spreadsheet Karen built in 2019 that now runs half the company. It cannot. But it can turn scattered change notes, release descriptions, workflow diagrams, affected-team lists, dependency notes, customer-impact concerns, risk registers, rollout plans, support tickets, training gaps, and acceptance criteria into a cleaner impact analysis. If you are already deciding whether to launch, use AI go/no-go decision prompts. If you are still figuring out what the change touches, stay here. AI can structure change impact analysis. A human still owns the evidence, judgment, approvals, customer promises, risk acceptance, and final sign-off. What is change impact analysis? Change impact analysis is the process of understanding what a proposed change affects before the change ships. A useful change impact analysis answers: What exactly is changing? What systems, workflows, reports, documents, customers, teams, and vendors are affected? What happens before the change, and what happens after it? What dependencies might break? What support, training, documentation, and customer communication need to change? What QA, regression, monitoring, and rollback checks are required? Who approves the risk? What evidence proves the analysis is not just vibes in a trench coat? A bad impact analysis says, “Should be fine.” That phrase has caused more damage than several minor natural disasters. AI change impact analysis prompts are useful because AI is good at structure. It can make checklists, compare current and future workflows, group affected audiences, draft stakeholder questions, turn messy notes into an impact brief, and flag missing evidence. But it follows the same rule as what AI can and can't do: AI can make the blast radius easier to inspect. It cannot prove the blast radius is safe. The reusable AI change impact analysis prompt formula Use this formula whenever you ask AI to help analyze a change: “You are a change impact analysis assistant. I am [your role] reviewing [product change, process change, workflow update, migration, policy change, integration release, pricing update, system cutover, client delivery, or internal rollout]. Here is privacy-safe context: [change summary, current workflow, future workflow, affected systems, affected teams, customer groups, dependency notes, rollout plan, acceptance criteria, risk notes, support concerns, training needs, documentation links, open questions]. Create [specific output]. Separate confirmed facts from assumptions. Flag missing evidence, likely downstream dependencies, customer-impact risks, support and training gaps, QA checks, rollback concerns, approval needs, and owner questions. Do not invent facts, owners, system behavior, customer commitments, security conclusions, legal conclusions, due dates, approvals, or final risk decisions.” That last sentence is not legal seasoning. AI likes to complete patterns. In impact analysis, a completed pattern can become a fake dependency, a fake owner, or the world’s most confident wrong “low risk” label. Before using any prompt, sanitize the input. Do not paste customer PII, credentials, access tokens, raw production logs, private tickets, unreleased strategy, legal disputes, security vulnerabilities, financial records, HR issues, regulated data, medical information, private client conversations, or sensitive personal information into unapproved AI tools. Use summaries, categories, redacted examples, approved screenshots, internal ticket IDs, and links to authorized systems instead. The model does not need a customer’s phone number to help you realize your support docs are stale. What to bring before asking AI for impact analysis AI cannot map a change it cannot see. If you feed it one sentence and a dream, it will give you a polished dream with bullets. Bring this first: InputWhy it mattersHuman check Change summaryDefines the thing being analyzedConfirm scope with the requester Current workflowShows what exists nowVerify against reality, not old docs Future workflowShows what will be differentMark proposed vs. approved clearly Affected systemsExposes technical dependenciesConfirm with system owners Affected teamsExposes operating dependenciesInclude support, sales, finance, legal, and ops when relevant Customer impactPrevents internal-only thinkingUse approved, privacy-safe summaries Rollout planShows timing and exposureCheck pilots, phases, cutover windows, and fallback paths QA criteriaDefines what must be testedInclude regression and edge cases Documentation and training needsPrevents support chaosCheck help docs, SOPs, scripts, and enablement Approval trailMakes risk ownership visibleCapture who signs off and why If you do not have these inputs, ask AI to create a missing-information checklist first. Do not ask it to analyze impact from “we are changing the onboarding flow” unless you enjoy receiving beautifully formatted fog. For adjacent work, pair this with AI risk assessment prompts and AI operational readiness prompts. Risk tells you what can hurt. Readiness tells you whether anyone is actually prepared. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI change impact analysis prompts you can use today Prompt 1: Turn a change summary into an impact brief Use this when the change is described in ten scattered places and none of them agree. “Act as a change impact analysis assistant. Using the privacy-safe notes below, create an impact brief with these sections: change summary, reason for change, current state, future state, affected systems, affected teams, customer impact, operational impact, support impact, documentation impact, QA checks, rollout risks, rollback concerns, open questions, and approval needs. Clearly label assumptions. Do not invent facts, owners, dependencies, customer commitments, or approvals. Notes: [paste sanitized notes].” This prompt gives you a first draft that humans can argue with productively. That is the goal. A good impact brief should create useful arguments before launch, not Slack archaeology after launch. Prompt 2: Map affected systems and teams Use this when everyone assumes someone else checked the dependency chain. “Review this proposed change: [paste sanitized change summary]. Create a table of potentially affected systems, workflows, teams, reports, documents, customer touchpoints, vendors, and automation. For each item, list why it might be affected, what evidence confirms or disproves the impact, who should verify it, and what question we need to ask next. Separate confirmed impact from possible impact.” AI is good at generating the map. Humans still verify the territory. If the model says billing might be affected, do not panic. Ask billing. The whole point is to find questions before users find consequences. Prompt 3: Compare current vs. future workflow Use this when a process change sounds simple but nobody has drawn the before-and-after. “Create a current-state vs. future-state comparison for this workflow change. Current workflow: [paste sanitized current workflow]. Proposed workflow: [paste sanitized future workflow]. Build a table with step, current owner, future owner, system used, data required, customer-visible impact, support impact, failure mode, and verification check. Flag any step where ownership, data source, approval, or customer expectation is unclear.” Workflow changes fail in the seams. The handoff moves. The permission changes. The report no longer updates. The customer receives a different message. The internal team says, “Wait, were we supposed to know?” This prompt makes the seams visible. Prompt 4: Find hidden dependencies Use this when the official dependency list looks suspiciously short. “Act as a skeptical dependency reviewer. Given this change: [paste sanitized change details], brainstorm likely hidden dependencies across systems, data fields, API behavior, permissions, reports, dashboards, automations, support scripts, SOPs, customer communications, training materials, analytics, billing, compliance, and vendor processes. For each dependency, list the evidence needed to verify it and the owner who should confirm it. Do not claim a dependency exists without evidence.” This is the prompt for finding the boring landmines. Dashboards. Zapier workflows. Revenue reports. Email templates. Permissions. That one CSV import no one admits is production. If the change touches release timing, combine this with AI cutover plan prompts. Cutovers are dependency festivals with snacks. Prompt 5: Check customer and support impact Use this when internal teams are excited and support has not been invited to the party. Again. “Analyze customer and support impact for this change: [paste sanitized change summary]. Create sections for customer-visible behavior, customer groups affected, possible confusion, support ticket drivers, help-center updates, macro/script updates, customer success talking points, escalation paths, and monitoring signals. Flag any place where we need approved customer language, legal review, security review, or account-team review.” A change can be technically correct and operationally rude. Customers do not experience your architecture. They experience changed buttons, missing fields, different invoices, surprise emails, broken habits, and support agents saying, “Interesting, I have not heard about that.” Do not do that to support. They have suffered enough. Prompt 6: Draft stakeholder review questions Use this when you need better questions before a review meeting. “Prepare stakeholder review questions for this proposed change: [paste sanitized context]. Group questions by product, engineering, QA, operations, support, customer success, sales, finance, legal/compliance, security/privacy, documentation, and leadership. For each question, explain what risk it reduces and what evidence would answer it. Keep questions practical and specific. Do not ask for sensitive data to be pasted into the AI tool.” The quality of an impact review depends on the quality of the questions. “Any concerns?” is not a review. It is a trapdoor. People remember concerns after the meeting, usually while brushing their teeth. Specific questions get specific answers. Prompt 7: Create QA and regression checks Use this when QA needs more than “test the change.” “Create a QA and regression checklist for this change: [paste sanitized change summary, affected workflow, and acceptance criteria]. Include happy-path tests, negative tests, boundary cases, permission checks, data validation, customer-visible behavior, reporting/dashboard checks, integration checks, support workflow checks, monitoring checks, and rollback verification. Format as a table with test area, scenario, steps, expected result, evidence to capture, owner, and priority.” Good QA is impact analysis with receipts. It does not only ask whether the new thing works. It asks what old thing might now be haunted. For more testing structure, use AI QA checklist prompts. For post-release watch plans, use AI post-launch monitoring prompts. Prompt 8: Plan training and documentation updates Use this when the change affects people who are expected to magically know things. “Review this change and identify training and documentation updates needed. Context: [paste sanitized change summary and affected audiences]. Create a table with audience, what changes for them, document or training asset affected, exact update needed, owner, deadline, review approver, and risk if not updated. Include help-center docs, internal SOPs, onboarding material, sales enablement, support macros, customer emails, and manager talking points where relevant.” Documentation is part of the product whether teams admit it or not. If the docs stay wrong, the change is not done. It is merely released into a fog machine. For handoff-heavy changes, pair this with AI support handoff prompts. Nobody enjoys inheriting a launch with no context except “ping Alex.” Prompt 9: Write a go/no-go impact summary Use this when leadership needs the short version without losing the important caveats. “Turn the impact analysis below into a go/no-go impact summary. Include: change overview, confirmed impacts, unresolved impacts, high-risk dependencies, customer/support readiness, QA status, rollback readiness, approval gaps, recommended decision, and conditions for proceeding. Use plain language. Do not hide unknowns. Do not claim approval or low risk unless the evidence supports it. Impact notes: [paste sanitized analysis].” This prompt helps prevent executive summaries from becoming executive sedatives. The goal is clarity, not comfort. If the change is launch-related, this belongs next to AI rollout plan prompts and the go/no-go guide linked earlier. Prompt 10: Create a decision log after review Use this when the meeting happened and now everyone’s memory is already rewriting history. “Create a change impact decision log from these review notes: [paste sanitized notes]. Include decision made, decision owner, date, evidence reviewed, confirmed impacts, accepted risks, mitigation actions, unresolved questions, follow-up owners, due dates, approval needs, and where evidence should be stored. Separate decisions from discussion. Flag any accepted risk that lacks a named owner or approval record.” Decision logs are boring until you need one. Then they become a miracle. A good log prevents “I thought we agreed” from becoming the team’s operating system. It also gives future-you a fighting chance when the change creates a weird side effect six weeks later. Common change impact analysis mistakes The first mistake is analyzing only the technical change. Systems do not break alone. They break workflows, expectations, reports, support promises, billing rules, customer habits, training material, and internal rituals with names like “the Friday export.” The second mistake is treating unverified AI output as evidence. AI can say a dependency is likely. That is not the same as confirming it. The sentence “billing may be affected” should trigger a billing review, not become a line item in the final truth ledger. The third mistake is leaving support and customer success out until the end. They will be the first people customers yell at. Invite them before the yelling. The fourth mistake is hiding unknowns to make the summary look clean. Unknowns are not embarrassment. Unknowns are work. Put them in the open where they can be handled by adults, or at least by adults with calendars. The fifth mistake is forgetting rollback. If the change cannot be reversed, paused, isolated, or mitigated, say that clearly. “No rollback” is a decision. It should have an owner. When not to use AI for change impact analysis Do not use public or unapproved AI tools for sensitive material. That includes customer PII, credentials, access tokens, raw logs, private contracts, security vulnerabilities, legal disputes, HR issues, financial records, medical data, regulated data, private client conversations, and anything your compliance team would describe with forehead veins. Do not ask AI to approve risk. It cannot. Do not ask AI to decide whether customers should be notified. It can draft options, but legal, customer success, product, leadership, and sometimes security need to decide. Do not ask AI to replace domain experts. The model does not know your production weirdness. It does not know which report the CFO quietly checks every Monday. It does not know that one enterprise customer has a special workflow because of a contract clause from the Bronze Age. Use AI to make the review sharper. Do not use it to skip the review. Frequently asked questions Can AI do change impact analysis by itself? No. AI can organize change notes, generate questions, map possible dependencies, and draft checklists. It cannot verify your systems, inspect customer behavior, approve risk, or know which internal workaround is secretly critical. Treat it as a facilitator, not the change advisory board wearing a hoodie. What should I paste into an AI change impact prompt? Paste privacy-safe summaries: change descriptions, current and future workflow notes, affected teams, redacted customer-impact themes, acceptance criteria, rollout steps, dependency notes, and open questions. Do not paste raw production logs, credentials, customer PII, private tickets, contracts, security vulnerabilities, legal issues, HR issues, or regulated data into unapproved tools. What is the best first prompt for a messy change? Start with Prompt 1. Ask AI to turn messy notes into an impact brief and label assumptions. Then have humans verify each section. If the draft contains too many assumptions, that is not failure. That is the map telling you where the fog lives. How is change impact analysis different from risk assessment? Impact analysis asks, “What does this change touch?” Risk assessment asks, “What could go wrong, how bad would it be, and what should we do about it?” They overlap, but they are not identical. Use impact analysis first when the blast radius is unclear, then use risk assessment to prioritize what deserves mitigation. How do I stop AI from inventing dependencies? Tell it to separate confirmed facts from possible impacts, require evidence for every dependency, and list who should verify each item. Never copy an AI dependency map directly into a final plan. Use it as a question generator. The robot can suggest doors to check. It cannot tell you which doors are real. Who should review an AI-assisted impact analysis? At minimum, the change owner, system owner, QA owner, support or customer success representative, and whoever owns the affected workflow. For higher-risk changes, add security, legal/compliance, finance, data, leadership, or account teams. The right reviewers depend on the blast radius, which is exactly why you are doing the analysis. Should I mention AI in the final change impact document? Usually, mention it only if your company policy requires it or if the document includes AI-assisted drafting notes. The important thing is not whether AI helped format the table. The important thing is whether humans verified the evidence, reviewed the risks, approved the decision, and kept an audit trail. The bottom line AI change impact analysis prompts are useful because most teams do not need more confidence. They need a better way to find the stuff they forgot. Use AI to structure the mess, surface missing questions, draft checklists, compare workflows, and make downstream impact visible. Then bring the humans back in to verify facts, judge risk, talk to customers, update support, approve rollout, and own the decision. The machine can help you see the blast radius. It cannot stand inside it for you. That is one of the core ideas in Don’t Replace Me by Dmitry Kargaev: stay useful by doing the human part AI cannot own. Judgment. Taste. Accountability. Knowing when a tidy matrix is helpful, and when it is just a very organized way to ignore the obvious. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Dependency Mapping Prompts: Find What Breaks AI dependency mapping prompts for finding upstream work, downstream risks, hidden handoffs, owners, schedule conflicts, and launch blockers. AI promptsdependency mappingproject management AI Remediation Plan Prompts: Fix Root Causes AI remediation plan prompts for turning findings, defects, incidents, owners, deadlines, fixes, and retest evidence into real corrective action. AI promptsremediationcorrective action AI Root Cause Analysis Prompts: Stop Fixing Symptoms AI root cause analysis prompts for turning messy incidents, bugs, timelines, evidence, and corrective actions into clearer RCA work. AI promptsroot cause analysisRCA --- ## AI Change Management Prompts: Get Buy-In Fast URL: https://dontreplace.me/articles/ai-change-management-prompts-get-buy-in-fast ai change managementchange management promptsai adoptionorganizational changeai at work AI Change Management Prompts: Get Buy-In Fast 10 copy-paste AI prompts for change management: announcements, stakeholder maps, resistance diagnosis, training, adoption metrics, and rollback plans. June 17, 2026 Updated June 17, 2026 Most AI rollouts fail for the same reason. Not the technology. Not the budget. The people. Your change management prompts need to account for the part no vendor demo covers: the team lead who's been doing this for 12 years and thinks the new tool is a personal insult, the frontline worker who's scared it means headcount cuts, the manager who nodded in the all-hands and then quietly told their team to ignore it. AI can help you draft your way through this. It cannot replace the judgment that knows which of those three people will tank the whole rollout if you get the messaging wrong. These 10 prompts give you real starting material. You still have to edit for your actual situation, run the outputs past the people who know your org, and keep anything sensitive out of tools that haven't been approved by your IT or security team. The one rule before you start Don't paste private data into AI tools that haven't been cleared for it. That means no employee names, performance records, private survey responses, customer PII, contract terms, legal disputes, confidential headcount plans, unreleased product details, security architecture, credentials, or proprietary workflows. If you're not sure whether something is safe to paste, don't paste it. Use placeholders like [ROLE] or [DEPARTMENT] instead of real names. This isn't paranoia. It's how you don't create a data incident on top of a change management headache. Now, the prompts. The reusable formula behind every AI change management prompt Before the templates, here's the pattern they're all built on. Good prompts for this kind of work have four parts. Role + Context + Task + Constraints "You are a change management consultant working with a [INDUSTRY] organization. We are rolling out [TOOL/CHANGE] to [AUDIENCE]. [SPECIFIC TASK]. Do not invent policy, legal requirements, compliance approvals, or specific productivity numbers. Flag anything that requires HR, legal, or security review." That last line matters more than people think. AI is very good at inventing authoritative-sounding details. A change plan that includes fake compliance citations or implied HR promises is worse than no plan at all. Prompt 1: Define the change clearly Vague change plans create vague adoption. Start by getting the definition tight. You are a change management consultant. We are introducing [TOOL NAME] to [TEAM/DEPARTMENT] at a [INDUSTRY] organization. The primary goal is [GOAL]. The main tasks it will change or replace are [LIST 2-3 TASKS]. The intended outcome is [OUTCOME]. Write a one-page change definition document covering: what is changing, what is not changing, who is affected, and what success looks like in 90 days. Do not make assumptions about headcount, performance review implications, or policy changes. Flag any section that needs review from HR, legal, or IT security before it can be shared with employees. This document becomes your source of truth. Every piece of communication you write after this should stay consistent with it. If your comms team and your operations team are saying different things about what the change means, you've already lost. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Prompt 2: Map your stakeholders You need to know who has influence over adoption before you start sending announcements. This prompt helps you think through it systematically without pretending AI knows your org chart. You are a change management consultant. I'm rolling out [TOOL/CHANGE] to [AUDIENCE]. Help me build a stakeholder map by asking me questions about: who makes the final adoption decision, who influences daily behavior on the ground, who has formally or informally resisted similar changes before, and who will benefit most and least from this change. Based on my answers, help me categorize stakeholders into: Champions, Neutrals, and Likely Resistors. Do not invent names, titles, or organizational details. Use placeholders I can fill in. Run the output past someone who actually knows these people. AI can give you a framework. It can't tell you that [TEAM LEAD NAME] has been burned by the last three "transformation initiatives" and needs a different conversation than everyone else. Prompt 3: Diagnose resistance before it surfaces The best time to handle objections is before they become a Slack thread with 47 reactions. You are a change management specialist. We are rolling out [TOOL/CHANGE] to [TEAM]. Based on common patterns in change resistance, generate a list of likely objections organized by: fear of job security, concerns about workload or learning curve, distrust of leadership intent, technical concerns, and fairness concerns (for example, some teams adopting faster than others). For each objection, suggest an honest response that acknowledges the concern without making promises I can't keep. Do not invent HR policy, legal protections, or headcount guarantees. Flag anything that requires legal or HR input before use. The output won't be perfect. But it'll surface things you haven't thought of, and it'll do it faster than waiting for the first town hall Q&A to go sideways. Prompt 4: Write the announcement The announcement is where most rollouts go wrong before they've started. Too corporate and nobody reads it. Too casual and people don't take it seriously. Too vague and the rumor mill fills in the gaps. You are a communications writer helping an internal team announce [TOOL/CHANGE] to [AUDIENCE]. The change is: [1-2 sentences from your change definition]. Write two versions of the announcement: one for a company-wide email from senior leadership (formal, clear, brief), and one for a team lead to share in a team meeting (conversational, specific to day-to-day impact). Both should include: what's changing, why now, what employees need to do next, and who to contact with questions. Do not make promises about job security, performance reviews, or compensation. Do not include statistics or productivity claims that haven't been verified. Flag any language that HR or legal should review before sending. One thing AI consistently gets wrong here: it makes announcements sound too optimistic. Real employees notice when a change memo reads like a press release. Edit for honesty. Prompt 5: Create manager talking points Managers are where change actually happens or doesn't. If they don't have the right language, they'll either over-promise or go silent, and both are bad. You are a change management consultant. We are rolling out [TOOL/CHANGE] and need to equip managers to have 1:1 conversations with their direct reports. Write a talking points guide that includes: a 2-3 sentence plain-language explanation of the change, honest answers to the three most common concerns (job security, workload, and 'why now'), what managers should escalate to HR or leadership rather than answering themselves, and what they should do if someone refuses to engage. Do not invent policy, job security guarantees, or legal protections. Flag anything that requires HR alignment before managers use it. For more on training managers to actually use AI tools themselves, the AI training prompts guide has templates that work in a similar format. Prompt 6: Design a pilot Rolling out to everyone at once is how you create a spectacular failure instead of a contained one. Pilots exist so you can learn before it's too late to adjust. You are a change management consultant. I want to run a 4-week pilot of [TOOL/CHANGE] with [PILOT GROUP SIZE] volunteers from [TEAM/DEPARTMENT]. Help me design a pilot structure that includes: selection criteria for pilot participants, what we're testing and how, weekly check-in cadence, what data we'll collect to decide whether to expand, and what a 'pause and adjust' trigger looks like. Do not invent adoption benchmarks, ROI estimates, or productivity percentages. Keep the design practical for a team that doesn't have a dedicated change management function. If you want to think through what could go wrong before you launch, the risk assessment prompts are worth running first. Prompt 7: Build training materials This is where "Garbage In, Garbage Out" applies hardest. Vague training materials create confused users who then become resistant ones. Give AI the specifics. You are a learning and development specialist. We are training [ROLE] on [TOOL/CHANGE]. Their main tasks are [LIST 3-5 TASKS]. The training needs to cover: what they'll do differently in their first week, three common mistakes to avoid, and where to get help when they're stuck. Write a one-page quick-start guide and a 5-question quiz to check basic comprehension. Do not invent feature names, product capabilities, or technical specifications. Use [PLACEHOLDER] for any tool-specific details I need to fill in. Flag anything that needs review from the tool vendor or IT before distribution. The onboarding prompts cover a similar format if you're rolling out to new hires alongside existing staff. Prompt 8: Collect feedback without making it theater Most "feedback sessions" on AI rollouts are theater. People say what they think you want to hear, and you collect it and call it adoption data. Here's a prompt for designing something that actually works. You are an organizational development consultant. We have completed [WEEK 2/MONTH 1] of rolling out [TOOL/CHANGE]. Help me design a feedback process that will surface honest concerns, not just satisfaction scores. Include: 3-5 anonymous survey questions that surface real friction (not just 'rate your experience 1-5'), a framework for a 20-minute focus group with frontline users, and a way to capture manager observations that protects the privacy of individual employees. Do not include employee names, performance data, or personal details in any prompt or output. The goal is actionable signals, not a dashboard that looks good. Anonymous beats attributed for early-stage rollouts. People won't tell you the tool is causing them to work longer hours if their name is on the feedback. Prompt 9: Measure adoption that means something "Adoption rate" usually means "we tracked logins." That's not the same as people actually using the tool in ways that help them do their jobs. This prompt helps you define better metrics. You are a change management analyst. We are measuring the adoption of [TOOL/CHANGE] across [TEAM/DEPARTMENT]. Help me define adoption metrics that go beyond login counts. Suggest leading indicators (early signals of real adoption), lagging indicators (proof it's working over time), and warning signs that adoption is superficial or coerced. For each metric, explain how I'd collect it without surveilling individuals or creating a compliance burden. Do not invent industry benchmarks or guaranteed outcome percentages. Flag any metrics that require IT system access or HR approval to collect. This is also where the AI policy prompts become relevant, because measuring adoption properly often requires having a usage policy already in writing. Prompt 10: Write a rollback or adjustment plan Most rollout plans don't have one. This is a mistake. Having a clear "what we do if this isn't working" plan isn't admitting defeat. It's how you maintain credibility with the people who were skeptical from the start. You are a change management consultant. We are 60 days into rolling out [TOOL/CHANGE]. Help me write a rollback and adjustment plan that covers: the conditions under which we'd pause or scale back the rollout, how we'd communicate a pause to employees without damaging trust, what a 'partial rollback' looks like (some teams continue, others pause), and how we'd capture lessons before resuming. Do not make this sound like failure. Frame it as responsible implementation. Do not include confidential strategy, legal details, or headcount decisions. Flag anything that needs HR or legal review before it's shared with the team. The people who were already skeptical will trust you more if you had this plan and stuck to it. The people who were enthusiastic will appreciate that you're taking it seriously. What AI can't do in change management AI can draft, organize, synthesize, and structure. It cannot build trust with a skeptical team lead. It cannot read the room in a town hall. It cannot tell you that the real resistance in your finance team isn't about the tool, it's about the manager who's been fighting with IT for eight months. Dee Kargaev covers this in the context of the "taste moat" in Don't Replace Me: the judgment about what people will actually accept, what language will backfire, and where a slick rollout plan is hiding real organizational friction. That part is still yours. For the broader picture on what AI is genuinely useful for at work and where it falls short, the plain-language breakdown of AI's capabilities is worth reading before you run any of these prompts. Use these templates as a starting point. Assign a named human owner to every decision AI helps you draft. Get HR and legal to review anything touching job security, policy, or compliance. Pilot small. Listen to frontline users before you call the rollout a success. That's not extra work. That's the actual job. Frequently asked questions What are AI change management prompts? AI change management prompts are structured instructions you give to tools like ChatGPT or Claude to help you draft rollout plans, stakeholder maps, training materials, and adoption frameworks. They don't replace operational judgment or HR/legal review, but they help you move faster from blank page to working draft when you're managing an organizational change. Can I use AI to help write an internal announcement about a new tool rollout? Yes, and it's one of the more useful applications. Give the AI your change definition, your audience, and a constraint not to make promises about headcount or policy, and it'll give you a solid first draft in minutes. Edit it for honesty and have HR or legal check anything that touches job security or compliance before you send it. What should I never put into an AI tool during a change management project? Don't paste employee names, performance records, private survey responses with identifying information, customer PII, contract terms, legal disputes, headcount plans, security architecture, credentials, unreleased product details, or proprietary workflows. Use placeholders like [ROLE] or [DEPARTMENT] instead. If your organization has approved specific tools for sensitive work, check which ones those are before you start. How do I measure AI adoption without just tracking logins? Look for leading indicators like whether people are completing the tasks the tool was meant to support, and lagging indicators like whether time spent on those tasks has actually changed. Warning signs of superficial adoption include high login counts alongside rising complaints, or people saying they use the tool but their outputs haven't changed at all. See the AI project management prompts for frameworks that help you track progress without surveillance. Should a manager or HR handle change management for AI rollouts? Both, with different roles. Managers handle the day-to-day conversations with their teams and surface frontline resistance early. HR handles anything touching job security, performance implications, policy, or compliance. Neither should rely on AI-generated output that hasn't been reviewed for accuracy and appropriateness before it goes to employees. Do I need a rollback plan for an AI tool rollout? Yes. Not because rollouts usually fail, but because having one maintains credibility with skeptical employees and forces you to define what success actually looks like before you start. A rollback plan isn't a plan to fail. It's a signal that you're taking the implementation seriously enough to know when to stop and adjust. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Rollout Plan Prompts: Turn Pilot Wins Into Adoption 10 copy-paste AI rollout plan prompts for phased adoption, stakeholder comms, training, data boundaries, success metrics, and rollback criteria. AI rollout planAI adoptionAI prompts AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Automation Prompts: Automate the Boring Stuff 10 copy-paste AI automation prompts for mapping workflows, writing SOPs, adding approval gates, and QA testing. Practical templates with safety rules included. ai automation promptsworkflow automationSOP templates --- ## AI Churn Analysis Prompts: Find Why Customers Leave URL: https://dontreplace.me/articles/ai-churn-analysis-prompts-find-why-customers-leave AI promptschurn analysiscustomer retentioncustomer successdata analysis AI Churn Analysis Prompts: Find Why Customers Leave 10 AI churn analysis prompts to investigate why customers leave, test assumptions, protect customer data, and turn evidence into accountable action. August 28, 2026 Updated August 28, 2026 Churn analysis often begins with a dashboard, a cancellation export, and somebody asking, “So why did they leave?” as if the spreadsheet contains tiny farewell letters. AI churn analysis prompts can organize evidence, code feedback, compare cohorts, surface missing data, and draft an investigation plan. They cannot read a customer's mind, establish causation from a correlation, or produce a trustworthy churn probability because you fed them twelve angry support tickets. AI can make churn evidence easier to inspect. Humans still own the definitions, statistical judgment, customer conversations, privacy decisions, and response. These ten templates help customer-success leaders, product managers, analysts, marketers, founders, and operations teams investigate customer churn without turning autocomplete into an oracle. If the immediate problem is saving current accounts rather than learning from past churn, start with AI customer retention prompts. What churn analysis should actually answer Good churn analysis does more than count departures. It clarifies what happened, to whom, when, under which definitions, and what evidence could explain it. It also states what the evidence cannot prove. Before prompting anything, define the metric: Logo churn: the share of customer accounts lost during a period. Revenue churn: recurring revenue lost, often separated into gross and net revenue retention. Voluntary churn: a customer actively cancels or does not renew. Involuntary churn: payment failure, expired cards, procurement failure, or another process issue ends service. Early-life churn: departure during onboarding or initial adoption. Mature-account churn: departure after the customer has had time to receive value. Mixing these categories produces useless averages. A failed credit card is not the same problem as an unusable feature. A tiny self-serve account and a multi-year enterprise customer should not silently become equal evidence for every decision. Churn analysis should produce testable explanations, not a cinematic villain monologue about “customers losing trust.” Direct customer statements matter, but cancellation surveys are incomplete and selected. Behavioral data matters, but an event log records activity—not intention. Support records matter, but customers who never opened a ticket may still have struggled. For broader grouping work, use customer segmentation prompts. For the experience around each stage, use customer journey mapping prompts. Churn analysis connects those views while keeping uncertainty visible. The reusable evidence-first churn prompt formula Add this instruction to any template below: “Act as a churn-analysis assistant. Use only the sanitized evidence I provide for [customer population] during [date range]. The metric definition is [definition], and the decision is [decision]. Produce [artifact]. Separate observed facts, direct customer statements, calculated measures, interpretations, hypotheses, and unknowns. For every important finding, include the source, cohort definition, sample size, date range, confidence note, contradictory evidence, competing explanation, owner, and next validation step. Do not invent quotes, customer motives, probabilities, causal claims, or sensitive attributes. Flag missing data, selection effects, tracking changes, unfair proxies, and conclusions requiring statistical, privacy, legal, commercial, or customer review.” This formula forces the output to show its homework. Without it, a model may transform “usage fell before cancellation” into “low usage causes churn.” Maybe. Or perhaps broken tracking, seasonality, an already-completed project, a pricing change, or a customer decision caused both. Never paste customer names, emails, account IDs, payment details, contracts, credentials, private support conversations, health information, confidential strategy, or legally sensitive data into an unapproved AI tool. Use approved systems, minimum necessary fields, aggregation, de-identification, access controls, retention limits, and human privacy review. Removing a name does not anonymize a record when the remaining details identify one company or person. What to collect before prompting A useful analysis starts with a documented evidence pack. InputWhy it mattersHuman check Churn definitionPrevents metric soupFinance or analytics owner confirms Population and exclusionsDefines who is countedAnalyst reviews filters Observation periodExposes seasonality and lagBusiness owner approves Cohort dimensionsEnables fair comparisonAnalyst checks sample sizes Product-event definitionsPrevents tracking fictionData owner verifies Billing and contract eventsSeparates commercial causesFinance or legal owner checks Support historyShows reported frictionSupport lead reviews Direct customer feedbackPreserves stated reasonsResearch owner checks source Product and pricing changesAdds timeline contextProduct owner verifies Missing-data notesLimits false confidenceAnalyst documents Privacy classificationControls sensitive processingPrivacy owner approves Decision and ownerMakes analysis actionableAccountable leader confirms Keep source links in approved systems. Document query versions and metric logic. If two teams define “active customer” differently, resolve that before asking AI for strategic insight. Otherwise the model will confidently summarize an argument nobody has noticed yet. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI churn analysis prompts Replace the brackets with sanitized facts. Treat each output as a draft for human review. 1. Prepare a clean churn dataset brief “Turn this sanitized data dictionary, metric logic, date range, table inventory, join plan, exclusions, and known limitations into a churn dataset brief: [paste]. Include the unit of analysis, churn event, observation window, outcome window, cohort fields, source owner, refresh timing, missingness, and privacy classification. Flag ambiguous definitions, possible leakage, duplicate entities, changing identifiers, and fields that should be removed or aggregated.” Run this before requesting findings. If the dataset accidentally includes events recorded after cancellation, a model may describe hindsight as prediction. If one account appears under three workspace IDs, logo churn can become performance art. Have an analyst verify joins, denominators, time zones, refunds, reactivations, and account hierarchies. AI data analysis prompts can structure the review, but generated SQL still needs testing. 2. Audit churn evidence quality “Audit these sanitized churn evidence sources: [paste]. For each source, list coverage, freshness, reliability, collection method, likely bias, sensitive content, contradictory evidence, and claims it can and cannot support. Check for missing-not-at-random feedback, survivorship bias, seasonality, tracking changes, small samples, duplicated reasons, and vague free-text labels. Return a prioritized repair list with an owner.” Cancellation surveys disproportionately represent people willing to complete them. Support tickets represent reported problems, not all problems. Exit interviews may overrepresent valuable accounts that received personal outreach. The goal is not to reject imperfect evidence. It is to stop imperfect evidence from wearing a fake mustache labeled “objective truth.” 3. Separate observations from assumptions “Classify these churn claims, notes, and dashboard annotations: [paste]. Put each into observed event, direct customer statement, calculated measure, contextual fact, interpretation, hypothesis, unsupported assumption, sensitive inference, or unknown. Cite its source and date. Rewrite unsupported conclusions as validation questions. Do not infer intent, emotion, budget, protected characteristics, or probability.” “Account canceled on August 4” is an event. “Customer cited missing export controls” is a statement. “They left because the product is too complex” is a causal claim unless stronger evidence supports it. This classification is especially useful in executive reviews, where five careful caveats can mysteriously become one confident bullet between the analyst's desk and the slide deck. 4. Code cancellation feedback without flattening it “Code these de-identified cancellation survey responses and interview notes: [paste]. Preserve each original excerpt ID. Propose a hierarchical theme codebook, apply primary and secondary codes, mark ambiguous cases, and report counts with denominators. Separate stated reason, contributing friction, trigger event, desired outcome, and unresolved unknown. Do not fabricate quotes or merge contradictory responses into one narrative.” A useful codebook distinguishes “price is high” from “budget disappeared,” “value was unclear,” and “procurement rejected the renewal.” Those conditions may share a surface word and demand completely different responses. Have a human review a sample, revise the codebook, and measure agreement. If you need better primary evidence, use customer interview prompts rather than asking the model to elaborate on thin notes. 5. Compare churned and retained cohorts “Compare these sanitized aggregate metrics for churned and retained cohorts: [paste]. Confirm cohort definitions, date windows, denominators, sample sizes, and missingness first. Report meaningful differences, uncertainty, contradictory measures, and plausible confounders. Do not call any difference causal. Suggest checks for cohort mix, tenure, plan, acquisition channel, company size, seasonality, pricing, and product changes.” A retained cohort has, by definition, survived long enough to be observed. Churned customers may have shorter histories and fewer opportunities to generate events. Raw totals therefore flatter retained accounts. Use rates and comparable windows. Check whether the cohorts entered during different product versions or campaigns. A statistically neat result can still describe who you acquired rather than what your product did. 6. Investigate when churn happens “Analyze this de-identified churn timing summary by tenure, onboarding milestone, renewal cycle, billing event, product release, and support incident: [paste]. Identify concentrations and gaps. For each pattern, provide competing explanations and the evidence needed to distinguish them. Separate early-life, mature-account, voluntary, and involuntary churn. Return a timeline and validation plan, not a prediction.” Timing can narrow an investigation. Churn before setup completion points toward a different evidence trail than churn after a pricing change. A spike after a release may matter—or coincide with annual renewals. Ask product, billing, support, and customer-success owners to annotate major process changes. The timeline is a map for investigation, not a conviction handed down by the graph. 7. Connect support and product friction carefully “Review these aggregated support themes, product events, incident dates, and churn outcomes: [paste]. Map where evidence overlaps and where it does not. Report issue frequency, affected cohort, resolution status, timing relative to churn, and missing silent-customer evidence. Flag severe but rare issues separately from frequent minor friction. Do not assume a ticket caused cancellation or that no ticket means no problem.” Customers vary in willingness and ability to report trouble. Enterprise customers may escalate through account teams while self-serve customers simply disappear. Your support system is a window, not the whole building. Use customer service prompts to improve issue summaries, then validate product patterns with owners who understand instrumentation and incident history. 8. Challenge a churn hypothesis “Red-team this churn hypothesis: [paste hypothesis]. Evidence: [paste sanitized summary]. Build the strongest supporting case, strongest opposing case, at least five competing explanations, evidence missing from each case, possible confounders, and conditions that would falsify the hypothesis. Label correlation versus causal evidence. End with the smallest responsible validation step and a named human owner.” This prompt fights confirmation bias. Teams naturally prefer explanations they can act on, especially explanations that justify work already planned. “Customers churn because we need Feature X” deserves the same scrutiny as any other convenient story. For deeper causal discipline, borrow the evidence ladder from root cause analysis prompts. The objective is not to prove everyone wrong. It is to make expensive confidence earn its keep. 9. Create a human-owned research plan “Turn these churn unknowns and competing hypotheses into a research plan: [paste]. Prioritize questions by decision value, customer impact, urgency, risk, and feasibility. For each question, recommend an appropriate method such as query validation, cohort analysis, interview, survey, usability study, billing audit, support review, or controlled experiment. Include sample considerations, consent, privacy controls, owner, timeline, and a stop condition. Do not recommend manipulative outreach or sensitive targeting.” Not every unknown deserves a dashboard. Some need five customer conversations. Others need a corrected event definition, a billing audit, or an experiment that protects current users. Run the plan through a risk assessment before customer-facing tests. A retention experiment that creates dark patterns, hides cancellation, or pressures vulnerable customers is not clever. It is a future apology drafted in advance. 10. Turn findings into an accountable action brief “Convert these validated churn findings into an action brief for [audience]: [paste]. Include the metric definition, affected cohort, evidence, uncertainty, customer impact, proposed action, alternatives considered, privacy and fairness risks, dependencies, owner, decision deadline, leading indicators, lagging indicators, guardrails, and review date. Separate approved actions from ideas. Preserve contradictory evidence and do not promise an outcome.” A good brief connects evidence to a reversible next step. It does not convert every theme into a feature request or every churned dollar into a fantasy revenue-save number. Prioritize with AI prioritization prompts, but keep the decision human-owned. Sometimes the responsible conclusion is that a segment is a poor fit, a normal lifecycle ended, or the available evidence cannot justify intervention. Common churn analysis mistakes Treating correlation as diagnosis Customers who use a feature may retain longer because committed customers are more likely to discover it. That does not prove forcing feature adoption will improve retention. State the relationship, list alternatives, and validate. Ignoring involuntary churn Payment failures, procurement delays, expired cards, and administrative mistakes can contaminate product conclusions. Separate them before analyzing behavior. Trusting stated reasons too literally A cancellation reason may be accurate, partial, socially convenient, selected from a bad dropdown, or entered by someone outside the real decision. Preserve it as evidence without promoting it to complete causation. Comparing unfair windows A twelve-month retained account naturally has more usage than an account that left in week three. Use aligned observation windows and comparable cohorts. Hiding uncertainty from leaders A polished summary that removes caveats is not concise; it is misleading. Keep sample size, missingness, confidence, and competing explanations beside the headline. Automating pressure Do not use churn analysis to create hidden cancellation barriers, deceptive discounts, invasive scoring, or harassment workflows. The action should help the customer or improve the product—not merely defend a metric. A practical human review checklist Before acting on AI-assisted churn analysis, confirm: The churn definition, population, date range, and denominator are explicit. Voluntary and involuntary churn are separated where relevant. Cohorts use comparable observation windows. Sources, queries, and event definitions have owners. Direct statements are distinct from interpretations. Sample size, missingness, selection effects, and seasonality are visible. Correlations are not described as causes. Sensitive fields and unfair proxies are excluded or governed. Important patterns have competing explanations. Customer-facing research has consent and privacy review. Proposed actions include guardrails and stop conditions. A named human owns the decision and review date. If the workflow itself is new, read the no-BS guide to using AI at work. If someone is treating fluent output as intelligence, the five-minute explanation of what AI can and cannot do is the cheaper intervention. Frequently asked questions Can ChatGPT analyze customer churn? It can help organize sanitized evidence, summarize themes, compare supplied aggregates, challenge hypotheses, and draft research plans. It cannot independently verify your warehouse logic, know an unstated customer motive, or establish causation. A qualified human must validate the data and interpretation. What data should I give an AI for churn analysis? Use the minimum approved data necessary: documented metric definitions, aggregate cohort measures, de-identified feedback, event definitions, date ranges, and known limitations. Avoid names, emails, account IDs, payment details, contracts, private messages, credentials, and sensitive personal information. Can AI predict which customer will churn? A language model should not be treated as a reliable individual churn predictor. Even purpose-built predictive models require careful labels, leakage checks, calibration, fairness review, monitoring, and lawful use. Predictions are not explanations, and high-risk customer treatment needs human governance. How do I prompt AI to find churn reasons? Do not ask simply, “Why did customers leave?” Supply a defined cohort, evidence sources, dates, sample size, limitations, and a decision. Require the output to separate facts, statements, hypotheses, and unknowns while listing competing explanations and validation steps. What is the difference between churn analysis and retention analysis? Churn analysis investigates departures and their surrounding evidence. Retention analysis examines what supports continued value and what actions may responsibly reduce preventable loss. They overlap, but neither should assume every departure is preventable or every retained account is healthy. How do I avoid bias in AI-assisted churn analysis? Inspect who is represented, who supplied feedback, which fields act as sensitive proxies, how cohorts differ, and which records are missing. Use aggregation, de-identification, documented exclusions, privacy review, and subgroup checks. Never let AI infer protected traits or recommend unequal treatment from weak proxies. Can AI churn analysis replace customer interviews? No. It can summarize existing notes and help prepare neutral questions. Direct research reveals context that event data and cancellation forms miss. Interviews still require consent, skilled facilitation, careful sampling, and human interpretation. What should the final churn report include? Include the metric definition, population, period, evidence sources, data-quality limits, cohort comparisons, customer statements, hypotheses, competing explanations, confidence, recommended validation, proposed actions, risks, owners, and review dates. Keep uncertainty next to each major conclusion. The bottom line AI churn analysis prompts are useful when they make evidence easier to inspect and assumptions harder to hide. They are dangerous when fluent summaries get promoted into customer motives, causal claims, or invasive action. Define the metric. Protect the data. Compare fair cohorts. Preserve disagreement. Test the explanation. Give every decision a human owner. That is less magical than an AI churn oracle. It is also how you learn something real. For a broader field guide to using AI without surrendering judgment, Don't Replace Me by Dmitry Kargaev keeps the same rule throughout: let the machine do speed; keep responsibility human. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Customer Retention Prompts: Keep Customers 10 AI customer retention prompts to diagnose churn signals, plan useful follow-up, protect customer data, and keep human judgment in charge. AI promptscustomer retentioncustomer success AI Account Planning Prompts: Build a Better Plan 10 AI account planning prompts to organize buyer evidence, map stakeholders, expose risks, test strategy, and plan next actions without inventing intel. AI promptsaccount planningsales strategy AI Customer Renewal Prompts: Prepare With Evidence 10 AI customer renewal prompts to verify outcomes, expose account risks, prepare options, and protect trust without inventing value or customer sentiment. AI promptscustomer renewalcustomer success --- ## AI Client Communication Prompts (10 Templates) URL: https://dontreplace.me/articles/ai-client-communication-prompts-say-the-hard-thing ai client communicationclient email promptsai prompts for workclient communication templatesdifficult client emails AI Client Communication Prompts (10 Templates) 10 copy-paste AI prompts for client emails: delays, scope creep, tense calls, and bad news. Draft faster, verify everything, own every word. June 11, 2026 Updated June 11, 2026 Client communication is the part of the job nobody teaches you. You learn it by sending one wrong email, watching a relationship go cold, and spending three weeks rebuilding trust you didn't know you'd burned. AI client communication prompts can shortcut the drafting part. They cannot shortcut the judgment part. That's still yours. This article gives you 10 copy-paste prompts for the situations people dread most: delays, scope creep, tense calls, bad news. Use them to draft faster. Then read what comes out, verify every fact, and own every word before you hit send. The formula before the prompts Every good client communication prompt has the same structure. Give the AI the facts, the audience, the constraint, the tone, and what you need. Vague input gives you corporate apology vapor. Specific input gives you a draft worth editing. Here's the reusable formula: [Context] I'm writing to a client [describe the relationship and project]. [Situation] The situation is [describe what happened factually]. [Constraint] We cannot [state any promises, timelines, or things you're not doing]. [Tone] The tone should be [direct/warm/brief/formal]. [Output] Write a [email/summary/update] that [states the goal clearly]. Fill in each bracket before you run the prompt. If you don't know what goes in a bracket, that's a sign you need to think for another five minutes before drafting anything. The most common mistake with this formula is skipping the constraint. People fill in context and situation, then leave the constraint blank because it feels awkward to write out. That's exactly when AI invents a delivery date your team never agreed to, or apologizes in a way that implies legal liability you don't have. The constraint bracket is the most important one. Why AI client communication prompts fail without you AI is fast. It's also relentlessly polite, structurally confident, and completely indifferent to whether what it says is true. That combination is fine for a draft. It's a disaster for a final email to a client whose launch just slipped two weeks. Rule #5 from Don't Replace Me puts it plainly: it's not smart, it's fast. Speed is useful. Speed with no one checking the facts is how you send a client an apology that promises a new delivery date your team never agreed to. The human work in client communication is knowing what to say, what to soften, what to escalate, and what cannot be delegated. That's the part AI genuinely can't replace. AI can give you three tone variants of the same hard message. Deciding which version to send, and whether to call first, is still yours. There's also a context problem. AI knows nothing about your relationship with this client. It doesn't know that your contact is under pressure from their own boss, that you missed a deadline six months ago and rebuilt trust slowly, or that this particular client reads directness as rudeness. You carry that history. The AI doesn't. So when you take a draft from AI and make it yours, that's where most of the real work happens. Before any of these prompts: do not paste confidential client data, private contracts, employee names, legal matters, unreleased strategy, security incidents, or customer PII into unapproved AI tools. Anonymize everything. Use "[Client]" instead of the actual company name. Use "[Project X]" instead of the real project. Then verify every date, commitment, and claim the AI produces before sending. AI client communication prompts for common situations 1. Weekly client update Prompt: I'm writing a weekly update to a client [describe relationship and project stage]. Progress this week: [list 3-5 factual bullet points]. Blockers: [list any]. Planned for next week: [list]. Tone: confident and brief. Write a professional weekly update email, under 200 words, with no filler phrases like "hope this finds you well." What to verify: Every claim in the bullets. No AI should be inventing progress you didn't make. Where this saves time: Weekly updates are low-stakes but time-consuming. Most people spend 20 minutes writing something that takes two minutes to read. AI gets you to a clean structure in under a minute. You spend the remaining time checking the facts and adjusting tone, not staring at a blank screen. 2. Explaining a delay without making excuses Prompt: A project deliverable is delayed by [X days/weeks]. The factual reason is [explain clearly]. We are responsible for [state what's on us]. Next step is [state the concrete action and realistic date]. Write an email to the client that acknowledges the delay directly, states what we're doing, gives the revised timeline, and does not blame external factors or use passive voice. What to verify: The revised timeline. Make sure your team agreed to it before it goes in an email. Where this saves time: Delay emails are the hardest to start because the stakes feel high and the instinct is to hedge everything. AI removes the blank-page problem. It'll usually produce something too soft on accountability. Your job is to put the accountability back in. 3. Saying no to scope creep Prompt: A client has asked for [describe the additional request] which is outside our original agreement covering [describe the original scope]. I want to say no without damaging the relationship, and offer a path forward. Write a short, direct email that acknowledges the request, explains it's outside scope, and offers either a change order discussion or a future phase. Keep it professional and not defensive. What to verify: That your original agreement actually covers what you think it does. Check the contract before sending. Scope creep emails have a specific failure mode: they come across as punitive. "That's out of scope" delivered badly sounds like you're nickel-and-diming a client who's been loyal. The good version acknowledges the value of the request first, then explains the path to getting it done. The AI is decent at this structure if you give it enough context about the relationship. 4. Asking for missing feedback Prompt: I've been waiting [X days/weeks] for feedback on [deliverable]. Without it, [state the concrete impact: timeline slips, next phase blocked, etc.]. Write a brief, polite follow-up email that creates urgency without sounding passive-aggressive. Include a clear action and deadline. What to verify: The date you sent the original. The actual impact of the delay, not a vague "this may affect timelines." Where this saves time: Chasing feedback is uncomfortable because it can sound accusatory. AI is good at threading the needle between urgent and polite. Where it fails is specificity. Make sure the concrete impact you give it is real, not hypothetical. 5. Summarizing a difficult call Prompt: We had a difficult call with a client about [describe the issue, no names]. Points raised by the client: [list]. Points we made: [list]. Agreements reached: [list, or "none yet"]. Open questions: [list]. Write a concise call summary email I can send within 24 hours that captures all of this accurately, stays neutral in tone, and ends with clear next steps. What to verify: Every point on both sides. The AI will smooth things over. Make sure the summary reflects what actually happened, including the uncomfortable parts. The instinct after a difficult call is to write something diplomatic that papers over the tension. That instinct is usually wrong. A call summary that buries the disagreement means you have the same conversation again in two weeks, except now nobody can remember what was said. Write it accurately. Let the AI help with neutrality, not omission. 6. Escalating a risk early Prompt: I need to alert a client to an early risk on [project]. The risk is [describe factually]. If unaddressed, the potential impact is [describe concretely]. We recommend [action/decision needed]. Write a brief risk escalation email that is direct, calm, and ends with a specific ask for a decision or meeting by [date]. What to verify: That the risk is real and documented. Don't escalate speculation as fact. Where this saves time: Risk escalation emails get delayed because nobody wants to deliver bad news before they have to. AI removes the drafting friction so you can stop stalling and send it. The earlier you flag a real risk, the better it lands. A problem you told the client about in week two is a problem you managed. The same problem in week eight is a problem you hid. 7. Turning messy notes into a decision recap Prompt: Here are rough notes from a client meeting: [paste anonymized notes]. Turn these into a clean decision recap email with four sections: Decisions Made, Open Items, Next Steps (with owners and dates), and Questions We Need Answered. Keep it under 300 words. Do not invent any decisions or owners that aren't in the notes. What to verify: Every owner and every date. The AI will fill in blanks if you leave them. That's where invented commitments come from. This prompt has the highest ROI of any on this list. Taking rough meeting notes and turning them into a structured recap is pure formatting work. It's also the task most people skip because it feels tedious right after a long meeting. Do it anyway. A clear recap sent within 24 hours prevents more misalignment than almost anything else you can do. 8. Rewriting a defensive draft into an accountable one Prompt: Here is a draft email I wrote: [paste your draft]. It sounds defensive and puts blame on the client. Rewrite it so it: takes clear ownership of what is ours, states facts without excuses, and ends with a concrete next step. Keep the length similar. Do not add any new commitments or facts. What to verify: That ownership is correctly assigned. If AI takes blame that isn't yours, fix it. If it softens blame that is yours, put it back. This is one of the more underused applications. Most people think of AI as a tool for generating things from scratch. But using it to fix a draft you already wrote is often more useful. You know what happened. You wrote it out in a defensive way because you're human and it's uncomfortable to own a mistake. AI rewrites it without the defensiveness. Then you check whether the accountability is accurate. 9. Preparing options for a client-facing tradeoff Prompt: We need to present a tradeoff to a client. The situation is [describe]. Option A is [describe, with pros and cons]. Option B is [describe, with pros and cons]. Option C (if any) is [describe]. Write a short email presenting these three options clearly, with a recommendation if I have one, and a specific deadline for the client to decide. Tone: direct and consultative. What to verify: That the options are real and your team can actually deliver any of them. Clients hate being surprised by a single bad option. They handle it much better when you present the situation, show the tradeoffs, and give them agency. The AI is good at formatting this clearly. What it can't do is sanity-check whether Option B is actually feasible. That's on you before this goes anywhere near a client inbox. 10. Follow-up after a tense exchange Prompt: I had a tense exchange with a client about [describe the issue without personal details]. I want to send a follow-up that acknowledges the difficulty of the conversation, reaffirms the relationship, and moves us toward a constructive next step. Write a brief, genuine email that is warm but not sycophantic. Do not make any new promises. Keep it under 150 words. What to verify: Tone. AI tends to over-apologize or under-acknowledge. Read it aloud before sending. The timing of this one matters. Send it too fast and it feels reactive. Wait too long and the silence compounds the tension. Usually 24 hours is right. The prompt tells AI not to make new promises, but check for this anyway. After a tense call, AI sometimes tries to smooth things over with vague commitments like "we'll make sure this doesn't happen again." If your team can't guarantee that, cut it. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What AI should never do in client communication This matters more than the prompts themselves. Do not let AI invent timelines your team hasn't agreed to. Don't let it quote the client's own words back to them from your paraphrase. Don't let it make legal statements, financial commitments, or delivery guarantees. Don't let it apologize on behalf of people who haven't approved the apology. If the output contains a specific date, a number, a name, or a promise, you are responsible for verifying it before it reaches the client. "The AI wrote it" is not a defense your client cares about, and it won't fix a relationship that breaks because of a commitment nobody made. There's a subtler problem too. AI writes confidently regardless of accuracy. A hedged, uncertain situation gets turned into clean, declarative language because that's what good writing looks like. But sometimes the right email is a little uncertain, because you genuinely don't know yet. If you're not sure when the revised deliverable is ready, don't let AI pick a date for you. Send the email without a date and say you'll confirm by Thursday. That's more honest, and honest is what keeps clients. For the broader picture on where AI genuinely helps and where it breaks, this plain-language explainer on what AI can and can't do is worth five minutes. And if you're doing this kind of work inside a team, the AI manager prompts article has templates for the internal side of difficult conversations. How to build this into a real workflow Using these prompts ad hoc works. Building them into a system works better. A few things that help. Keep a running doc with your most-used prompt templates pre-filled with the static parts: your project type, your standard tone, your formatting preferences. When a situation comes up, you only fill in the variable parts. That cuts the time from "I need to write this email" to "I have a draft to edit" from ten minutes to two. If you manage a team that communicates with clients, standardize on two or three prompts that fit your most common situations. A weekly update format, a delay template, a scope discussion email. When everyone uses the same structure, your clients get consistent communication regardless of who's writing. That consistency reads as professionalism, even when different people have very different writing styles. One more thing: keep the outputs. A folder of your edited AI drafts is also a folder of your communication history with that client. When a dispute comes up about what was agreed or when something was communicated, you have a paper trail that started with structure. That's useful. Using these prompts well Think of AI as a fast communications editor, not a communications director. It's good at structure, tone variants, clarity, brevity, and catching passive voice. It's bad at context, history, relationship dynamics, and knowing when the most important thing is to pick up the phone instead. The broader guide to using AI at work covers how to build this habit across your whole job, not just client emails. The pattern is the same everywhere: you own the thinking, AI speeds up the drafting, you own the output. The reason to use these prompts is not to sound more polished. It's to spend less time staring at a blank email and more time on the judgment that actually keeps clients. That judgment is yours. Don't outsource it. Frequently asked questions Can I use AI to write emails to clients? Yes, with conditions. AI is useful for drafting, restructuring, and softening tone. You need to verify every fact, date, commitment, and claim before sending. Don't paste real client data into unapproved tools, and don't let AI invent promises your team hasn't agreed to. What's the biggest mistake people make with AI client communication prompts? Sending the first draft. AI produces clean, confident language regardless of whether the content is accurate. The most common problem is AI inventing timelines, softening accountability, or adding commitments nobody made. Treat every AI draft as a starting point, not a final product. How do I explain a delay to a client using AI? Give the AI the factual reason for the delay, what your team is responsible for, the revised timeline your team has actually agreed to, and the next concrete step. Then verify the output against your real situation before sending. A template is in the prompts above. More general email templates are in the AI email prompts guide. Should I tell clients their communications were drafted with AI? That depends on your relationship and any contractual obligations. For most professional contexts, AI-assisted drafting is similar to using templates or having a colleague review. What matters is that you own the content, the facts are correct, and the commitments are real. Don't hide it if directly asked. What client information should I never put into AI tools? Don't paste client names, confidential project data, contract terms, employee issues, legal matters, security incidents, unreleased strategy, financial details, or any personally identifiable information into AI tools that haven't been approved by your organization. Anonymize everything. Use placeholders like [Client] and [Project X]. How do I write a no-to-scope-creep email that doesn't kill the relationship? The key is acknowledging the request genuinely before explaining why it's outside scope, then offering a real path forward like a change order or a future phase. Prompt #3 above handles this directly. Check your actual contract before sending anything, and have the conversation by phone first if the relationship is at all fragile. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Prompts for Work: 12 Templates You Can Use Today 12 copy-paste AI prompt templates for emails, reports, and decisions. Real examples for normal people, with what to check before you send anything. ai prompts for workchatgpt promptsworkplace ai 10 AI Research Prompts That Actually Work 10 copy-paste AI research prompts for desk research. Learn to summarize sources, extract claims, spot gaps, and build decision memos without hallucinations. ai research promptsdesk researchchatgpt research AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation --- ## AI Compatibility Testing Prompts: Catch Breakage URL: https://dontreplace.me/articles/ai-compatibility-testing-prompts-catch-cross-platform-breakage AI promptscompatibility testingcross-browser testingquality assurancesoftware testing AI Compatibility Testing Prompts: Catch Breakage 10 AI compatibility testing prompts for browser, device, OS, responsive, locale, network, defect triage, and evidence-based release decisions. August 18, 2026 Updated August 18, 2026 Compatibility testing is where the sentence “works on my machine” goes to receive adult supervision. A product can work perfectly in one developer’s browser and fail on an older phone, a different rendering engine, a narrow viewport, a translated interface, a keyboard-only flow, or a network that is not office Wi-Fi. AI compatibility testing prompts can help organize that mess. A model can draft a support matrix, suggest test conditions, expose missing assumptions, and summarize sanitized defect evidence. It cannot open every browser, reproduce a real device, verify what a screen reader announces, or declare the release safe. AI can draft the map. Humans still choose the territory, run the tests, inspect the evidence, fix the product, and own the release decision. Start with an AI test plan if the overall strategy is unclear. Use AI test case prompts when you need detailed checks. This guide focuses on compatibility: deciding which environments matter and proving critical journeys work there. What compatibility testing actually covers Compatibility is broader than checking Chrome and Safari at two window sizes. A useful plan considers combinations that can change behavior: Browser engine and version: Chromium, WebKit, and Gecko can differ in layout, APIs, media, storage, and form behavior. Operating system: fonts, permissions, file pickers, keyboards, notifications, and system controls vary. Device class: desktop, tablet, phone, foldable, and constrained hardware create different risks. Viewport and display: width, height, pixel density, zoom, text scaling, orientation, and safe areas matter. Input method: mouse, touch, keyboard, stylus, switch control, and assistive technology expose different failures. Locale and timezone: longer text, right-to-left layouts, number formats, calendars, currencies, and date boundaries break assumptions. Network and dependency conditions: latency, packet loss, offline transitions, content blockers, CDNs, identity providers, and payment services affect outcomes. Configuration: feature flags, permissions, account state, themes, and privacy settings create additional branches. You cannot test every possible combination. The job is to build a defensible, risk-based sample. Use actual audience evidence, a documented support policy, technical change, critical journeys, failure severity, and business obligations. “Most people use Chrome” is not a complete strategy. It ignores browser engines, high-value minorities, contractual commitments, accessibility, and the possibility that your analytics stopped recording users whose browser could not load the page. The reusable compatibility testing prompt formula Use this base prompt, then attach one specialized template below: “Act as a compatibility test-planning assistant. I am evaluating [named product, build, and critical journeys]. Use only the supplied support policy, sanitized analytics, requirements, architecture notes, change list, environment inventory, and verified defects. Produce [artifact] with coverage, priorities, assumptions, setup, expected behavior, evidence fields, exclusions, and open questions. Separate planned checks, observed facts, hypotheses, and decisions. Do not invent execution, market share, device behavior, test results, defect causes, or release approval.” That last sentence prevents a common failure: the model turns a proposed matrix into fictional proof. A test idea is not a test result. An emulator screenshot is not automatically real-device evidence. A defect seen on one configuration does not establish its cause or full blast radius. Never paste credentials, API keys, customer PII, production logs, private source code, proprietary analytics, unreleased screenshots, regulated data, or confidential client material into an unapproved AI tool. Use synthetic accounts, redacted evidence, approved test environments, placeholder secrets, and your organization’s privacy, security, vendor, and retention rules. What to collect before prompting A model needs verified context, not “test it everywhere.” Build a small sanitized input packet. InputWhy it mattersHuman check Support policyDefines promised environmentsProduct and support approve Sanitized usage analyticsGrounds prioritizationAnalytics owner validates gaps Critical user journeysFocuses testing on outcomesProduct owner confirms priority Recent change listIdentifies new compatibility riskEngineers verify affected layers Browser and device inventoryShows available evidence sourcesQA confirms real vs. emulated Design breakpoints and tokensGrounds visual expectationsDesign reviews intended behavior Accessibility requirementsPrevents compatibility tunnel visionAccessibility owner verifies scope Locale and timezone listReveals formatting and layout riskRegional owners confirm Third-party dependenciesExposes browser-specific integrationsService owners verify support Evidence and severity rulesMakes results actionableQA and engineering agree Treat analytics cautiously. Record the time range, consent effects, unknown traffic, bot filtering, and whether broken clients can report at all. If the data is incomplete, label the uncertainty and test a reasonable safety set rather than pretending the chart is scripture. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI compatibility testing prompts Replace bracketed text with verified, sanitized information. These prompts create planning and analysis artifacts. Humans must execute the checks in real tools and environments. 1. Turn promises into a support policy “Using these product requirements, customer commitments, usage data, vendor requirements, technical constraints, and support history: [paste], draft a versioned compatibility support policy. Separate fully supported, best-effort, and unsupported environments. Include browser engines and versions, operating systems, device classes, accessibility assumptions, locale coverage, review cadence, exception process, and evidence needed before dropping support. Label every proposed choice for human approval.” A support policy prevents the matrix from changing according to whoever filed the latest bug. It should say what “supported” means: perhaps critical journeys must work, content must remain available, and cosmetic differences may be accepted. It should also define how old versions age out. Do not let AI invent browser market share or contractual obligations. Product, legal, customer support, sales, engineering, and accessibility owners may all hold relevant facts. Dropping a platform can exclude users and violate a promise; autocomplete does not get that vote. 2. Build a risk-based browser and device matrix “Using this approved support policy, sanitized analytics, critical journeys, revenue or service impact, recent changes, historical defects, browser engines, operating systems, device classes, and available test inventory: [paste], create a prioritized compatibility matrix. Score each configuration by usage, impact, technical risk, and evidence confidence. Mark real-device, cloud-device, emulator, and manual-only coverage. Explain exclusions and unresolved gaps.” Group browsers by engine before multiplying logos. Chrome and Edge often share Chromium, but integrations, policies, update channels, and OS behavior can still differ. Safari adds WebKit risk; Firefox adds Gecko risk. Mobile browsers also inherit operating-system and embedded-webview constraints. A matrix should be small enough to run and broad enough to catch meaningful variation. Tier one might receive full critical-journey coverage. Tier two gets smoke and changed-area checks. Long-tail environments may get graceful-degradation checks. The tiers are human decisions backed by evidence, not an excuse to ignore inconvenient users. 3. Map critical journeys across the matrix “Given these critical user journeys, support tiers, browser-device matrix, account states, permissions, integrations, and failure severity rules: [paste], create a journey-by-environment coverage table. Include preconditions, representative data, expected outcome, compatibility-specific risks, evidence to capture, cleanup, and escalation rules. Identify redundant combinations and dangerous gaps.” Test outcomes, not isolated pages. A checkout journey includes navigation, form controls, validation, payment handoff, return flow, confirmation, and persisted order state. A file upload includes selection, permission, progress, cancellation, error recovery, and the downloaded result. Prioritize authentication, account recovery, payments, creation and editing, search, navigation, data export, and any journey tied to safety or legal obligations. Use AI smoke test prompts for a compact changed-build pass, but keep deeper checks for risky combinations. 4. Cover browser engines and web APIs “Using this frontend change list, browser support policy, build targets, polyfills, CSS features, JavaScript APIs, storage behavior, media usage, permissions, and third-party scripts: [paste], draft cross-browser checks. Map each risky feature to browser engines, fallback behavior, feature detection, expected console and network evidence, and a manual verification step. Do not assume transpilation guarantees runtime compatibility.” A successful build does not prove browser compatibility. Problems can hide in CSS layout, date parsing, storage restrictions, cookies, autoplay, clipboard access, file handling, focus behavior, popovers, dialogs, and third-party widgets. Check graceful degradation. When a capability is unavailable, users need a clear alternative or explanation instead of a dead button. Capture browser version, OS, viewport, settings, console output, and network evidence. “Broken in Safari” is not enough for diagnosis. 5. Test responsive layouts and input methods “Using these design breakpoints, component states, content samples, navigation patterns, supported viewport ranges, zoom and text-scaling requirements, and input methods: [paste], create responsive compatibility scenarios. Include narrow and wide extremes, portrait and landscape, virtual keyboards, touch targets, hover absence, keyboard navigation, focus visibility, sticky elements, safe areas, overflow, and evidence checkpoints.” Do not test only pristine English copy on a fashionable phone. Use long names, validation errors, empty states, large values, translated strings, browser zoom, and increased text size. Watch for clipped actions, horizontal scrolling, overlays behind keyboards, unreachable controls, and content hidden beneath fixed headers. Compatibility and accessibility overlap, but one does not replace the other. Use AI accessibility testing prompts for semantic, assistive-technology, contrast, and keyboard analysis with qualified human review. 6. Plan operating-system and real-device checks “Given these supported operating systems, device classes, hardware constraints, permissions, file and camera flows, notifications, biometric or authentication features, installed-app handoffs, and available labs: [paste], draft a real-device test plan. Distinguish what emulators can cover from evidence requiring physical devices. Include setup fingerprints, battery and memory observations, permission states, interruptions, orientation, backgrounding, and cleanup.” Emulators and cloud grids are useful for breadth and repeatability. They are weaker evidence for cameras, sensors, Bluetooth, thermal throttling, memory pressure, keyboards, browser chrome, installed-app handoffs, biometrics, and vendor-specific behavior. You do not need a museum of every handset. Select devices that represent meaningful hardware, OS, viewport, engine, and user segments. Record exact versions and settings. “Tested on Android” is roughly as precise as “tested somewhere on Earth.” For native and hybrid products, adapt the deeper AI mobile app testing prompts. 7. Cover locale, language, and timezone variation “Using these supported locales, languages, writing directions, currencies, number and date formats, timezone rules, content limits, and critical journeys: [paste], create localization compatibility checks. Include text expansion, wrapping, truncation, sorting, pluralization, decimal input, currency display, daylight-saving boundaries, date rollover, right-to-left layout, fallback fonts, and untranslated content.” Locale bugs are not merely cosmetic. A date interpreted as month-first instead of day-first can alter a booking. A decimal separator can change money. A timezone boundary can move a deadline. A missing glyph can hide a person’s name. Use representative translated content reviewed by people who understand the language and product context. Machine-generated strings can exercise layout, but they do not certify translation quality. Separate linguistic review from technical compatibility evidence. 8. Test networks, blockers, and third-party dependencies “Using these supported network conditions, offline behavior, CDN design, authentication, payment, analytics, consent, content blockers, corporate proxies, and third-party dependencies: [paste], draft compatibility scenarios for slow, intermittent, blocked, and recovering connections. Include observable user behavior, retry and idempotency checks, timeout expectations, fallback paths, duplicate-action prevention, evidence, and safe dependency limits.” Users switch networks, lose signal, block trackers, reject cookies, use private browsing, and work behind restrictive corporate systems. The product should fail clearly and recover safely. Test interrupted submissions and repeated clicks. A retry must not create duplicate payments, records, or messages. Do not hammer real vendors without permission. Use approved sandboxes and controlled failure injection. A stub can verify your fallback logic, but it cannot prove the real integration works in every environment. Label that limitation. 9. Triage a sanitized compatibility defect “Using this sanitized defect evidence—product build, browser and OS versions, device, viewport, settings, account state, exact steps, expected and observed behavior, screenshots, console and network summaries, reproduction rate, and comparison environments: [paste]—create a compatibility defect brief. Separate facts from hypotheses, identify the smallest reproduction matrix, suggest evidence to collect, and list safe isolation experiments. Do not invent a root cause.” Compare one variable at a time where practical: same browser on another OS, another engine on the same OS, private mode, extensions disabled, clean account, different viewport, or previous build. This narrows the cause without converting correlation into certainty. Keep raw logs and screenshots in approved systems. Use AI bug report prompts to format verified evidence for engineering, but remove identities, tokens, private URLs, and customer content before prompting. 10. Turn results into a release brief “Using only these verified compatibility results, approved support policy, environment fingerprints, passed and failed journeys, defect severities, user impact, workarounds, evidence links, coverage gaps, and owner decisions: [paste], draft a compatibility release brief. Include scope, exclusions, failures by supported tier, unresolved risk, remediation owners, retest plan, monitoring, support communication, rollback triggers, and a decision table. Do not invent results or make the final decision.” Keep observation and decision separate. “Checkout fails on supported Safari 18 when returning from the payment provider” is evidence. “Delay release” depends on affected users, workaround, severity, contractual promises, timing, remediation confidence, monitoring, and rollback. A green matrix also has limits. State whether evidence came from physical devices, cloud browsers, emulators, or automation. Name untested combinations. Compatibility testing reduces uncertainty; it does not prove absence of defects. A workflow that produces evidence instead of matrix theater Define support. Agree which environments receive full, best-effort, or no support. Sanitize the context. Remove secrets, private analytics, customer data, and confidential screenshots. Prioritize risk. Combine usage, business impact, technical change, history, and obligations. Design the matrix. Cover meaningful engines, operating systems, devices, viewports, inputs, locales, and networks. Choose critical journeys. Test user outcomes rather than every screen equally. Fingerprint environments. Record build, browser, OS, device, viewport, settings, flags, and account state. Run real checks. Use automation for repeatability and humans for visual, behavioral, assistive, and exploratory evidence. Capture proof. Store expected and observed behavior, screenshots, video, console, network, and reproduction rate safely. Fix and retest. Verify the original configuration and nearby combinations; check regressions. Make the decision. Humans weigh evidence, users, commitments, mitigations, support burden, and rollback. Use AI UI testing prompts for deeper visual and interaction coverage. Use AI performance testing prompts when speed, stability, or resource behavior varies across environments. Compatibility can reveal those risks, but it does not replace specialized testing. Common ways compatibility testing goes wrong Building the matrix from vibes The loudest stakeholder, newest phone, or developer’s browser is not a support strategy. Ground choices in policy, users, impact, technical risk, history, and obligations. Counting browser brands instead of engines Ten Chromium browsers do not provide the same diversity as testing Chromium, WebKit, and Gecko. Brands still matter when integrations or enterprise policies differ, but engine coverage should be explicit. Treating emulators as physical-device proof Emulators are excellent tools, not magical certificates. Document what they covered and schedule real-device checks for hardware, performance, input, permissions, interruptions, and high-risk journeys. Testing screenshots instead of outcomes A page can look correct while keyboard focus disappears, a payment callback fails, a date shifts, or a file never uploads. Validate complete journeys and persisted state. Ignoring zoom, text scaling, and long content A layout that survives one viewport may collapse under 200% zoom, larger system text, translated copy, validation messages, or a virtual keyboard. Include realistic extremes. Calling automation complete coverage Automation repeats known assertions efficiently. It can miss visual clipping, confusing fallback behavior, assistive-technology output, device-specific interaction, and new failure modes. Combine it with human exploration. Pasting sensitive evidence into public AI tools Analytics, screenshots, logs, URLs, and network traces can expose customer identities, tokens, unreleased features, architecture, and business data. Redact aggressively and follow approved tooling rules. Understanding what AI can and cannot do makes the boundary simpler: generation and organization are fast; real-world execution, evidence, judgment, and accountability remain human work. Frequently asked questions Can AI perform compatibility testing automatically? AI can help orchestrate approved browser automation or device-grid tools when engineers deliberately connect it, but a chat response has tested nothing. Connected automation still needs a reviewed matrix, secure environments, reliable assertions, stored evidence, and human interpretation. What is the best AI prompt for cross-browser testing? Include the named product and build, support policy, critical journeys, browser engines, OS versions, device classes, recent changes, test inventory, expected behavior, evidence requirements, and exclusions. Require assumptions and gaps to be explicit and forbid invented results. How many browsers and devices should I test? There is no universal number. Select a risk-based set using actual audience evidence, browser-engine diversity, critical journeys, technical changes, business impact, support commitments, accessibility needs, and available evidence. Add configurations when uncertainty or consequences justify them. Is browser emulation enough? It is enough for some layout checks and broad automation, but not all compatibility evidence. Physical devices matter for real browser chrome, keyboards, touch, sensors, permissions, memory, performance, interruptions, installed-app handoffs, and vendor-specific behavior. Is compatibility testing the same as accessibility testing? No. They overlap because browsers, operating systems, input methods, zoom, and assistive technologies affect access. Accessibility testing also evaluates semantics, keyboard behavior, screen-reader output, contrast, cognition, and standards. Use qualified human review rather than treating one matrix as certification. Should I test old browser versions? Test versions covered by the approved support policy and any high-risk versions justified by users, contracts, controlled environments, or known defects. Define an aging policy so support decisions are deliberate. Never drop a version solely because an AI suggested it. Can analytics tell me exactly what to support? Analytics informs the decision but can be incomplete, consent-limited, biased, or unable to record clients that fail early. Combine it with support tickets, contracts, market context, product obligations, technical risk, and a safety set of representative environments. Can AI decide whether a compatibility defect blocks release? No. It can organize verified evidence against severity and support rules. Humans must judge user impact, affected tiers, obligations, workarounds, remediation, monitoring, support burden, rollback, timing, and accountability. The useful boundary AI is good at turning scattered requirements and environment lists into a structured compatibility plan. It can generate variations, find obvious gaps, and format evidence for review. That saves time. It is bad at knowing whether the support policy is fair, whether a cloud browser reproduces a real customer’s setup, whether a screenshot proves the journey works, or whether remaining risk is acceptable. Those are the valuable parts: context, observation, taste, and responsibility. Use the prompts to accelerate planning and analysis. Keep sensitive evidence, real execution, diagnosis, remediation, and release authority with accountable people. That is the broader argument in Dmitry Kargaev’s Don’t Replace Me: use the machine for speed, then apply the human judgment it cannot supply. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance AI API Testing Prompts: Catch Broken Integrations AI API testing prompts for contracts, edge cases, authorization, retries, error responses, and human-verified integration evidence. AI promptsAPI testingquality assurance AI Exploratory Testing Prompts: Find Hidden Bugs AI exploratory testing prompts for focused charters, unusual user journeys, sharper observations, reproducible bugs, and human-owned release decisions. AI promptsexploratory testingquality assurance --- ## AI Conflict Resolution Prompts: 10 Templates URL: https://dontreplace.me/articles/ai-conflict-resolution-prompts-say-the-hard-thing ai conflict resolutionworkplace conflict promptsai at workdifficult conversationsconflict management AI Conflict Resolution Prompts: 10 Templates 10 copy-paste AI conflict resolution prompts for workplace disputes. Prep hard conversations, rewrite heated messages, and document agreements. June 27, 2026 Updated June 27, 2026 Conflict at work is uncomfortable enough without typing "help me not sound like a lunatic" into ChatGPT at 11pm. But that's exactly what people do. And honestly? Done right, AI conflict resolution prompts are one of the most underrated uses of these tools. Not because AI understands your coworker. It doesn't. But because getting your thoughts out of your head and onto a page, in a structured way, before you send the message that ends your career, is genuinely useful. The AI is the rehearsal. You're still the one who has to walk into the room. Here's the full playbook: what works, what doesn't, when to stop entirely, and 10 copy-paste AI conflict resolution prompts you can use today. What AI can actually do in a conflict (and what it can't) AI is good at one thing in conflict situations: helping you think more clearly than your nervous system will allow in the moment. When you're angry, your brain narrows. You remember every slight, forget every nuance, and write emails that would make HR cry. AI can slow that down. It can take your one-sided rant and reflect back something closer to neutral facts. It can help you separate what happened from what you assumed. It can draft a follow-up that doesn't sound like it was written by someone who hasn't slept. What it can't do is read the room. It doesn't know your company's culture, your manager's triggers, the history between you and this person, or the power dynamics in play. It invents plausible-sounding motives and flattens complex relationships into tidy paragraphs. This is one of the core warnings in Don't Replace Me: garbage in, garbage out. Feed AI a one-sided rant and you get a cleaner, more articulate one-sided rant. More dangerous, not less. Use it as a thinking tool. Not as a mediator. When to stop and escalate instead Before the prompts, the most important thing in this article: some conflicts don't belong in a chatbot. Do not paste any of the following into an AI tool, approved or not: employee health or disability information, harassment or discrimination complaints, retaliation claims, active legal disputes, HR investigation details, security incidents, customer PII, contracts, credentials, confidential strategy, or performance records covered by company policy. If your conflict involves any of these, stop. Go to your HR team, your legal department, your manager, or your company's employee assistance program. These are not situations where smoother phrasing solves anything. They're situations where the wrong move creates serious personal and legal exposure, for you, for your company, and sometimes for the person you're in conflict with. Same goes for anything involving safety, threats, or a colleague's wellbeing. AI can't protect anyone. A human can. Knowing the difference between a problem AI can help with and one it can't is probably the most important skill you'll build this year. The reusable AI conflict resolution prompt formula Every prompt in this list follows the same structure. You can use this formula to build your own: [Context] Here is the situation: [what happened, using neutral facts, no interpretations]. [Relationship] My relationship with this person is [colleague/direct report/manager/client] and we have [describe relevant history]. [Goal] I want to [what outcome you're seeking]. [Constraints] I need to avoid [what you can't say or do]. [Ask] Please help me [specific task: draft a message / prepare questions / role-play pushback / etc.]. That's it. The more specific you are, the more useful the output. The more you vent without structure, the more useless (and potentially harmful) the output becomes. One thing worth knowing: ChatGPT and Claude will both work for these prompts. Claude tends to be better at nuanced, emotionally sensitive drafting. ChatGPT is slightly more direct. Neither one will tell you you're wrong when you are, so build that skepticism in yourself. Ask explicitly: "What might I be missing here?" and "What's the strongest argument for the other person's position?" You'll get better material. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI conflict resolution prompts you can actually use 1. Turn a rant into neutral facts You've written the email. It's not good. Paste it here instead. "I've written a message about a conflict I'm having at work. Before I send anything, help me rewrite it as a neutral list of observable facts only: what was said, what was done, and when. Remove interpretations, emotional language, assumptions about intent, and anything I can't directly verify. Here is my draft: [paste draft]." 2. Separate facts from interpretations Useful before any difficult conversation, especially if you're not sure what you actually know versus what you're assuming. "I'm preparing for a difficult conversation. Help me separate this situation into two columns: (1) verifiable facts I can confirm, and (2) my interpretations or assumptions about what those facts mean. Here's what happened: [describe situation in plain language]." 3. Prepare for a difficult conversation Use this before a 1:1, a performance discussion, or a tense meeting. For more on prepping hard manager conversations, the AI manager prompts guide has related templates. "I need to have a difficult conversation with [colleague/manager/report] about [topic]. The specific issue is [factual description]. My goal is [what I want to happen after]. Help me prepare: draft an opening statement, three clarifying questions I should ask, and what I should listen for. Don't assume the other person's motives." 4. Rewrite a heated message For when you've typed something that would absolutely be screenshotted. "Here is a message I drafted but haven't sent. Rewrite it so it: states the issue clearly and factually, doesn't assign blame or accuse, makes a specific request, and leaves room for dialogue. Don't make it passive-aggressive or corporate-sounding. Here is the original: [paste message]." 5. Name impact without blame This is the difference between "you made me feel" and "here's what happened and here's the effect." Much harder to argue with. "Help me write a short statement that describes the impact of someone's behavior without attributing intent or blame. The behavior was: [what happened]. The impact on me or my work was: [what effect it had]. I want to name this directly without making it an accusation." 6. Map both sides' needs Before you go in assuming the other person is just wrong, this prompt forces some useful perspective-taking. "I'm in a conflict with [role/relationship] about [topic]. Here's my perspective and what I need: [your view]. Here's what I know or can reasonably assume about their perspective and what they might need: [what you know about their situation]. Help me identify where our needs overlap and where they conflict, so I can look for solutions that address both." 7. Role-play the pushback Possibly the most useful prompt in the list. You need to hear the hard version before you're in the room. See also: AI decision-making prompts for pressure-testing assumptions before committing to a course of action. "I'm going to have a difficult conversation where I raise [topic]. I want you to role-play as the other person and push back hard on what I'm saying. Be realistic, not a cartoon villain. Use the following background on their likely perspective: [what you know]. I'll go first." 8. Draft a repair or apology note For when you were, in fact, the problem. "I need to write a genuine apology to [role/relationship] for [what happened]. I want the apology to: acknowledge specifically what I did, name the impact it had without minimizing it, not include excuses or justifications, and state what I'll do differently. Help me draft this. Here is the context: [brief description]." 9. Document agreements after a conflict Conflict resolution doesn't end when the conversation does. Write it down. "I just had a difficult conversation with a colleague and we reached some agreements. Help me write a clear, neutral summary of what was decided that I can send as a follow-up email. The agreements were: [list what was agreed]. The tone should be professional and confirmatory, not punitive. I'll send this for their review before treating it as final." 10. Decide whether to escalate This prompt won't make the decision for you. But it'll help you think it through. "I'm dealing with a workplace conflict and I'm not sure whether to handle it directly, involve my manager, or escalate to HR. Here is the situation in neutral terms: [factual description]. Help me think through: what the options are, what the risks are in each direction, and what questions I should be asking before I decide. Don't give me legal or HR advice." Common mistakes people make with AI conflict resolution prompts These prompts only work if you use them honestly. Most people don't, at least not at first. The most common mistake is using the AI to validate your position rather than examine it. You describe the situation in a way that makes you look reasonable and the other person look like an idiot, and then you're surprised when the output confirms that you're reasonable and the other person is an idiot. That's not AI helping you. That's AI reflecting your own framing back to you with better grammar. The second mistake is skipping the facts-from-interpretations step. It feels slow. It isn't. If you can't articulate what specifically happened without editorializing, you're not ready to have the conversation. The AI can't save you from that, but it can force you to slow down if you let it. The third mistake is using AI to avoid the conversation entirely. Someone writes a message, AI polishes it, they send it, the other person responds, they paste that into AI, AI suggests a reply, and suddenly they've had three rounds of "conflict resolution" without ever having an actual conversation. This produces nothing except the illusion of resolution and a record that will confuse everyone later. Use the prompts to prepare. Then actually talk to the person. A useful comparison here: workplace mediation is built around direct conversation, not endless drafted messages. Even when writing helps you prepare, the AI draft only gets you ready. It doesn't replace the human moment. When AI conflict preparation actually pays off There are three situations where these prompts make a real, measurable difference. The first is when you're dealing with a power imbalance. If you need to address something with your manager, or someone more senior, or a major client, the cost of getting the tone wrong is high. You can't just fire off the first thing that comes to mind. Taking 20 minutes to prepare your opening, anticipate their response, and draft a follow-up note can be the difference between a resolved issue and a damaged relationship that affects your next performance review. The second is when you've been ruminating. You've been replaying the incident for three days, you've told four friends about it, and now your version of events has calcified into something that bears only a passing resemblance to observable reality. Running it through a facts-versus-interpretations prompt is uncomfortable precisely because it works. It surfaces the parts of your narrative that are made up. The third is cross-functional or cross-cultural conflicts. Different teams have different norms. Different backgrounds mean different assumptions about directness, hierarchy, and what "professional" looks like. AI won't perfectly capture cultural nuance, but asking it explicitly to flag where your communication style might land differently than intended is a useful check. The WEF's Future of Jobs report consistently identifies interpersonal and cross-functional collaboration skills as among the most durable human competencies. That's not coincidental. These are exactly the situations where human judgment, not AI output, determines the outcome. What you still have to do yourself AI can draft the message. It can't send it. More importantly, it can't sit across from another human being and mean what it says. This matters more than the words. In conflict situations, the words matter less than whether the other person believes you actually mean them. A perfectly worded apology from someone who clearly doesn't mean it is worse than a clumsy one that's real. AI can help you find the words. It can't manufacture the intention behind them. For anything involving clients, the AI client communication prompts guide walks through a similar approach for external relationships, where the stakes and the tone are different from internal ones. Before you send anything AI helped you write, read it back out loud. Ask yourself: does this sound like me? Does it say what I actually mean? Would I be comfortable if the other person knew I used AI to draft it? If the answer to any of these is no, rewrite it. The goal isn't a perfect document. The goal is a real conversation. The best conflict resolution prompt in the world is still just preparation. You're the one who has to show up. Frequently asked questions Can AI help me resolve a workplace conflict? AI can help you prepare for difficult conversations by clarifying facts, separating interpretations from observations, drafting messages, and role-playing pushback. It can't resolve the conflict itself, and it doesn't know your workplace, your relationships, or the power dynamics involved. The conversation still has to happen between humans. Is it safe to paste conflict details into ChatGPT or Claude? It depends on what's in those details. Never paste employee health information, harassment or discrimination claims, HR investigation details, legal disputes, customer PII, performance records covered by company policy, or security incidents into any unapproved AI tool. Check your company's AI usage policy before you paste anything sensitive. What's the difference between using AI conflict resolution prompts versus just venting to a chatbot? Venting gives you a more articulate version of your original position. Actually using AI conflict resolution prompts for real prep means giving structured, factual input and asking it to help you see both sides, prepare questions, or draft something neutral. The quality of the output depends entirely on how honestly and specifically you describe the situation. When should I escalate a workplace conflict instead of trying to handle it myself? Escalate to HR, legal, or your manager when the issue involves harassment, discrimination, retaliation, safety threats, legal exposure, protected employee information, or anything your company's policy requires to be reported. AI can't make that call for you, and attempting to handle these situations with better phrasing is not the right move. How do I make sure an AI-drafted message still sounds like me? Read it out loud before you send it. Cut anything that sounds like corporate-speak or that you'd never actually say to someone's face. Add back specifics that only you would know. The AI draft is a starting point, not a final product. Can I use these prompts for conflicts with clients, not just colleagues? Yes, with the same caveats. Don't include confidential client information, contract details, or anything covered by an NDA. The AI client communication prompts guide has templates specifically built for external relationships where the tone and power dynamics are different from internal ones. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Automation Prompts: Automate the Boring Stuff 10 copy-paste AI automation prompts for mapping workflows, writing SOPs, adding approval gates, and QA testing. Practical templates with safety rules included. ai automation promptsworkflow automationSOP templates AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts --- ## AI Contingency Plan Prompts: 10 Templates URL: https://dontreplace.me/articles/ai-contingency-plan-prompts-10-templates-that-work contingency planningAI promptsbackup plansrisk managementproject management AI Contingency Plan Prompts: 10 Templates 10 copy-paste AI contingency plan prompts for ChatGPT and Claude. Map failures, assign owners, draft comms, and verify assumptions before a crisis hits. July 25, 2026 Updated July 25, 2026 A contingency plan that lives only in someone's head fails the moment that person is on a plane. A contingency plan that exists as a tidy document but was never stress-tested fails about 10 minutes into the actual crisis. AI contingency plan prompts give you a faster way to build the structure, map the scenarios, and draft the comms. What they don't give you is a substitute for knowing your actual constraints, your actual owners, and your actual budget. That gap is the whole game. AI is fast. It will generate a five-scenario failure tree in 90 seconds. Every scenario will sound plausible. Some will be completely wrong for your situation, and you won't know which ones until you check them against reality. So use these prompts to do the heavy lifting on structure, then do the human work of verifying every assumption before you call it a plan. Here's how to do that without wasting your afternoon. What makes AI contingency plan prompts actually useful? The short answer: they're useful for structure, not substance. AI is genuinely good at generating scenario lists, turning vague worries into explicit trigger criteria, drafting stakeholder comms, and surfacing gaps you forgot to think about. It's dangerous when it invents operational facts, assumes resources you don't have, or implies that having a document equals having a plan. Rule 13 from Don't Replace Me applies here: garbage in, garbage out. If you paste in vague assumptions, you get a polished backup plan built on sand. Feed the AI your actual constraints: budget limits, approved tooling, real team headcount, timing windows, known dependencies. The output gets dramatically more useful when it has something real to work with. One more thing before you start. Don't paste customer PII, credentials, access tokens, raw production logs, private tickets, unreleased strategy, legal disputes, security vulnerabilities, financial records, HR issues, regulated data, medical information, or private client conversations into unapproved AI tools. Describe the situation in general terms. You don't need to name the client to draft a customer-impact response plan. The reusable prompt formula for contingency planning Every prompt in this list follows the same skeleton. Learn it once, adapt it for any situation. "I'm building a contingency plan for [specific project or process]. The context is [brief description]. My constraints are [budget / timeline / team size / approved tools]. The failure scenario I'm most worried about is [specific risk]. Help me [specific output: map scenarios / assign owners / draft stakeholder message / compare tradeoffs / etc.]. Flag any assumptions that need verification before this goes live." That last sentence matters more than the rest. Asking AI to flag its own assumptions forces it to surface uncertainty instead of hiding it inside confident-sounding prose. The other thing worth building into every prompt: ask for priority order. A list of 10 failure scenarios without any indication of which one deserves your attention first is a list you won't act on. Tell the AI to rank outputs by likelihood, by severity, or by your ability to actually control the outcome. You'll cut the analysis time in half. 10 copy-paste AI contingency plan prompts Prompt 1: Map failure scenarios I'm building a contingency plan for [project name or process]. The main deliverable is [what success looks like]. Known dependencies include [list 3-5 dependencies]. Generate a list of 8-10 realistic failure scenarios, ordered from most to least likely. For each scenario, note the likely trigger, the team most affected, and whether it's recoverable without external escalation. Flag any scenarios where you've made assumptions about our environment. Use this before anything else. You want a full failure map before you start assigning owners or drafting comms. Check every scenario against what you know about your actual environment. Delete the ones that don't apply. Add the ones it missed. The output will typically cluster into three categories: execution failures (things your team controls), dependency failures (things your vendors or partners control), and assumption failures (things that turn out not to be true). Knowing which cluster a scenario falls into tells you who needs to own the response. Prompt 2: Turn assumptions into trigger criteria Here are the key assumptions behind [project or plan]: [list 5-10 assumptions]. For each assumption, suggest a measurable trigger criterion that would indicate the assumption has failed, a monitoring method to detect it early, and a suggested response window (immediate, same day, next business day). Note where verification requires human judgment rather than a metric. Assumptions dressed up as facts are how plans fail quietly. This prompt forces them into the open and attaches a tripwire to each one. Have your actual monitoring owner review the detection methods before you publish anything. The specific value here is in the response window column. Most teams treat all failures as equally urgent. They're not. A monitoring alert that fires at 2am for a non-critical process creates alert fatigue and gets ignored. Building in explicit response windows trains people to take the right things seriously. Prompt 3: Identify critical dependencies The process I'm planning around is [description]. Key inputs include [list inputs]. Key outputs include [list outputs]. Map the critical path dependencies, identify single points of failure (one owner, no backup, no alternative supplier), and flag which dependencies involve third parties whose response times we don't control. Mark any dependencies where you don't have enough context to assess risk accurately. Third-party dependencies are where contingency plans go to die. If your plan assumes a vendor will respond in four hours, you need to verify that SLA exists. AI can't check that. You can. Pay particular attention to the "response times we don't control" flag. That list is your actual risk surface. Those are the dependencies that need either a real backup option or an honest conversation with leadership about what happens if they fail. Prompt 4: Create fallback options My primary plan for [specific goal] involves [Plan A description]. If Plan A fails due to [most likely failure reason], I need at least two viable fallback options. My constraints are: budget [amount or range], timeline [hard deadline], approved tools [list], available team [headcount or roles]. For each fallback, estimate effort, estimated timeline impact, cost delta, and key risk. Flag which fallbacks require approvals I may not have yet. Fallbacks that require new budget approvals, new vendor contracts, or executive sign-off aren't ready fallbacks. They're ideas. This prompt surfaces that distinction before the crisis happens. The cost delta column is worth extra scrutiny. AI will often suggest fallbacks that sound reasonable but carry hidden costs: overtime, expedited shipping, premium support tiers, licensing for tools you don't currently have. Check each cost estimate against your actual procurement process before you commit to calling it a fallback. Prompt 5: Assign owners and backup owners Here are the key response actions in my contingency plan: [paste list of actions]. For each action, suggest: a primary owner role (not name, just role title), a backup owner role in case the primary is unavailable, an estimated response time from trigger to action, and what that owner needs to execute (access, tools, information). Flag any actions where a single person is both primary owner and backup owner. Single-person ownership is the same as no contingency. This prompt finds it. You fill in the actual names and verify the access requirements with your real team. There's a version of this that bites teams regularly: the person who knows how to do the thing is also the person who built the thing, which means they're also probably the person most likely to be unavailable when the thing breaks. Redundancy only works if the backup has actually practiced the action, not just been listed as backup. Prompt 6: Draft stakeholder escalation messages A [severity: low / medium / high / critical] incident has occurred: [describe the situation in general terms, no PII or confidential data]. Draft three stakeholder update messages: one for internal leadership (direct, factual, no spin), one for affected customers (clear, empathetic, specific about impact and timeline), one for the wider team (calm, directive, focused on what to do now). Each message should include: what happened, what we're doing, what stakeholders should and shouldn't do, and when we'll update them next. Flag any phrases that might create legal or compliance exposure. You'll need to edit these before sending. But having a draft in the first five minutes of a crisis is worth more than a perfect message 90 minutes in. See also the AI escalation plan prompts for a deeper template set on incident comms. The internal leadership message and the customer message need different editing passes. Leadership wants completeness. Customers want clarity. What's reassuring in a leadership briefing can sound evasive in a customer email, and vice versa. Don't use the same tone for both. Prompt 7: Compare Plan A vs. Plan B tradeoffs I'm deciding between two approaches for handling [situation]: Plan A is [description]. Plan B is [description]. My decision criteria are: speed of execution, cost, team disruption, customer impact, reversibility, and regulatory or compliance risk. Build a comparison table scoring each plan on these criteria, with a brief rationale for each score. Flag criteria where you don't have enough information to score accurately, and note which criteria require human or expert judgment. Comparison tables generated by AI are starting points, not verdicts. The scores are illustrative. Your actual constraints are the only thing that settles the decision. If your team keeps revisiting decisions that were supposedly already made, adding reversibility as a scored criterion helps. Teams undervalue reversibility when things are going well and desperately wish they'd prioritized it when things aren't. A plan you can undo is worth more than a slightly better plan you can't. Prompt 8: Build a resource-shortage contingency My plan assumes [X amount of resource: people / budget / infrastructure / tooling]. If that resource is reduced by 30% or becomes unavailable, I need a contingency. Suggest: what to cut first (minimum viable scope), what to cut only if necessary (important but recoverable), what must be protected at all costs, and what the customer-facing impact of each cut would be. Note where any of these decisions require management approval before action can be taken. Resource-shortage contingencies are the ones nobody writes until they're in the middle of a shortage. Write this one in advance. Verify the approved decision authority before you file it. The 30% reduction number is a useful default, but run it at 50% too. A 30% reduction is a bad month. A 50% reduction is a reorganization, a budget freeze, or someone resigning at a critical moment. Plans that only survive mild disruption aren't contingency plans. Prompt 9: Create a customer-impact response plan [Describe the customer-facing failure scenario in general terms. No PII, no account details, no confidential client data.] Draft a customer-impact response plan that includes: who is affected and how (segment and nature of impact), what we will communicate and when, what remediation or compensation options exist, what support needs to be ready before we communicate anything, and what success looks like 48 hours after the incident. Flag any steps that require legal or compliance review before customer communication goes out. Don't communicate to customers before support is ready to handle the volume. This prompt reminds you to sequence it correctly. Your legal and compliance teams get final review before anything goes out. The "what support needs to be ready" step is the one most teams skip. You announce the problem, customers flood your support channels, and suddenly you have two crises instead of one. Sequence matters more than speed in customer-facing incidents. Getting it out an hour later with support staffed is better than getting it out immediately with no one to answer questions. Prompt 10: Convert lessons into prevention notes Here's what happened during [incident or near-miss]: [brief description, no sensitive details]. The contributing factors were [list what you know]. Convert this into a prevention note that includes: the specific conditions that created the risk, what early warning signs were visible in retrospect, what process or ownership change would prevent recurrence, and whether this is a one-time fix or a systemic pattern. Draft it for a team retrospective, not a blame document. Post-incident, most teams produce a timeline and a vague "we'll do better" paragraph. This prompt turns that into something the next project team can actually use. For the broader retrospective format, the AI retrospective prompts article covers that in full. The "one-time fix vs. systemic pattern" distinction is where most post-incident work breaks down. Teams fix the specific thing that broke. They don't ask whether the same failure mode exists in five other places. AI is actually useful here because it can take one incident description and suggest analogous failure points elsewhere in a system, if you give it enough context about how that system works. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What AI can't do in contingency planning AI can't verify that your budget approval is real. It can't check whether a vendor SLA actually covers your failure scenario. It can't tell you whether the backup owner has the system access they need, or whether your customer communication will create a legal problem in a regulated market. These aren't edge cases. They're the most common reasons contingency plans fail. The document looks complete. The assumptions underneath it were never checked. Someone filed it, no one tested it, and the first real incident proved that the plan described a version of the organization that doesn't exist. The fix is straightforward: every assumption in the plan needs a named verifier and a verification date. Not AI-generated. Not "TBD." A real person who confirmed the thing is true as of a specific date. When the plan is six months old and something goes wrong, you want to know which assumptions were verified in March and which ones were just copied over from the previous version. If you're working on a project with launch risk, the AI launch checklist prompts are a useful companion to this article. And if you're in the middle of a migration, the AI migration plan prompts have a parallel structure for that context. Every contingency plan needs a final human sign-off from someone who has actually read it and is accountable for the outcome. AI generates the draft. You own the decision. The audit trail matters too. When something goes wrong (and eventually something will), you want a record of when the plan was created, who reviewed it, what assumptions were verified, and what approvals were obtained. A document with no version history and no named reviewer isn't a plan. It's a hope. Frequently asked questions What are AI contingency plan prompts? AI contingency plan prompts are structured instructions you give to tools like ChatGPT or Claude to generate contingency planning documents: failure scenario maps, fallback options, owner assignments, stakeholder messages, and tradeoff comparisons. They speed up the structural work but require human verification of all assumptions, constraints, and approvals before use. Can AI write a contingency plan for me? AI can draft the structure of a contingency plan quickly, but it can't verify your actual constraints, confirm your budget approvals, check vendor SLAs, or ensure compliance with legal or regulatory requirements. Treat the AI output as a first draft that needs human review before it becomes an operational plan. What should I never paste into an AI tool when building a contingency plan? Don't paste customer PII, credentials, access tokens, raw production logs, unreleased strategy documents, legal disputes, security vulnerabilities, financial records, HR records, regulated data, medical information, or private client conversations. Describe situations in general terms to protect sensitive information. How do I make AI contingency plan prompts more specific to my situation? Include your actual constraints in the prompt: budget range, hard deadlines, approved tooling, real team headcount, and known dependencies. The more specific your inputs, the more useful the output. Asking AI to flag its own assumptions is also worth adding to every prompt. What's the difference between a contingency plan and a rollback plan? A contingency plan covers a range of failure scenarios and defines responses, owners, and fallback options before a crisis happens. A rollback plan is specifically about reversing a deployment or change after something goes wrong. They overlap but aren't the same. The AI rollback plan prompts cover the rollback use case in detail. Do I need to do a risk assessment before building a contingency plan? Yes. A contingency plan built without a prior risk assessment is guessing at which failures to prepare for. The AI risk assessment prompts are the logical first step before you use any of the prompts in this article. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Dependency Register Prompts: Keep Projects From Surprising You Use AI dependency register prompts to map owners, blockers, assumptions, and downstream impact before one small change becomes a three-week incident. AI dependency mappingproject managementAI prompts AI Assumption Log Prompts: Stop Planning on Guesswork Use 10 AI assumption log prompts to expose project guesswork, design evidence checks, assign review owners, and stop unverified beliefs becoming commitments. AI promptsproject managementassumption logs AI Dependency Mapping Prompts: Find What Breaks AI dependency mapping prompts for finding upstream work, downstream risks, hidden handoffs, owners, schedule conflicts, and launch blockers. AI promptsdependency mappingproject management --- ## AI Customer Interview Prompts: Learn Without Leading URL: https://dontreplace.me/articles/ai-customer-interview-prompts-learn-without-leading AI promptscustomer interviewscustomer discoveryuser researchproduct research AI Customer Interview Prompts: Learn Without Leading 10 AI customer interview prompts for research goals, neutral questions, follow-ups, notes, synthesis, and findings—without manufacturing customer truth. August 23, 2026 Updated August 23, 2026 Customer interviews look simple: ask people what they want, write down the answer, then build it. That is also how teams spend six months shipping a feature five unusually polite customers said sounded “interesting.” AI customer interview prompts can help turn a fuzzy assumption into a research question, draft a neutral guide, critique leading language, prepare follow-ups, organize de-identified notes, and surface candidate themes. They cannot recruit a representative group, build trust, read discomfort reliably, verify what someone meant, or decide what your team should build. AI can help prepare and organize the interview. Humans still earn consent, conduct the conversation, protect participants, interpret the evidence, and own the decision. These ten templates are for product managers, UX researchers, founders, marketers, customer-success teams, and service designers who want evidence rather than customer-flavored fan fiction. For broader source work, start with AI research prompts. If a structured questionnaire is the better method, use these AI survey design prompts. What customer interviews can actually tell you An interview is good at revealing context: what someone recently tried, what happened, where they struggled, what workaround they invented, and why the situation mattered. It is weaker at estimating prevalence, predicting future purchases, or proving that one product change caused an outcome. A useful interview needs: One learning goal: the uncertainty the team needs to reduce. A real decision: what might change after the research. Relevant participants: people with recent experience of the behavior. Neutral questions: no answer hidden inside the wording. Concrete recall: recent events instead of fantasy futures. Responsive follow-ups: curiosity, not a robotic script. Consent and privacy: clear rules for recording, storage, and use. An evidence trail: observations separated from interpretation. Visible limitations: who was missing and what remains unknown. Interviews do not produce “the customer’s opinion.” They produce accounts from specific people, in a specific context, speaking to a specific interviewer. Courtesy bias, memory limits, recruitment bias, incentives, power differences, and interviewer behavior all affect the result. If you need to observe someone using a product, pair interviews with AI usability testing prompts. Watching a participant fail to find a button is different evidence from asking whether navigation feels easy. The reusable customer interview prompt formula Add this instruction to any template below: “Act as a customer-research preparation assistant. I am investigating [learning goal, decision, audience, recent behavior, existing evidence, recruitment method, constraints, and deadline]. Use only verified information I provide. Produce [artifact] with assumptions, exclusions, open questions, privacy risks, accessibility needs, and required human review. Separate participant statements, observed behavior, interviewer notes, interpretations, and decisions. Do not invent participants, quotes, prevalence, motivations, consensus, or confidence.” That separation prevents a common mess. “Participant 4 said setup took two hours” is a statement. “Setup is too slow” is an interpretation. “Rebuild onboarding” is a decision. Each step needs more evidence and judgment than the one before it. Never paste names, email addresses, recordings, identifiable transcripts, health or financial details, employer-confidential information, credentials, unreleased strategy, or legally sensitive comments into an unapproved AI tool. Removing a name does not automatically anonymize a distinctive story. Use approved research systems, explicit consent, access controls, retention limits, synthetic examples, and human privacy review. What to collect before prompting “Write customer discovery questions” is an invitation for generic sludge. Give the model a compact research brief instead. InputWhy it mattersHuman check Decision to informStops research theaterDecision owner confirms Learning goalDefines the uncertaintyResearch lead reviews Recent behaviorAnchors questions in realityDomain owner verifies Participant criteriaConnects people to the questionResearcher approves Existing evidenceAvoids asking settled questionsEvidence owner checks Recruitment channelExposes selection biasResearcher documents Sensitive topicsTriggers safety planningPrivacy/legal reviews Recording planSets consent and storage rulesParticipant explicitly agrees Accessibility needsMakes participation possibleResearch ops confirms Analysis approachReduces cherry-picking laterResearch lead approves Keep the brief focused. An interview guide is not a database schema for every fact the company may someday want. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI customer interview prompts Replace brackets with sanitized, verified details. These templates create drafts, not validated research. 1. Turn an assumption into a research question “Using this product assumption, proposed decision, existing evidence, target audience, constraints, and unknowns: [paste], draft three neutral research questions. For each, explain what an interview could reveal, what it cannot establish, which recent behavior matters, suitable participants, and what decision the evidence may inform. Reject questions that ask interviews to estimate market prevalence, prove causality, or predict purchases with certainty.” “Would customers pay for automation?” is mostly an invitation to be agreeable. A stronger question investigates how people currently complete the task, what it costs them, what they have already tried, and which tradeoffs drove their choices. The human researcher should select one learning goal. Ten fuzzy goals create a long interview and shallow evidence. 2. Define participant and screening criteria “Given this research question, relevant behavior, recency window, product access, geography, accessibility needs, exclusions, recruitment channels, and known gaps: [paste], draft participant criteria and a short screener. Separate essential criteria from convenient preferences. Flag professional-participant risk, power dynamics, channel bias, missing groups, and sensitive questions. Do not claim the resulting sample is representative.” Interviewing active power users because they answer Slack quickly tells you about active power users. It may tell you nothing about people who abandoned setup, churned, could not access the product, or never trusted the invitation. Screen only for what the study needs. Collecting extra demographics “just in case” creates privacy risk without automatically creating insight. 3. Draft a neutral interview guide “Using this approved research question, participant criteria, interview length, known evidence, prohibited topics, and decision context: [paste], draft a conversational interview guide. Include an introduction, consent check, easy context questions, recent-behavior questions, neutral probes, closing reflection, and time boxes. Use one idea per question. Avoid pitching, praise-seeking, jargon, hypotheticals presented as evidence, and language that reveals the preferred answer.” A guide is a map, not courtroom testimony. The interviewer should follow relevant threads and skip questions already answered naturally. Ask “Tell me about the last time you…” before “Would you ever…?” Recent behavior is imperfect, but usually more useful than imagined behavior in an imagined future with an imaginary budget. 4. Replace leading and loaded questions “Audit this draft interview guide: [paste]. Create a table with exact question, issue, likely effect, neutral rewrite, and a suggested behavior-based follow-up. Check for leading wording, false assumptions, double-barreled questions, praise-seeking, solution pitching, social-desirability pressure, jargon, absolutes, blame, and questions that make disagreement awkward. Preserve the original wording for an audit trail.” “How helpful would our smart dashboard be?” has already declared the dashboard smart and helpful. “Walk me through how you currently monitor this work” leaves room for reality. AI can catch textbook leading language. It may miss internal politics, cultural meaning, power differences, or the fact that the interviewer designed the feature being discussed. Human review remains mandatory. 5. Generate behavior-based follow-ups “For each core question in this guide: [paste], suggest three optional neutral probes that ask for a recent example, sequence of events, workaround, trigger, consequence, artifact, or tradeoff. Include reminders to tolerate silence and ask for clarification using the participant’s own words. Do not suggest ‘why didn’t you’ blame questions, debate the participant, diagnose motivation, or push toward the product concept.” Useful probes include “What happened next?”, “Can you show me how you handled that?”, and “When you said ‘confusing,’ what did that mean in that moment?” Do not machine-gun every follow-up. Listening beats completing the template. A participant’s unexpected detour may contain the actual answer. 6. Prepare consent, recording, and safety checks “Using this research setting, participant relationship, location, topics, incentive, recording method, storage system, retention period, observers, and intended uses: [paste], draft a pre-interview checklist and plain-language consent script for human legal/privacy review. Cover voluntary participation, recording choice, withdrawal, confidentiality limits, observers, quote use, storage, access, retention, incentive, accessibility, and escalation if distress or sensitive information appears. Mark legal requirements unverified.” Consent is not a decorative sentence buried before the first question. Participants should understand what happens to their words and be able to decline recording without being tricked. If someone shares credentials, personal health details, legal allegations, or another person’s private information, do not casually feed it into the synthesis pipeline. Stop, follow the approved protocol, and minimize exposure. 7. Create an evidence-first note-taking template “Create a note-taking template for this interview guide and research question: [paste]. Include timestamp or question reference, direct quote, observed behavior or artifact, factual context, interviewer observation, interpretation, confidence, contradiction, follow-up needed, and privacy flag. Keep participant statements separate from analyst conclusions. Add a post-interview section for immediate reflections and possible interviewer bias.” Notes become dangerous when “participant said export took 20 minutes” silently turns into “exports are broken.” Preserve what was said, what was observed, and what the researcher inferred. If recording is allowed, notes still matter. They capture context, nonverbal events cautiously, and what the interviewer thought was important at the time. Never treat automated transcription as perfect. 8. De-identify notes before AI-assisted synthesis “Using this approved de-identification policy and a synthetic example of interview notes: [paste], create a redaction checklist and transformation procedure. Cover direct identifiers, contact details, company and product names, precise job titles, locations, dates, rare attributes, linked events, credentials, health or financial details, third-party information, and distinctive quotations. Recommend participant IDs and a separate access-controlled key. Do not claim text is anonymous merely because names were removed.” A quote can identify someone through context. “The only pediatric surgeon using Product X in a small Montana hospital” does not need a name attached. The safest dataset is the smallest one that answers the research question. Keep raw recordings and identity keys outside general-purpose model workflows unless the approved system and consent explicitly support that use. 9. Compare themes without flattening disagreement “Using these de-identified, human-checked interview notes and the approved research question: [paste], create a candidate-theme matrix. For each theme, cite participant IDs and source excerpts, list supporting and contradicting evidence, relevant context, possible alternative explanations, missing groups, and follow-up questions. Distinguish repeated patterns from isolated but high-impact cases. Do not invent quotes, count vague mentions as prevalence, or resolve disagreement by majority vote.” Three people mentioning a workaround is a pattern in your interviews. It is not automatically “most customers.” One participant describing a severe accessibility blocker may deserve action even if nobody else encountered it. AI is useful for sorting. Humans must return to the source notes, check whether excerpts were stripped of context, and decide what differences matter. 10. Draft a findings brief with limitations “Using this verified theme matrix, participant criteria, recruitment method, research question, decision context, and unresolved contradictions: [paste], draft a findings brief. Include executive summary, method, participant boundaries, evidence-backed findings, counterevidence, limitations, open questions, implications, and options for next research or product action. Label observations, interpretations, and recommendations. Every claim must trace to supplied evidence. Do not fabricate consensus, prevalence, certainty, quotes, or ROI.” A useful brief helps a decision owner see what changed, why, and where uncertainty remains. It does not pretend six interviews are a referendum. Recommendations should be framed as options with tradeoffs and evidence needs. Product, research, legal, accessibility, security, and business owners still make the call. A practical interview workflow Use the prompts in a controlled sequence: Define one decision and one learning goal. Review existing evidence before recruiting anyone. Choose participants connected to recent relevant behavior. Draft and human-review the guide. Pilot the guide with one or two appropriate people. Confirm consent, accessibility, recording, and retention rules. Conduct interviews with a human interviewer who listens. Secure, minimize, and de-identify notes under policy. Use AI only for approved organization and candidate synthesis. Verify every theme against source evidence and document limits. For turning findings into clearer product boundaries, see AI requirements gathering prompts. For the eventual choice among options, use AI decision-making prompts without outsourcing accountability. Common ways teams fool themselves They interview whoever is easiest. Friendly customers, internal champions, and loud community members are convenient. Document who is absent. They pitch instead of research. If half the session explains the concept, the participant is reviewing your presentation, not describing their world. They ask about hypothetical intent. “Would you use this?” costs nothing to say yes to. Ask about recent behavior, existing spending, workarounds, and tradeoffs. They count mentions as market size. Interviews reveal mechanisms and context. Surveys, behavioral data, experiments, or other methods may be needed for magnitude. They summarize too early. A neat theme created after the second interview can become a lens that hides later contradictions. They paste raw transcripts into a consumer chatbot. Convenience does not cancel consent, contracts, policy, or privacy obligations. They let generated prose outrank evidence. A polished findings brief can still be wrong. Trace every claim back to a participant statement, observation, or verified artifact. If your organization needs clearer rules for approved tools and data classes, these AI policy prompts can help draft the policy. Humans with actual authority must review it. Frequently asked questions Can ChatGPT write customer interview questions? Yes, it can draft and critique questions when given a specific learning goal, audience, recent behavior, constraints, and existing evidence. A trained human should review the guide for bias, sensitivity, accessibility, context, and research fit before using it. Can AI conduct customer interviews by itself? A bot can collect structured responses, but that is not equivalent to a skilled human interview. It may miss discomfort, rapport, ambiguity, unexpected context, accessibility needs, and ethical boundaries. For meaningful or sensitive research, use an accountable human interviewer and approved protocol. How many customer interviews are enough? There is no universal magic number. The answer depends on the decision, audience diversity, recruitment quality, risk, method, and whether new interviews still change understanding. Report the actual participant boundaries and avoid pretending a small qualitative sample estimates prevalence. Should I record customer interviews? Only with clear consent and an approved plan for storage, access, use, retention, deletion, and withdrawal. Recording laws and organizational requirements vary. Participants should know who can access the recording and whether quotes may be used. Can I paste interview transcripts into an AI tool? Only if participant consent, contracts, organizational policy, data classification, vendor terms, retention controls, and the approved research system allow it. Prefer de-identified notes and data minimization. Do not assume removing names makes a transcript anonymous. How do I stop AI from inventing customer quotes? Explicitly forbid invented quotations, require participant IDs and source excerpts for every claim, and verify the output against source notes. If the tool cannot provide a traceable source, treat the statement as unsupported and remove it. What is the difference between interviews and surveys? Interviews explore context, behavior, language, and mechanisms in depth. Surveys collect structured self-reports at broader scale when sampling and questionnaire design are defensible. Use the method that matches the question; sometimes the right sequence is interviews first, then a survey designed from what you learned. Who owns the final product decision? A human decision owner. AI may organize evidence and outline options, but strategy, ethics, feasibility, customer impact, and accountability remain human work. If that boundary feels fuzzy, read the five-minute guide to what AI can and cannot do. The point is better listening, not automated certainty AI can make interview preparation faster and synthesis less chaotic. That is useful. The value disappears when teams use generated questions to lead participants, upload sensitive transcripts without permission, or turn six conversations into a universal truth. Use the machine for structure. Use humans for trust, curiosity, judgment, privacy, and responsibility. That division of labor is the larger argument in Don’t Replace Me by Dmitry Kargaev: the goal is not to cosplay as a machine. It is to use one without surrendering the parts of work that require an accountable person. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Customer Journey Mapping Prompts: Find the Friction 10 AI customer journey mapping prompts to organize evidence, map touchpoints, find friction, plan research, and turn journey maps into accountable action. AI promptscustomer journey mappingservice design AI Survey Design Prompts: Ask Better Questions 10 AI survey design prompts for objectives, neutral questions, response scales, pilots, analysis, and reporting—without turning biased feedback into facts. AI promptssurvey designuser research AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation --- ## AI Customer Journey Mapping Prompts: Find the Friction URL: https://dontreplace.me/articles/ai-customer-journey-mapping-prompts-find-friction-not-fiction AI promptscustomer journey mappingservice designuser researchcustomer experience AI Customer Journey Mapping Prompts: Find the Friction 10 AI customer journey mapping prompts to organize evidence, map touchpoints, find friction, plan research, and turn journey maps into accountable action. August 24, 2026 Updated August 24, 2026 Customer journey maps have a dangerous superpower: they can make guesses look official. Add a curved arrow, three emoji faces, and a gradient labeled “delight,” and suddenly nobody asks whether a real customer did any of it. AI customer journey mapping prompts can help organize research, draft a map structure, label evidence gaps, compare segments, and turn workshop notes into actions. They cannot observe customers, verify touchpoints, infer emotions as fact, represent people your research excluded, or decide which problem deserves investment. AI can format the map. Humans must supply the evidence, challenge the fiction, protect customer data, and own the decisions. These ten templates are for product managers, UX researchers, service designers, marketers, customer-success teams, and operations leads. Use them to create a working model of a bounded experience—not a decorative mural of “the customer.” If you still need source material, start with AI customer interview prompts or AI survey design prompts. What a useful customer journey map contains A journey map describes how a specific group moves through a specific experience to reach a specific goal. It should connect customer actions to touchpoints, backstage processes, evidence, friction, and opportunities. It should also admit what the team does not know. A defensible map includes: A bounded journey: one starting condition, goal, and ending condition. A defined audience: the people represented and explicitly excluded. Stages grounded in behavior: not whatever fits five pretty columns. Customer goals and actions: what people are trying to accomplish and do. Touchpoints and channels: where interactions actually happen. Backstage dependencies: teams, systems, vendors, and policies customers cannot see. Evidence references: research, analytics, tickets, observations, or process records. Confidence labels: verified, supported, weak, assumed, or unknown. Friction and accessibility barriers: where effort, failure, exclusion, or risk appears. Owners and next steps: because a map without action is expensive wallpaper. A map is not universal truth. New customers and experienced customers may take different routes. A mobile user with a screen reader may experience a different journey from a desktop user. A support ticket reveals a failure, not the prevalence of that failure. A conversion funnel shows where people leave, not necessarily why. If you are mapping a digital task, pair this work with AI usability testing prompts. Journey mapping connects the experience across time and channels; usability testing examines what happens when someone tries to use a specific interface. The reusable journey mapping prompt formula Add this instruction to any template below: “Act as a customer-journey synthesis assistant. I am mapping [audience, situation, starting condition, goal, ending condition, channels, time period, decision, and constraints]. Use only the verified evidence I provide. Produce [artifact] and label every item as observed, reported, measured, documented, inferred, assumed, or unknown. Preserve conflicting evidence and segment differences. Include sources, confidence, accessibility concerns, privacy risks, backstage dependencies, open questions, and required human review. Do not invent customer actions, quotes, emotions, touchpoints, prevalence, causes, or consensus.” Those evidence labels matter. “Seven support tickets mention failed verification” is documented evidence. “Customers feel betrayed” is an interpretation unless customers said that. “Replace the verification vendor” is a proposed decision. AI tends to glide across those boundaries because a smooth answer looks complete. Do not let it. Never paste customer names, email addresses, raw recordings, identifiable transcripts, unredacted support tickets, account details, health or financial information, credentials, unreleased strategy, or legally sensitive material into an unapproved AI tool. Use approved systems, de-identification, consent, access controls, retention limits, and human privacy review. A rare job title plus a distinctive incident can identify someone even after their name is removed. What to collect before prompting “Create a customer journey map” produces generic stages such as awareness, consideration, purchase, and loyalty. That may describe a marketing slide. It does not necessarily describe your customer’s work. InputWhy it mattersHuman check Journey boundaryPrevents mapping an entire lifetimeResearch lead approves Customer goalAnchors stages in customer intentCustomers or evidence verify Audience and exclusionsShows who the map representsResearch owner documents Decision to informStops mapping theaterDecision owner confirms Research and analyticsGrounds actions and frictionSource owners verify Channels and touchpointsExposes cross-channel gapsOperations validates Backstage processConnects pain to causesProcess owners review Accessibility evidenceFinds exclusion hidden by averagesAccessibility expert reviews Known contradictionsPrevents false consensusResearch lead preserves Evidence datesReveals stale assumptionsOwners confirm currency Privacy constraintsLimits unsafe data usePrivacy or legal reviews Success measuresConnects changes to outcomesAnalytics owner approves Sanitize this material before using it. Summaries should retain useful context without retaining identities. When possible, reference source IDs that authorized reviewers can trace inside approved systems rather than copying sensitive content into the prompt. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI customer journey mapping prompts Replace brackets with sanitized, verified details. Each output is a draft for human review. 1. Define the journey scope and decision “Using this business question, customer group, triggering situation, customer goal, proposed start and end points, channels, time period, existing evidence, exclusions, and decision deadline: [paste], draft a journey scope statement. Include what is in scope, out of scope, the decision the map will inform, represented and missing groups, evidence required, and five ways the scope could become misleading. Offer two narrower alternatives if the proposed journey is too broad.” “Map onboarding” is not a scope. Does onboarding begin when someone signs a contract, receives an invitation, or first attempts setup? Does it end after account creation, first value, or successful adoption? Different boundaries produce different maps and investment choices. A human decision owner should approve the boundary. Otherwise the map expands until it contains every customer problem and explains none of them. 2. Build an evidence inventory “Review this sanitized list of interviews, observations, survey results, analytics, support themes, process documents, accessibility findings, and prior maps: [paste]. Create an evidence inventory with source, date, audience, method, sample or coverage, journey relevance, limitations, privacy restrictions, and confidence. Identify unsupported claims, stale sources, overrepresented channels, missing groups, and questions requiring new research. Do not merge different evidence types into one confidence score.” This prompt prevents the workshop’s loudest anecdote from becoming the map’s organizing principle. Analytics can show a repeated drop-off. Interviews can explain possible reasons. Support tickets can expose severe edge cases. Each source contributes something different. Use AI data analysis prompts to structure quantitative checks, but have an analyst verify definitions, filters, denominators, and date ranges. 3. Draft evidence-backed stages and customer goals “Using only this approved evidence inventory and journey scope: [paste], propose stages based on changes in customer goal, behavior, context, channel, or responsibility. For each stage, list the customer goal, entry condition, observed actions, exit condition, evidence references, confidence, and open questions. Label conventional lifecycle stages as hypotheses unless evidence supports them. Provide an alternate stage model where contradictory behavior suggests more than one route.” Stages should help the team reason about meaningful transitions. They should not exist because a template arrived with six empty boxes. Ask researchers and frontline staff to review the draft independently before a workshop. If everyone edits it together immediately, hierarchy can turn a hypothesis into “what we all know.” 4. Separate facts, interpretations, and assumptions “Audit this draft journey map: [paste]. Put every claim into one of seven categories: observed behavior, customer-reported statement, measured event, documented process, interpretation, assumption, or unknown. Preserve its source reference. Flag emotional labels without direct evidence, causal claims without a test, universal language, invented transitions, and places where one segment is treated as everyone. Rewrite unsupported claims as explicit hypotheses or research questions.” This is the prompt that removes “customer is delighted” from a box supported only by a completed transaction. Completion may indicate delight, resignation, urgency, habit, or a lack of alternatives. Do not delete uncertainty to make the map cleaner. Visible uncertainty is useful: it tells the team where research can change a decision. 5. Map touchpoints, channels, and backstage handoffs “Given these verified customer actions, channel records, service processes, system dependencies, policies, and team responsibilities: [paste], create a touchpoint and handoff matrix. For each journey stage, show customer-facing touchpoints, channel changes, backstage activities, systems, vendors, queues, owners, handoff conditions, failure modes, evidence, and unknowns. Flag ownership gaps and differences between the documented process and observed reality.” Customers experience one company. Internally, their request may bounce among sales, billing, product, an outsourced verification vendor, and support. Journey mapping earns its keep when it connects visible friction to those invisible handoffs. For a deeper process check, use AI workflow audit prompts. The humans doing the work still need to confirm what actually happens, especially where the official process and survival process differ. 6. Find friction, failure, and accessibility barriers “Using these verified observations, task failures, abandonment events, support themes, wait times, accessibility findings, and customer statements: [paste], draft a friction register. For each item, record the affected stage and segment, evidence, severity, frequency if measured, customer effort, accessibility impact, business impact, workaround, likely contributing conditions, confidence, and owner. Keep causes separate from symptoms. Do not estimate prevalence or emotion when the data does not support it.” Average completion can hide exclusion. A workflow may look acceptable overall while failing keyboard users, people using assistive technology, customers with limited bandwidth, non-native speakers, or anyone who cannot switch channels. An accessibility specialist and affected users should review accessibility findings. AI can remind you to ask; it cannot certify an experience as accessible. 7. Compare segments without inventing personas “Compare these evidence sets for [segments]: [paste]. Create a matrix of shared stages, different goals, routes, constraints, touchpoints, friction, accessibility needs, and evidence gaps. Keep each claim tied to its segment and source. Distinguish a measured difference from an anecdotal difference. Do not invent names, biographies, motivations, quotes, demographics, or personality traits. Recommend when separate journey maps are justified.” A fictional persona can make a workshop easier to narrate and harder to audit. “Busy Brenda hates forms” is not evidence. “Five of eight interviewed sole operators abandoned setup when tax documentation was requested” is evidence with visible limitations. Segment only when the distinction changes the journey or decision. Decorative demographic splits create extra maps without creating extra understanding. 8. Turn gaps into a research plan “Using this journey map’s assumptions, unknowns, contradictions, weak evidence, missing groups, and upcoming decisions: [paste], create a prioritized research backlog. For each question, include the decision it affects, risk of being wrong, suitable method, participants or data needed, accessibility and consent needs, owner, deadline, and expected evidence. Identify questions interviews cannot answer and where analytics, observation, usability testing, service data, experiments, or policy review are more appropriate.” Prioritize gaps by decision risk, not visual emptiness. An unknown that could reverse a costly roadmap choice matters more than a missing quote for a presentation. Use direct research where trust, context, or observation matters. AI may help draft materials, but humans must recruit responsibly, conduct the work, and interpret findings. 9. Design an evidence-first mapping workshop “Using this approved scope, evidence inventory, draft map, participant roles, power dynamics, remote or in-person format, accessibility needs, available time, and decisions: [paste], draft a journey-mapping workshop plan. Include pre-reading, evidence review, silent individual critique, stage validation, contradiction capture, assumption labeling, friction prioritization, breaks, accessible participation options, decision rules, parking lot, owners, and follow-up. Prevent senior opinions from silently overriding customer evidence.” A workshop should reconcile evidence and expose disagreement. It should not ask twenty people to generate sticky notes from memory and then call the cluster “research.” Send the evidence inventory before the session. Let participants critique independently before discussion. Record dissent and unresolved questions. Consensus achieved through exhaustion is still not evidence. 10. Create an accountable action plan “Convert this human-reviewed journey map, friction register, research gaps, constraints, and decision criteria into an action plan: [paste]. For each proposed action, include the customer problem, evidence, affected groups, hypothesis, owner, dependencies, effort range supplied by humans, risk, accessibility and privacy review, success and guardrail measures, validation method, deadline, and stop or rollback condition. Separate quick fixes, experiments, research, policy changes, and structural work. Do not rank actions using invented numbers.” A journey map becomes valuable when it changes research, operations, product behavior, policy, or priorities. Assign owners and decision dates. Define how the team will know whether a change helped and whether it harmed another part of the journey. If several problems compete for attention, use AI prioritization prompts to structure the discussion. Humans still supply the criteria and accept the tradeoffs. How to review an AI-assisted journey map Before sharing the map, run a human review: Trace claims to sources. Can a reviewer locate the evidence behind each important item? Check the boundary. Does the map stay inside the agreed journey and audience? Preserve routes and contradictions. Did synthesis flatten meaningful differences? Challenge emotional labels. Are feelings reported, observed carefully, or merely guessed? Inspect missing groups. Who could not participate, was not recruited, or disappeared in aggregation? Review accessibility. Were disabled users and channel constraints considered directly? Validate backstage reality. Do frontline teams recognize the process shown? Separate symptoms and causes. Is a proposed cause supported or labeled as a hypothesis? Protect customer data. Are outputs sanitized, access-controlled, and retained appropriately? Assign action. Does every priority have an owner, measure, and next decision? Also check whether the map is current. Product releases, policy changes, vendor changes, seasonality, and new channels can make an old map confidently wrong. Common journey mapping mistakes AI makes faster It fills every empty cell Blank space feels incomplete to a language model. Sometimes blank space means “we have no evidence.” Keep it blank and label the research need. It turns sequence into causation A customer contacted support after verification failed. That does not prove the failure caused every later action. Treat causal explanations as hypotheses until evidence supports them. It invents a smooth, linear route Real journeys loop, pause, restart, switch devices, involve colleagues, and escape through unofficial workarounds. Show alternate paths where evidence supports them. It averages away exclusion A majority route can hide severe barriers for smaller groups. Review accessibility, language, geography, account type, channel access, and power differences explicitly. It confuses polished language with confidence A complete sentence is not a verified claim. Require source and confidence fields in the artifact itself, not in a forgotten appendix. For a broader reality check, read what AI can and cannot do. It is fast pattern production, not customer understanding in a tasteful blazer. Frequently asked questions Can ChatGPT create a customer journey map? It can draft the structure of a map from evidence you provide, organize stages and touchpoints, label assumptions, and format workshop outputs. It cannot conduct the missing research, verify customer behavior, or know whether the map represents the people affected. Treat the output as a draft requiring source checks and human review. What should I include in a customer journey mapping prompt? Include the customer group, triggering situation, goal, start and end conditions, channels, decision, evidence inventory, exclusions, constraints, accessibility considerations, and required output. Tell the model to cite supplied sources, preserve contradictions, and label facts, interpretations, assumptions, and unknowns separately. How is a journey map different from a process map? A journey map centers the customer’s goals, actions, touchpoints, friction, and experience across time. A process map usually centers internal steps, systems, roles, and handoffs. Connecting the two is powerful because customer friction often originates in backstage processes. Can AI infer customer emotions from journey data? Not reliably. A drop-off, delay, support contact, or completed purchase does not prove a particular emotion. Use direct customer statements carefully, consider context, and label inferred emotions as hypotheses. Do not perform fake sentiment archaeology on people who never consented to it. How many stages should a customer journey map have? Use as many stages as the evidence and decision require. A useful stage marks a meaningful change in goal, behavior, context, channel, or responsibility. Do not force the journey into a standard number of columns because a template looked tidy. Should every customer segment have a separate map? No. Create separate maps when segment differences materially change goals, routes, barriers, touchpoints, or decisions. Keep evidence tied to each segment and avoid fictional personas that add personality without adding proof. How often should we update a journey map? Review it when products, channels, policies, vendors, customer groups, or major processes change—and before using it for a consequential decision. Record source dates and map version so stale evidence is visible. A map is a maintained decision tool, not a ceremonial artifact. Is it safe to paste customer research into an AI tool? Only when the tool and use are approved for that data. Remove direct and indirect identifiers, use consent and access controls, respect retention rules, and avoid sensitive or confidential material. For high-risk research, use synthetic summaries or work inside approved research systems with privacy review. The rule that keeps the map honest Make the model show its work: source, evidence type, confidence, segment, date, and reviewer. If it cannot tie a claim to evidence, the claim becomes a hypothesis or an unknown—not a pastel box pretending to be customer truth. AI is useful here because journey mapping contains tedious synthesis, comparison, formatting, and documentation. Human value remains in earning trust, observing reality, noticing who is missing, interpreting contradictions, making tradeoffs, and accepting accountability for what changes next. That division of labor is the point of Don’t Replace Me by Dmitry Kargaev: use the machine for speed, then keep judgment, taste, responsibility, and the uncomfortable questions stubbornly human. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Customer Interview Prompts: Learn Without Leading 10 AI customer interview prompts for research goals, neutral questions, follow-ups, notes, synthesis, and findings—without manufacturing customer truth. AI promptscustomer interviewscustomer discovery AI Survey Design Prompts: Ask Better Questions 10 AI survey design prompts for objectives, neutral questions, response scales, pilots, analysis, and reporting—without turning biased feedback into facts. AI promptssurvey designuser research AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation --- ## AI Customer Persona Prompts: Build Useful Profiles URL: https://dontreplace.me/articles/ai-customer-persona-prompts-build-useful-profiles AI promptscustomer personasbuyer personascustomer researchUX research AI Customer Persona Prompts: Build Useful Profiles 10 AI customer persona prompts to turn verified research into useful profiles, expose assumptions, reduce stereotypes, and plan human validation. August 26, 2026 Updated August 26, 2026 Customer personas have a fiction problem. Give a profile a name, a stock photo, and a favorite podcast, and a room full of adults may start treating “Operations Olivia” like a customer who personally approved the roadmap. AI customer persona prompts can organize research, compare needs, expose missing evidence, draft profile structures, and plan validation. They cannot meet customers, infer motivations as fact, repair a biased sample, prove market demand, or decide which customer differences matter. AI can draft the profile. Humans must prove that the profile represents evidence, changes a real decision, and does not flatten people into stereotypes. These ten templates help product managers, marketers, UX researchers, founders, customer-success teams, and operations leads create evidence-backed personas rather than demographic fan fiction. If you need source material first, start with AI customer interview prompts or AI survey design prompts. What a useful customer persona actually is A persona is a compact, revisable model of a meaningful pattern in customer evidence. Its purpose is to improve a specific decision: product design, onboarding, service delivery, research recruitment, messaging, or something equally concrete. A useful persona usually has five properties: Evidence-backed: important claims trace to observations, research, or verified data. Decision-specific: the profile exists to improve a named choice. Behavioral: it emphasizes needs, contexts, constraints, and outcomes. Honest about uncertainty: assumptions and contradictions remain visible. Revisable: new evidence can change or retire it. A persona is not a screenplay. “Administrators setting up their first integration without engineering support” describes a context that may change onboarding. “Tech-Timid Tina, 42, drinks oat lattes and hates change” is a bundle of guesses wearing a lanyard. Personas and segments are related but different. Segmentation defines groups using explicit rules; personas communicate meaningful patterns within or across those groups. Use AI customer segmentation prompts when you need defensible grouping logic. A persona should never quietly replace that work with vibes. People also move between contexts. A confident buyer can become a confused first-time administrator. A small team can become a regulated organization. A persona describes a useful situation, not a permanent species of human. The reusable customer persona prompt formula Add this instruction to any template below: “Act as a customer-research synthesis assistant. I am making [specific decision] for [defined population, context, and time period]. Use only the sanitized evidence I provide. Produce [artifact]. Separate observations, direct statements, interpretations, assumptions, and unknowns. For every material claim, include the source, date range, confidence, contradictory evidence, and validation method. Do not invent names, quotes, demographics, motivations, prevalence, causality, or market demand. Flag sensitive data, stereotypes, weak samples, and details that do not change the decision.” That separation matters. “Six of eight interviewed administrators asked for setup examples” is evidence. “Administrators lack confidence” is an interpretation. “They want a friendly mascot” is an idea. Models glide between those categories because plausible continuation is their entire little party trick. Never paste names, email addresses, account IDs, raw transcripts, exact locations, health or financial details, protected characteristics, credentials, confidential strategy, or legally sensitive data into an unapproved AI tool. Use approved systems, minimum necessary data, aggregation, de-identification, access controls, retention limits, consent, and human privacy and fairness review. Rare combinations can re-identify people even after names are removed. What to collect before prompting “Create our ideal customer persona” asks the model to invent a person from internet-shaped averages. Give it a decision and an evidence pack instead. InputWhy it mattersHuman check Decision to supportDefines what belongs in the personaDecision owner confirms Population and exclusionsShows who is and is not representedResearch lead approves Research questionsPrevents unrelated biographyTeam agrees Interview and survey themesGrounds reported needsResearcher verifies Behavioral evidenceConnects claims to actionsAnalyst checks Context and constraintsExplains when needs appearDomain expert reviews Source datesExposes stale findingsOwner verifies Sample limitationsPrevents false generalizationResearcher documents Contradictory casesKeeps diversity visibleTeam reviews Sensitive fieldsPrevents harmful inferencePrivacy reviewer approves Existing decisionsTests actionabilityProduct or service owner confirms Validation planStops drafts becoming doctrineResearch lead owns Keep source references in approved systems so authorized people can inspect them. Summaries should preserve disagreement instead of laundering it into a neat average. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI customer persona prompts Replace brackets with sanitized, verified details. Every output is a draft for human review. 1. Define the decision the persona must support “Using this proposed decision, population, product or service context, date range, stakeholders, and constraints: [paste], draft a persona scope. State the exact decision, what evidence could improve it, who is represented or missing, prohibited uses, success measures, and review owner. Flag goals that combine unrelated decisions and suggest narrower alternatives.” “Help marketing” is not a scope. A persona for onboarding design may emphasize setup context; one for service planning may emphasize support needs. One universal persona usually becomes generic enough to justify whatever somebody already wanted. A human owner should approve the decision before synthesis begins. Otherwise the team can cherry-pick details around a preferred solution. 2. Audit the available evidence “Review this sanitized evidence inventory, methods, sample details, dates, recruitment criteria, data definitions, consent limits, and known quality issues: [paste]. Create a persona-readiness audit. For each source, record what it can support, coverage, freshness, reliability, missing groups, possible bias, sensitive content, and claims it cannot support. Flag duplicate participants, leading questions, survivorship bias, inaccessible methods, and conflicting definitions.” This catches the dull failures behind exciting profiles. Sales notes overrepresent prospects who reached sales. Support tickets overrepresent people willing and able to complain. Analytics miss offline work and blocked tracking. Research participants selected into the study. For quantitative evidence, use AI data analysis prompts and have an analyst verify queries, joins, filters, denominators, and missing-value treatment. 3. Separate observations from assumptions “Classify these proposed persona claims: [paste]. Put each into observed behavior, direct customer statement, contextual fact, calculated measure, researcher interpretation, business assumption, sensitive inference, or unsupported detail. Cite the available source and confidence. Rewrite unsupported motivations as research questions and remove details that do not affect the decision.” A participant abandoned setup after an error. That is behavior. “They fear technology” is a story. Maybe the error message was useless, they lacked permission, or they had a meeting in two minutes. Fiction often enters through innocent fields: age, job title, personality, favorite brands, or a fabricated quote. If a detail lacks evidence or does not change action, delete it. 4. Cluster needs without stereotyping people “Using only these verified behaviors, reported needs, contexts, constraints, and outcomes: [paste], propose three to five candidate need patterns relevant to [decision]. For each, list supporting evidence, sample coverage, contradictory cases, overlap, likely context changes, unknowns, and the action it might change. Do not assign names, demographic identities, personality types, or prevalence.” Candidate patterns are hypotheses, not discovered tribes. Two people can share a need for faster setup for completely different reasons. One person can fit multiple patterns at different times. If the patterns do not imply different legitimate actions, you may not need separate personas. Decorative variety is still decoration. 5. Extract jobs, constraints, and desired outcomes “Analyze this sanitized research evidence: [paste]. Build a table of customer situations, tasks or jobs, triggers, desired outcomes, barriers, workarounds, dependencies, and directly stated success criteria. Separate frequency from importance. Include source references, confidence, disagreement, and missing evidence. Do not convert behavior into personality or infer an unstated emotional motive.” This produces more useful material than “busy professional who values convenience.” Almost everybody values convenience. “Must reconcile three systems before a Friday compliance deadline” gives a team something to design around. Use AI customer journey mapping prompts when sequence and handoffs matter. Keep alternative journeys visible rather than forcing one heroic funnel. 6. Map evidence to persona fields “Using this approved persona template and evidence table: [paste], map every proposed field to its supporting sources. Include context, goal, behavior, constraint, workaround, desired outcome, decision implication, confidence, date range, contradictions, and unknowns. Mark any field with no evidence as REMOVE or RESEARCH. Do not fill blanks creatively.” The instruction not to fill blanks is important. Language models experience an empty field as an invitation. Your research process should experience it as honesty. Keep citations or source IDs near important claims. A persona deck without traceability becomes organizational folklore after the people who ran the research leave. 7. Find contradictions and missing research “Stress-test this candidate persona against the supplied evidence: [paste]. Identify disconfirming cases, differences by context, inaccessible or absent populations, sample bias, outdated findings, proxy assumptions, disputed interpretations, and claims supported by only one source. Rank research gaps by how likely they are to change the decision. Propose ethical, accessible validation methods and clear failure criteria.” Contradiction is useful information. If half the evidence points elsewhere, do not average it into a smooth paragraph. You may have multiple contexts, a weak pattern, changing behavior, or a measurement problem. Desk research can supplement the picture, but it should not impersonate your customers. AI research prompts help preserve sources and uncertainty. 8. Stress-test actionability and harm “Evaluate this candidate persona for [decision]: [paste]. Score evidence strength, distinctness, actionability, accessibility, freshness, maintainability, privacy risk, stereotyping risk, exclusion risk, and potential misuse. Explain every score and uncertainty. Identify details that encourage discriminatory targeting, worse service, manipulative messaging, or false precision. Recommend removal, revision, validation, or qualified review.” A profile can be memorable and harmful. It can also be accurate but irrelevant. Ask what changes because the persona exists, who benefits, who might be excluded, and whether the action is proportionate. For consequential use, run AI risk assessment prompts and require accountable human approval. 9. Draft an evidence-backed persona card “Turn this validated evidence into a concise persona card for [decision]: [paste]. Include a descriptive context-based label, situation, jobs, verified behaviors, reported needs, constraints, desired outcomes, current workarounds, decision implications, supporting sources, confidence, contradictions, exclusions, privacy limits, last-reviewed date, and a prominent ‘what we do not know’ section. Do not add a fictional name, photo, quote, age, biography, or personality.” A label such as “First-time admins without engineering support” may be less adorable than “DIY Dana.” Good. It tells readers when the profile applies and resists becoming a cartoon. Review the card with researchers, analysts, frontline staff, accessibility and privacy stakeholders, and—when appropriate—people represented by the evidence. 10. Create a validation and maintenance plan “Create a human-owned validation and maintenance plan for this persona and its claims: [paste]. Include methods, participant coverage, accessibility and language needs, quantitative checks, contradictory cases, decision tests, privacy and fairness review, owners, dates, evidence thresholds, version history, change triggers, and retirement criteria. Prioritize claims most likely to alter the proposed action.” Validation is not asking whether customers “identify with” a catchy card. Test whether the pattern reproduces, whether it improves the target decision, and whether the resulting action helps without unacceptable harm. Set expiration dates. Products, markets, policies, and behavior change. A persona from three years ago should not haunt the roadmap because nobody knows who owns the slide deck. A simple human review workflow Name one decision. Do not model humanity. Audit evidence and permissions. Exclude unsafe or undefined material. Separate facts from interpretations. Label assumptions visibly. Draft need patterns. Keep overlap and context changes. Seek disconfirming evidence. Do not reward a tidy story. Review privacy and fairness. Inspect fields, proxies, uses, and effects. Test actionability. Confirm that the profile changes a legitimate choice. Validate with customers and data. Use accessible, appropriate methods. Approve and version. Record owners, sources, dates, and permitted uses. Update or retire. Do not preserve a mascot after the evidence dies. If a persona supports prioritization, AI prioritization prompts can structure tradeoffs after humans validate the evidence and criteria. Common persona mistakes AI makes faster It invents a complete person The model turns three findings into a name, salary, hobbies, fears, and quote because complete profiles resemble its training examples. Completeness is not accuracy. It converts behavior into personality Repeated help-center visits become “low technical confidence.” A delayed purchase becomes “risk averse.” Keep observations separate from possible explanations. It hides diversity inside an average A polished paragraph can erase conflicting needs, accessibility differences, and changing contexts. Preserve ranges and contradictions. It sneaks in sensitive proxies Location, device, language, browsing history, and job information can correlate with protected traits. Review what fields imply and how outputs will be used. It makes the persona impossible to disprove “Values quality but watches budget” describes nearly everyone. Require explicit evidence and failure criteria. It optimizes for memorability over usefulness Rhyming names and stock photos win workshops. Decision implications, source links, and unknowns win actual work. Frequently asked questions Can ChatGPT create a customer persona from interview notes? It can help synthesize sanitized notes, classify evidence, surface themes, and draft a traceable card. A researcher must verify the transcript handling, coding, interpretation, contradictions, and sample limitations. Do not paste identifiable raw transcripts into an unapproved tool. What should an AI customer persona include? Include the decision served, context, verified behaviors, reported needs, constraints, desired outcomes, evidence references, confidence, contradictions, unknowns, permitted uses, and review date. Include demographic details only when they are evidenced, necessary, lawful, and responsibly governed. Should customer personas have names and photos? Usually not. Fictional names and stock photos encourage teams to mistake a communication device for a real person and can reinforce stereotypes. Context-based labels are clearer and easier to revise. How many customer personas do we need? As few as needed to support distinct decisions. Three evidence-backed patterns beat twelve mascots. If profiles lead to the same action or teams cannot distinguish them reliably, merge or remove them. What is the difference between a segment and a persona? A segment groups customers using explicit variables or rules. A persona communicates a meaningful pattern of needs, behaviors, contexts, and constraints. Personas may sit within segments or cross them, but they should not substitute fictional narrative for grouping evidence. How do I validate an AI-generated persona? Trace every claim to sources, look for disconfirming cases, recruit missing populations, verify quantitative patterns where appropriate, review privacy and fairness, and test whether using the persona improves the named decision. Define what evidence would cause revision or retirement. Can AI remove bias from buyer personas? No. AI can list risks and expose unsupported claims, but it can also reproduce stereotypes and historical bias. Qualified humans must inspect data, methods, interpretations, uses, and effects while including affected perspectives. How often should personas be updated? Review them after major product, market, policy, pricing, channel, or customer changes and on a scheduled date. Monitor whether evidence, contexts, and decision implications still hold. Retire stale profiles instead of endlessly polishing them. Build a decision tool, not an imaginary friend AI customer persona prompts are useful when they make evidence, uncertainty, contradictions, safety, and validation easier to see. They are dangerous when fluent prose upgrades assumptions into customer truth. Start with one decision. Use the minimum approved evidence. Separate observations from interpretations. Keep missing people and contradictory cases visible. Test whether the profile changes a legitimate action, then keep a human accountable for the consequence. That is the broader skill behind using AI at work without the nonsense: let the machine provide speed and structure while humans retain context, judgment, responsibility, and taste. For the short version, read what AI can and cannot do. If you want the blunt field guide for that division of labor, Don’t Replace Me by Dmitry Kargaev takes it further. The model can write a convincing profile. You still have to decide whether a customer is in it—or whether the machine merely wrote better fan fiction. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Customer Segmentation Prompts: Find Useful Groups 10 AI customer segmentation prompts to audit data, find behavioral groups, test bias, validate segments, and turn evidence into responsible action. AI promptscustomer segmentationmarket research AI Usability Testing Prompts: Find Friction Fast 10 AI usability testing prompts for research plans, neutral tasks, moderator guides, observation notes, evidence synthesis, and human-reviewed priorities. AI promptsusability testingUX research AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation --- ## AI Customer Renewal Prompts: Prepare With Evidence URL: https://dontreplace.me/articles/ai-customer-renewal-prompts-prepare-without-faking-the-relationship AI promptscustomer renewalcustomer successaccount managementrevenue operations AI Customer Renewal Prompts: Prepare With Evidence 10 AI customer renewal prompts to verify outcomes, expose account risks, prepare options, and protect trust without inventing value or customer sentiment. September 08, 2026 Updated September 08, 2026 Customer renewals get strange when a calendar date becomes a substitute for customer value. The CRM says “healthy,” product usage says “maybe,” support has three unresolved issues, and somebody asks an AI tool to write an upbeat renewal email before anyone has called the customer. AI customer renewal prompts can organize verified evidence, expose missing information, prepare questions, compare approved options, and draft accurate follow-ups. They cannot know whether a customer is happy, invent return on investment, predict churn, authorize a discount, interpret a contract, or repair trust on your behalf. AI can prepare the renewal conversation. It cannot manufacture the reason a customer should renew. These ten templates help customer success managers, account managers, founders, revenue operations teams, and commercial leaders prepare renewals without faking the relationship or outsourcing judgment to autocomplete. What good renewal preparation actually does A renewal is not an invoice with better typography. It is a checkpoint where both sides decide whether the relationship still solves a real problem on workable terms. Useful preparation answers: What outcomes did the customer originally seek? Which outcomes are supported by current evidence? What changed in usage, priorities, stakeholders, or risk? Which support or delivery issues remain unresolved? Who can decide, approve, influence, and sign? Which commercial options are actually authorized? What questions must be answered before an offer is made? What commitments need owners and dates? Keep these categories separate: Verified fact: supported by a current, dated source. Customer statement: something the customer said, not proof of motive. Internal interpretation: a hypothesis that needs testing. Measured outcome: a result with an agreed definition and source. Approved option: a commercial path authorized for discussion. Open risk: a known issue with an owner or unresolved status. Unknown: information that requires a question or specialist review. Decision: a choice made by a named human with authority. Models blur those categories because confident prose sounds complete. Your process should make uncertainty visible instead. For broader relationship context, start with AI account planning prompts. If retention problems need pattern analysis across many accounts, use AI churn analysis prompts. A renewal brief should use those records, not invent a shinier parallel reality. The evidence-first customer renewal prompt formula Add this instruction to every template below: “Act as a customer renewal preparation assistant. Use only the sanitized, approved sources I provide for [account, renewal period, product, customer goals, outcomes, risks, stakeholders, and commercial options]. Cite a source label and date for every usage claim, outcome, issue, deadline, price, term, commitment, and stakeholder statement. Separate verified facts, customer statements, internal interpretations, measured outcomes, approved options, open risks, decisions, and unknowns. Mark unsupported or stale claims [VERIFY]. Do not invent customer sentiment, ROI, authority, urgency, product capabilities, approvals, legal interpretations, competitor claims, or future outcomes.” That instruction makes the output less magical and more useful. Magic is expensive when it reaches a customer deck. Never paste customer names, personal contact details, contracts, credentials, raw usage exports, private CRM notes, support transcripts, payment information, confidential pricing or margins, security findings, regulated data, health data, or legally sensitive information into an unapproved AI tool. Use approved systems, minimum necessary data, de-identification, access controls, retention limits, and human legal, finance, security, privacy, customer success, and commercial review. What to collect before prompting Build a focused source pack. Do not upload the company’s entire customer nervous system because the chatbot looked lonely. InputWhy it mattersHuman owner Current contract and renewal datesEstablishes actual terms and timingLegal or commercial operations Original goals and success planDefines what value was supposed to meanCustomer success owner Verified usage summaryShows activity without pretending activity equals valueProduct or analytics owner Outcome evidenceGrounds claims about business impactCustomer and account owner Support and incident historyExposes unresolved friction and trust risksSupport owner Delivery commitmentsChecks whether promises were completedDelivery owner Stakeholder mapIdentifies users, champions, approvers, and unknownsAccount owner Recent customer statementsPreserves the customer’s actual words and contextAccount owner Approved pricing and optionsPrevents unauthorized commercial promisesFinance or revenue operations Product capability sourcesPrevents roadmap fan fictionProduct owner Security and legal statusReveals specialist dependenciesSecurity and legal owners Restricted-data rulesControls what may enter the AI toolPrivacy or security owner When sources conflict, preserve the conflict. A renewal plan that quietly selects the happiest version is not a plan. It is wishful thinking wearing a blazer. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI customer renewal prompts Replace brackets with sanitized, approved information. Treat every output as a draft until an authorized human verifies it. 1. Audit the renewal inputs “Review this sanitized renewal source pack against these readiness requirements: [paste]. Create a table with required input, supplied evidence, source, date, status, conflict, missing information, owner, and next action. Do not infer missing values. Mark stale or unsupported claims [VERIFY].” Run this before writing a renewal narrative. It catches expired success plans, missing usage periods, contradictory dates, vague value claims, and stakeholder gaps while there is still time to investigate. The account owner should resolve relationship questions. Finance, legal, product, support, security, and delivery owners should confirm their own facts. The model is checking the folder, not certifying the account. 2. Separate facts from assumptions “Classify every material statement in these notes as verified fact, customer statement, internal interpretation, measured outcome, approved option, open risk, decision, or unknown. Include source and date. For each interpretation, draft a neutral question that could test it. Do not infer emotion, intent, satisfaction, budget, authority, or urgency.” “The customer loves us” may mean a champion praised one feature six months ago. “They are a churn risk” may mean usage dropped, the champion left, an invoice is late, or one executive had a bad call. Those situations require different responses. This prompt turns hidden stories into visible hypotheses. Humans can then ask respectful questions instead of steering the renewal around imaginary sentiment. 3. Summarize verified customer outcomes “Using only these approved sources: [paste], create an outcome summary with original goal, agreed measure, baseline, current result, measurement period, source, customer confirmation status, caveat, and owner. Put product activity in a separate section from business outcomes. Mark any causal or ROI claim not explicitly supported [VERIFY].” Logins, clicks, exports, and seats activated can show activity. They do not automatically show saved money, faster work, reduced risk, or happier employees. Never turn “used feature” into “achieved value” without evidence. If the customer has not confirmed an outcome, say so. A useful renewal review can include “we have evidence of adoption, but we still need to validate business impact.” Honesty builds more trust than a synthetic victory lap. 4. Map stakeholders and unknowns “Build a renewal stakeholder map from these dated sources: [paste]. Include role, stated goals, known concerns, participation, decision authority evidence, relationship owner, last verified date, and unanswered question. Keep inferred influence or sentiment in a separate hypothesis column. Do not profile people from demographic or unrelated personal data.” Renewals often stall because the daily user, champion, budget owner, procurement contact, security reviewer, and signer are different people. A tidy job title does not prove authority. Use the output to identify conversations that need to happen. Do not use it to generate fake familiarity or manipulative personalization. The correct way to learn what a stakeholder cares about remains annoyingly old-fashioned: ask them and listen. 5. Identify adoption, support, and trust risks “Review these sanitized usage summaries, support records, delivery notes, and customer statements: [paste]. Create a risk register with observed signal, source, date, affected goal, severity rationale, current owner, mitigation status, customer visibility, unresolved question, and next checkpoint. Do not assign churn probability or root cause without an approved method and sufficient evidence.” A dormant feature may be irrelevant, poorly introduced, blocked by permissions, replaced by another workflow, or simply seasonal. An open ticket may be minor—or the visible edge of a trust problem. The model cannot determine which from a spreadsheet alone. For material risks, use AI risk assessment prompts and involve the people accountable for support, delivery, product, security, and the customer relationship. Do not bury a known failure beneath a discount proposal. 6. Prepare renewal discovery questions “Using this verified renewal brief and unknowns list: [paste], draft open, neutral discovery questions grouped by outcomes, adoption, unresolved issues, changing priorities, stakeholders, security or procurement, commercial constraints, and decision process. For each question, state which unknown it tests. Avoid leading wording, pressure tactics, and assumptions about satisfaction or budget.” Good renewal discovery is not a trap designed to get the customer to repeat your value proposition. It is a reality check. Ask what changed. Ask which outcomes still matter. Ask where the product creates friction. Ask whether the current scope fits. Ask who needs to participate and how decisions are made. Then update the source pack with the customer’s actual answers rather than the answer the team hoped to hear. 7. Build an approved renewal options matrix “Using only these approved products, prices, terms, service levels, and exception boundaries: [paste], create a renewal options matrix. For each option, show customer goal addressed, included scope, verified dependency, price and term source, trade-off, risk, required approval, and prohibited promise. Label brainstormed ideas NOT APPROVED. Do not recommend options unsupported by customer evidence.” Options can clarify trade-offs: maintain scope, right-size seats, adjust term, sequence implementation, add approved support, or expand only where a verified need exists. More options are not automatically better. Three confusing bundles can be a PowerPoint-shaped avoidance of one honest conversation. Complex exceptions belong in AI deal desk prompts. If negotiation begins, use AI sales negotiation prompts. Finance and commercial owners authorize the offer; the model arranges approved facts into rows. 8. Rehearse difficult renewal objections “Using only these verified account facts and approved boundaries: [paste], role-play these customer objections one at a time: [list]. After each response, critique it for unsupported claims, defensiveness, evasiveness, pressure, excessive length, unapproved concessions, and failure to acknowledge a real issue. Suggest a clearer response and follow-up question. Do not invent proof or promise a resolution.” Rehearsal is useful for price pressure, weak adoption, unresolved incidents, leadership changes, missing functionality, procurement delays, or a customer considering alternatives. It lets the human find bad wording before the customer has to hear it. Do not train a team to “overcome” a legitimate complaint. If your company failed, acknowledge the failure, explain verified remediation, and state what remains uncertain. AI client communication prompts can help structure the draft, but an accountable human must own the message. 9. Draft an accurate renewal recap “Draft a concise renewal conversation recap from these approved notes: [paste]. Include confirmed goals, verified outcomes, customer concerns in neutral language, decisions, unresolved questions, commitments with named owners and dates, commercial items requiring approval, and the next checkpoint. Add a [VERIFY] marker beside anything not supported by the notes. Do not add enthusiasm, agreement, deadlines, or commitments that were not stated.” Send recaps quickly, but never let speed erase disagreement. If the customer said the usage metric does not reflect value, preserve that. If pricing is under review, do not write “pricing agreed.” If legal owns the contract answer, do not improvise one. Record material decisions with AI decision log prompts. The final recap should be reviewed by the meeting owner before it becomes part of the account record. 10. Run the final human sign-off checklist “Check this renewal brief, customer-facing draft, and proposed options against the following approval rules: [paste]. Return pass, fail, or needs review for evidence accuracy, date freshness, customer quote accuracy, outcome claims, privacy, security, product capability, pricing, terms, legal language, accessibility, unresolved risks, approval owners, commitments, and next steps. Cite the exact passage for every issue. Do not approve the renewal.” This is the boring prompt that prevents exciting mistakes. Run it before a customer-facing deck, proposal, recap, or contract request leaves the building. Use AI QA checklist prompts for a broader release-style review. A human account owner still signs off, and specialist owners approve their domains. “The chatbot found no issues” is not an approval record. A compact workflow that does not create renewal theater Use the prompts in a controlled sequence: Define the decision. State what the renewal process needs to decide and by when. Collect minimum necessary sources. Use current, approved, sanitized evidence. Audit inputs. Find gaps before drafting a story. Separate evidence from interpretation. Keep unknowns visible. Validate with owners. Customer, product, support, finance, legal, security, and delivery owners confirm their facts. Ask the customer. Use discovery to test assumptions and update the record. Prepare approved options. Tie each option to evidence and authority. Rehearse difficult conversations. Critique wording without inventing proof. Document accurately. Capture decisions, owners, dates, and unresolved questions. Require human sign-off. No autonomous offer, promise, discount, or customer communication. This is less glamorous than an “AI renewal copilot” that claims to predict every account. It is also much less likely to send a confident lie to a customer. If your team is new to this, the no-BS guide to using AI at work covers the basic operating model. The short version: give the tool bounded work, inspect the output, and keep accountability attached to a person. Common failure modes Treating activity as value A customer can use a product frequently and still fail to achieve the intended outcome. Another customer can use one feature monthly and receive substantial value. Report activity as activity until the customer and measurement method support a stronger claim. Predicting churn from vibes Risk signals can prioritize investigation. They do not reveal a customer’s decision. Never present a model-generated probability as certainty, especially when the inputs are sparse, biased, stale, or poorly defined. Hiding unresolved problems behind commercial options A discount does not fix broken trust, missing functionality, bad support, or failed delivery. Resolve or honestly frame the issue before asking the customer to recommit. Letting old CRM notes impersonate current truth Stakeholders leave. Goals change. Budgets move. Policies evolve. Put dates beside claims and revalidate material assumptions rather than recycling last quarter’s optimism. Giving the tool authority it does not have AI cannot approve pricing, terms, security exceptions, product commitments, legal interpretations, or customer promises. If an output crosses one of those boundaries, label it as a draft and route it to the named owner. For a plain-language explanation of these limits, read what AI can and cannot do. Fast pattern generation is useful. It is not judgment, authority, or a relationship. Frequently asked questions Can AI predict whether a customer will renew? AI can organize known risk signals or apply an approved analytical method to suitable data. It cannot know the customer’s decision. Predictions depend on data quality, definitions, representativeness, and changing circumstances. Use scores to guide investigation, never as permission to ignore the customer or declare an outcome certain. What information should I give an AI renewal tool? Provide the minimum sanitized information needed for the task: source labels, dates, de-identified goals, approved outcome summaries, issue categories, stakeholder roles, and authorized commercial boundaries. Do not paste raw contracts, personal data, private messages, credentials, payment details, confidential margins, regulated information, or unrestricted usage and support exports into an unapproved tool. Can AI write a renewal email for me? It can draft one from verified notes and approved language. A human must check the recipient, facts, tone, commitments, pricing, dates, privacy, and relationship context before sending. Never let the model invent customer praise, agreement, urgency, or a promise from your company. How do I stop AI from inventing customer value? Require a source and date for every outcome claim. Separate product activity from business results. Mark unsupported claims [VERIFY], preserve caveats, and ask the customer to validate the outcome. If you cannot support a claim, remove it or present it honestly as an open question. Should AI recommend discounts or renewal terms? Only within current options and boundaries supplied by authorized finance or commercial owners. It may compare approved scenarios, but it cannot authorize a concession or infer what a customer will accept. Route exceptions through the real approval process and involve legal when contract language changes. Can AI analyze support tickets for renewal risk? Yes, if your organization has an approved system, a legitimate purpose, appropriate access, and a sound method. Minimize and de-identify data where possible. Verify themes against source records, account for missing context, and do not treat ticket sentiment as the customer’s complete position. Who owns the final renewal decision and communication? Named humans do. The customer decides whether to renew. Your account owner manages the relationship. Finance, legal, security, product, support, privacy, and delivery owners approve their domains. AI can prepare drafts and checks; it cannot own a promise, approval, signature, or consequence. The useful boundary The strongest use of AI in renewals is not synthetic charm. It is disciplined preparation: finding missing evidence, separating facts from assumptions, organizing risks, generating better questions, comparing approved options, and checking drafts before they reach the customer. The human work remains the work that matters most: listening, verifying value, acknowledging failures, balancing trade-offs, protecting sensitive information, making commitments, and preserving trust. That is the broader argument in Don’t Replace Me by Dmitry Kargaev: use the machine for speed and structure, then keep judgment and accountability where they belong. A polished renewal brief may arrive in seconds. A durable customer relationship still has to be earned. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Customer Retention Prompts: Keep Customers 10 AI customer retention prompts to diagnose churn signals, plan useful follow-up, protect customer data, and keep human judgment in charge. AI promptscustomer retentioncustomer success AI Customer Success Plan Prompts: Use Real Evidence 10 AI customer success plan prompts to verify goals, map adoption gaps, assign milestones, and build an account plan without inventing customer value. AI promptscustomer success plancustomer success AI QBR Prompts: Run Reviews Without the Theater 10 AI quarterly business review prompts to verify outcomes, expose risks, build a useful agenda, and plan next quarter without inventing customer value. AI promptsquarterly business reviewQBR --- ## AI Customer Retention Prompts: Keep Customers URL: https://dontreplace.me/articles/ai-customer-retention-prompts-keep-customers-without-guessing AI promptscustomer retentioncustomer successcustomer churnaccount management AI Customer Retention Prompts: Keep Customers 10 AI customer retention prompts to diagnose churn signals, plan useful follow-up, protect customer data, and keep human judgment in charge. August 27, 2026 Updated August 27, 2026 Customer retention gets weird when a health score turns red. Suddenly everyone becomes a detective, a fortune teller, and a coupon printer. AI customer retention prompts can organize verified signals, summarize account history, surface missing evidence, draft check-in questions, and format an action plan. They cannot know why a customer is unhappy, predict an individual decision, authorize a discount, or repair a relationship without a human conversation. AI can organize retention evidence. Humans must decide what it means, what promise is responsible, and whether the proposed action actually helps the customer. These ten templates help customer-success leaders, account managers, product teams, marketers, founders, and operations leads do retention work without pretending the model has psychic access to the customer. If your inputs are still vague, use customer interview prompts before drafting a save plan. What customer retention work should actually do Retention work should identify preventable friction, understand changing customer needs, and coordinate useful follow-through. It should not trap customers, manufacture urgency, hide cancellation paths, or spray discounts at anyone who looks restless. A useful retention review separates at least five things: Verified signals: usage changes, support history, renewal dates, survey responses, and direct statements. Interpretations: plausible explanations that still need validation. Unknowns: information nobody currently has. Available actions: product, support, education, commercial, or relationship responses. Human decisions: who can approve each action and own the outcome. That separation matters because the same signal can have several causes. Lower usage could mean poor adoption, seasonal work, a finished project, missing tracking, a changed team, or a customer getting exactly what they needed with fewer sessions. Calling every dip “churn risk” is how dashboards become workplace astrology. Retention also differs from segmentation. A segment can show patterns across groups; an account decision concerns a particular relationship and current context. Use AI customer segmentation prompts for group analysis, then validate individual assumptions directly. The reusable AI customer retention prompt formula Add this instruction to any template below: “Act as a customer-retention analysis assistant. Use only the sanitized, verified evidence I provide for [customer group or account context] during [date range]. Produce [artifact] for [decision]. Separate facts, customer statements, calculated signals, interpretations, assumptions, and unknowns. For each material claim, include its source, date, confidence, contradictory evidence, owner, and next validation step. Do not invent churn reasons, quotes, probabilities, causal claims, customer intentions, or commitments. Flag sensitive data, unfair proxies, weak samples, stale evidence, and actions requiring human approval.” This formula blocks the most common failure: polished certainty built from half a CRM note and a declining chart. AI is fast at compressing material. It is not a witness, a relationship owner, or a commercial authority. Never paste customer names, emails, account IDs, contracts, payment details, credentials, private support conversations, confidential roadmap information, health data, or legally sensitive material into an unapproved AI tool. Use approved systems, minimum necessary data, aggregation, de-identification, access controls, retention limits, and human privacy review. “We removed the name” is not enough when the remaining details identify one customer. What to collect before prompting A useful retention prompt needs an evidence pack, not “Why is this customer leaving?” InputWhy it mattersHuman check Decision and deadlineDefines the actual outputAccount owner confirms Product or service scopePrevents irrelevant analysisDelivery owner verifies Direct customer statementsPreserves the customer's voiceSource link checked Usage definitions and datesExposes stale or noisy signalsAnalyst verifies Support historyShows unresolved frictionSupport lead reviews Adoption milestonesDistinguishes setup from valueSuccess lead confirms Contract factsPrevents invented obligationsCommercial owner checks Known product incidentsAdds operational contextProduct owner verifies Prior promisesPrevents contradictory outreachAccount owner reviews Missing stakeholdersShows whose view is absentTeam identifies Sensitive fieldsLimits harmful processingPrivacy owner approves Action constraintsStops impossible recommendationsDecision-maker confirms Keep links to source records in approved systems. Summaries should preserve disagreement and uncertainty rather than flattening everything into one suspiciously tidy story. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI customer retention prompts Replace the brackets with sanitized facts. Treat every output as a draft that needs an accountable human owner. 1. Audit the available retention evidence “Review this sanitized evidence inventory, date range, data definitions, collection methods, access limits, and known quality problems: [paste]. Create a retention-evidence audit. For each source, list what it supports, freshness, coverage, reliability, contradictory evidence, sensitive content, and claims it cannot support. Flag missing periods, tracking changes, duplicate events, survivorship bias, selection bias, and vague CRM notes.” Run this before interpreting risk. A health score may combine usage, tickets, survey responses, and commercial status using weights nobody remembers. A clean-looking number can hide changed event tracking or a support backlog. Have an analyst verify filters, joins, denominators, and missing values. AI data analysis prompts can help structure that review, but a generated query is not automatically a correct query. 2. Separate signals from guesses “Classify these retention claims and notes: [paste]. Put each into verified event, direct customer statement, calculated signal, contextual fact, interpretation, assumption, sensitive inference, or unknown. Cite the source and date. Rewrite unsupported conclusions as questions for validation. Do not infer emotion, intent, budget, authority, or likelihood to churn.” “Weekly active users fell 25 percent” can be a calculated signal if the metric is sound. “The champion has lost interest” is an interpretation. “They will cancel” is a prediction. Those statements do not belong in the same column. This prompt is especially useful before an executive review, where guesses can harden into company folklore after one confident slide. 3. Summarize account health without fake precision “Using these verified product, service, support, relationship, and commercial signals: [paste], draft an account-health brief. Organize it into confirmed strengths, confirmed friction, recent changes, unresolved questions, conflicting evidence, upcoming decision points, and human owners. Use qualitative confidence labels with reasons. Do not create a churn probability or overall score unless I provide an approved method.” A good brief helps somebody act. It does not turn six uncertain inputs into “73.4% churn risk,” a number precise enough to look scientific and meaningless enough to survive every meeting. Keep commercial facts separate from product value. A renewal date can increase urgency, but it does not prove dissatisfaction. A large contract does not make a weak inference stronger. 4. Find recurring friction across customer evidence “Review these de-identified support themes, interview findings, survey comments, adoption notes, and product events: [paste]. Identify recurring friction relevant to retention. For each theme, show supporting evidence, affected contexts, date range, frequency only when valid, contradictory cases, likely measurement gaps, and a proposed validation step. Do not claim causation or generalize beyond the supplied population.” This can reveal patterns that no single account owner sees. It can also amplify whatever your systems collect most easily. Tickets show reported problems, not every problem. Surveys show responses from people who chose to respond. Compare findings with the broader customer journey mapping prompts before deciding that one visible complaint explains the whole relationship. 5. Prepare a useful customer check-in “Using this verified account context, previous commitments, open questions, and meeting goal: [paste], draft a 30-minute customer check-in agenda. Include a plain-language opening, questions about outcomes and friction, confirmation of changed priorities, review of prior commitments, space for unexpected concerns, and a closing recap. Avoid leading questions, defensive phrasing, upsell pressure, and claims not supported by the record.” The goal is to learn, not to corner the customer into validating your dashboard. “Are you getting value from feature X?” is weaker than “What work are you trying to complete now, and where does the current process slow you down?” A human should choose the questions and conduct the conversation. Tone, history, power, and trust do not fit neatly into a prompt window. 6. Draft a save-plan hypothesis “Given these verified concerns, desired outcomes, constraints, prior attempts, and available actions: [paste], draft three retention hypotheses. For each, state the evidence, assumption, customer benefit, proposed action, required approval, cost or tradeoff, success measure, stop condition, and validation question. Rank by customer usefulness and reversibility, not by pressure.” A save plan is a hypothesis until the customer confirms it. More training might help an adoption gap. It will not fix a missing capability. A discount might address budget pressure. It will not repair broken trust. Do not promise roadmap work, credits, pricing, security changes, legal terms, or service levels without the owner authorized to make that promise. 7. Build a human-owned risk review “Turn this sanitized account context into a retention risk register. Use columns for verified risk, source, affected outcome, likelihood basis, impact basis, uncertainty, mitigating action, owner, approval needed, review date, and escalation trigger. Keep customer risk, delivery risk, privacy risk, commercial risk, and reputational risk separate. Flag any action that could manipulate, discriminate, or create an unauthorized promise.” This makes risk discussable without asking AI to pronounce a verdict. Owners and review dates matter more than a colorful heat map. For high-impact actions, run a separate AI risk assessment prompt and involve legal, security, privacy, finance, or leadership where appropriate. 8. Compare customer groups without stereotyping “Using these approved aggregate metrics and verified research themes for defined groups: [paste], compare retention patterns. Report definitions, sample sizes, time periods, missing data, observed differences, uncertainty, and alternative explanations. Do not infer protected traits, personality, intent, or individual behavior. Flag proxy variables, small groups, selection effects, and comparisons that need statistical or fairness review.” Group differences can guide research and product work. They should not become a shortcut for treating a person as their segment average. Avoid labels like “low-value customers” when the measure is really current revenue, feature usage, or support volume. Name the metric. Moral adjectives are not analysis. 9. Create a retention experiment brief “Using this verified friction hypothesis and available interventions: [paste], draft a retention experiment brief. Include the customer problem, target population, evidence, intervention, comparison approach, primary and guardrail metrics, sample and duration questions, consent or privacy needs, operational owner, analysis plan, stop conditions, and decision rule. Flag novelty effects, contamination, selection bias, small samples, and harms.” An experiment should test a useful change, not make cancellation harder and celebrate the delayed click. Guardrail metrics might include support burden, complaints, accessibility, refunds, or unintended pressure. A qualified analyst should approve design and interpretation. AI can format the brief; it cannot rescue a biased experiment after the fact. 10. Write the follow-up action plan “Convert these verified meeting notes and approved decisions into a customer retention action plan. Separate customer commitments from our commitments. For each action include owner, due date, dependency, approval, evidence link, customer-visible status, and escalation trigger. Draft a concise follow-up message that states what was heard, what was agreed, what remains unknown, and when the next update will arrive. Do not add promises or deadlines absent from the notes.” This is where useful retention work becomes ordinary execution. The customer should not need another meeting to discover who owns the fix. Have the account owner compare the draft with the source notes and prior commitments. A polished follow-up containing one invented promise is worse than a slightly ugly accurate email. How to review AI-generated retention work Before using an output, ask: Can every material claim be traced to a source and date? Are direct customer statements separate from interpretations? Did the output invent intent, emotion, probability, or causality? Are missing stakeholders and contradictory signals visible? Does the action help the customer, or merely protect a metric? Does anyone need to approve pricing, legal, security, privacy, or roadmap language? Is sensitive data minimized and processed in an approved system? Is there a named human owner and a next validation step? If the answer to several of these is no, do not ask the model to “make it more confident.” Fix the evidence and ownership. The useful mental model is simple: AI is a fast junior analyst with no relationship, authority, or accountability. In Don't Replace Me, Dmitry Kargaev calls out the difference between speed and judgment. Retention work is exactly where confusing those two gets expensive. Frequently asked questions Can ChatGPT predict which customers will churn? Not reliably from a prompt and a handful of notes. Valid prediction requires defined outcomes, representative historical data, evaluation, monitoring, and expert review. Even then, a score estimates a pattern; it does not reveal why one person will leave. What data should I give an AI retention prompt? Use the minimum approved, sanitized evidence needed for the decision: metric definitions, aggregate patterns, de-identified themes, dates, known limitations, and prior commitments. Keep names, contact details, contracts, payment data, raw private conversations, and credentials out of unapproved tools. Can AI write a churn-prevention email? It can draft options from verified context. A human must confirm the facts, tone, promises, audience, and whether email is the right channel. Sensitive relationships usually deserve a real conversation, not automated concern theater. Should I offer a discount to every at-risk customer? No. A discount only addresses some forms of budget or value pressure, and it can hide product, support, or trust problems. Diagnose the issue, confirm it with the customer, and require commercial approval. How often should a retention review run? Match the cadence to the business and decision cycle. High-touch accounts may need regular human reviews; aggregate product signals may be reviewed weekly or monthly. Review sooner when a verified trigger appears, but avoid constant noisy alerts that train everyone to ignore them. Are customer health scores useful? They can be useful summaries when definitions, weights, data quality, and intended decisions are documented. They become dangerous when treated as truth, used outside their validated context, or allowed to replace direct evidence and customer conversation. Can AI replace a customer-success manager? No. It can reduce synthesis and drafting work. It cannot build trust, read a room, negotiate responsibly, own a promise, understand every hidden constraint, or accept accountability for the relationship. Use AI for the paperwork, not the relationship Good retention work is not magical prediction. It is disciplined listening, evidence review, responsible decisions, and follow-through. Use AI customer retention prompts to clean up the paperwork: sort signals, expose gaps, draft questions, and format plans. Keep humans in charge of the customer conversation, sensitive context, commercial judgment, and every promise that leaves the building. If you are still calibrating where the tool helps and where it absolutely does not, read what AI can and cannot do and the practical no-BS guide to using AI at work. The machine can make you faster. It should not make you less accountable. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Churn Analysis Prompts: Find Why Customers Leave 10 AI churn analysis prompts to investigate why customers leave, test assumptions, protect customer data, and turn evidence into accountable action. AI promptschurn analysiscustomer retention AI Customer Renewal Prompts: Prepare With Evidence 10 AI customer renewal prompts to verify outcomes, expose account risks, prepare options, and protect trust without inventing value or customer sentiment. AI promptscustomer renewalcustomer success AI Customer Success Plan Prompts: Use Real Evidence 10 AI customer success plan prompts to verify goals, map adoption gaps, assign milestones, and build an account plan without inventing customer value. AI promptscustomer success plancustomer success --- ## AI Customer Segmentation Prompts: Find Useful Groups URL: https://dontreplace.me/articles/ai-customer-segmentation-prompts-find-useful-groups AI promptscustomer segmentationmarket researchcustomer researchdata analysis AI Customer Segmentation Prompts: Find Useful Groups 10 AI customer segmentation prompts to audit data, find behavioral groups, test bias, validate segments, and turn evidence into responsible action. August 25, 2026 Updated August 25, 2026 Customer segmentation has a branding problem disguised as an analysis problem. Give three clusters cute names, add stock photos, and everyone starts discussing “Budget Brenda” as if she walked into the research lab carrying a laminated biography. AI customer segmentation prompts can help audit data, propose testable groupings, compare patterns, expose missing evidence, and format segment briefs. They cannot observe customers, infer identity or motivation as fact, repair a biased sample, validate demand, decide who deserves service, or make targeting choices for you. AI can organize possible groups. Humans must prove the groups are real, useful, lawful, fair, and worth acting on. These ten templates are for product managers, marketers, UX researchers, analysts, customer-success teams, and operations leads. They focus on behavior, needs, context, and outcomes rather than demographic fan fiction. If you need better source material first, use AI customer interview prompts or AI survey design prompts. What useful customer segmentation actually does A segment is a group of customers who share a meaningful characteristic that should change a real decision. That last part matters. If two groups receive the same product, message, support model, and experience, separating them may produce a colorful dashboard without producing value. Useful segmentation usually has five properties: Observable: the grouping uses evidence you can define and inspect. Distinct enough: differences are meaningful, not statistical confetti. Reachable: the organization can identify or serve the group appropriately. Stable enough: membership lasts long enough to support the decision. Actionable: the segment changes a product, service, research, or communication choice. Segmentation is not the same as creating personas. A segment might be “teams attempting their first integration with no dedicated administrator,” supported by setup behavior and account context. A persona called “Scrappy Sam, 34, loves podcasts and hates meetings” is mostly a screenplay unless research supports each detail—and even then, those details may not affect the decision. Customers can belong to more than one segment. They can also move between segments as their needs, skill, circumstances, or relationship with the product change. A first-time buyer becomes a repeat buyer. A small team becomes a regulated enterprise. A customer who needs speed today may need certainty tomorrow. Good segmentation represents that movement instead of pretending everyone receives a permanent sorting hat. If your evidence spans multiple channels and stages, AI customer journey mapping prompts can help connect segment differences to the experience without inventing a universal journey. The reusable customer segmentation prompt formula Add this instruction to any template below: “Act as a customer-segmentation analysis assistant. I am making [specific decision] for [defined customer population, market, time period, and context]. Use only the sanitized evidence I provide. Produce [artifact]. Separate observed variables, calculated measures, reported needs, interpretations, and hypotheses. For every proposed segment, include the rule, evidence, sample size, confidence, limitations, overlap, likely movement, excluded customers, fairness and privacy risks, validation method, and action it would change. Do not invent identities, motivations, quotes, protected characteristics, prevalence, causality, or market demand.” This formula makes uncertainty visible. “Accounts that used feature X at least twice in 30 days renewed more often in this dataset” is an observation with definitions and a date range. “Power users love innovation” is a story. “Promote feature X to improve retention” is a hypothesis requiring a test. AI often slides from the first statement to the third while sprinkling in the second. Your prompt should stop that little magic trick. Never paste names, email addresses, account IDs, raw transcripts, exact locations, health or financial details, protected characteristics, credentials, confidential strategy, or legally sensitive data into an unapproved AI tool. Aggregation is safer than row-level data. De-identification helps, but rare combinations can still identify people. Use approved systems, minimum necessary data, access controls, retention limits, consent, and human privacy and fairness review. What to collect before prompting “Segment our customers” is not a useful request. The model will happily return enterprise, mid-market, and small business because those labels are nearby in language, not because your evidence supports them. InputWhy it mattersHuman check Decision to informDefines what “useful” meansDecision owner confirms Population and exclusionsSets the denominatorResearch or data owner approves Time periodPrevents mixing stale and current behaviorAnalyst verifies Behavioral eventsGrounds groups in actionsInstrumentation owner checks Outcome definitionsConnects groups to consequencesBusiness owner approves Research evidenceExplains needs and contextResearcher verifies Data provenanceReveals quality and consent limitsData steward reviews Missingness and sample biasShows who disappears from the analysisAnalyst documents Sensitive variables and proxiesPrevents harmful inferencePrivacy or legal reviews Operational constraintsTests whether groups are reachableDelivery owner confirms Existing segment rulesEnables comparison and migrationCRM owner verifies Validation planKeeps hypotheses from becoming doctrineResearch lead approves Sanitize and aggregate inputs before prompting. Keep traceable source IDs in approved systems so authorized humans can check the evidence. If fields have unclear definitions, resolve that before clustering. A column called active_user may mean logged in once, used a paid feature, or merely avoided cancellation. Those are not interchangeable. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI customer segmentation prompts Replace brackets with sanitized, verified details. Treat every output as a draft for human review. 1. Define the decision and segmentation scope “Using this proposed business decision, customer population, market, date range, available evidence, operational constraints, and excluded use cases: [paste], draft a segmentation scope. State the decision segments must improve, who is represented and missing, the unit of analysis, required stability, acceptable overlap, sensitive uses that are prohibited, and success measures. Flag vague goals and offer two narrower scopes if this one combines unrelated decisions.” “Improve marketing” is too broad. Segments for onboarding design may differ from segments for pricing research or support staffing. Trying to build one segmentation model for every department usually creates bland groups that satisfy nobody and linger in slide decks forever. A human owner should approve the decision before analysis begins. Otherwise teams can reverse-engineer segments to justify an idea they already prefer. 2. Audit the available customer data “Review this sanitized data dictionary, source inventory, collection methods, date coverage, samples, known quality issues, and consent restrictions: [paste]. Create a segmentation-readiness audit. For each variable, record its definition, source, coverage, missingness, freshness, reliability, sensitivity, possible proxy effects, and relevance to the decision. Flag leakage, duplicate customers, inconsistent units, survivorship bias, inaccessible populations, and fields that should not be used.” This prompt helps you find the boring problems that destroy exciting models. Your CRM may overrepresent customers who talked to sales. Product analytics may exclude blocked scripts or offline activity. Support records contain people who had enough time and confidence to complain. Survey respondents selected themselves. For quantitative work, pair the audit with AI data analysis prompts, then have an analyst verify queries, joins, filters, denominators, and missing-value treatment. 3. Separate observed variables from inferred attributes “Classify these proposed segmentation variables and labels: [paste]. Put each into observed behavior, transaction, stated preference, reported need, contextual fact, calculated measure, inferred attribute, protected or sensitive characteristic, or unsupported label. Explain the evidence required for each. Flag variables that may act as proxies for protected characteristics and rewrite invented motivations as testable questions.” A customer opened five help articles. That is observed behavior. They are “confused” is an interpretation. They are “not technical” may be an unsupported identity claim. Perhaps the documentation is bad. Perhaps five colleagues shared an account. Perhaps they were researching for someone else. Do not infer sensitive traits because a model can produce them. Ability is not permission. Privacy, legal, and fairness reviewers should approve allowed variables and uses. 4. Propose testable behavioral segments “Using only these approved behavioral variables, needs evidence, contexts, and outcomes: [paste], propose three to six candidate segments relevant to [decision]. For each, provide an explicit membership rule, supporting evidence, sample or coverage, distinguishing behavior, need hypothesis, overlap with other groups, expected movement over time, unknowns, and the different action the segment could justify. Reject clusters that lack a distinct action.” Ask for candidate segments, not final truth. A useful output might distinguish customers who fail before first value from those who reach value but cannot repeat it. The action changes: one group may need setup repair while the other needs workflow support. Avoid asking the model to assign personality names. Plain labels such as “first-time admins blocked during configuration” are less memorable, but much harder to mistake for a human essence. 5. Compare segment needs without inventing motives “Compare these candidate segments using the supplied interview themes, survey results, behavioral data, outcomes, and journey evidence: [paste]. Build a matrix of observed behavior, directly reported needs, constraints, triggers, desired outcomes, barriers, evidence strength, contradictory findings, and missing research. Keep motivations as hypotheses unless customers stated them. Do not write fictional quotes or demographic backstories.” The comparison should preserve contradictions. Ten customers in one candidate segment may share behavior while describing different reasons. That could mean the grouping is too broad, context changes the need, or the behavior does not reveal motivation. If the evidence is mostly desk research, use AI research prompts to track sources and uncertainty instead of letting a summary become folklore. 6. Check sample, measurement, and proxy bias “Stress-test this proposed segmentation and its source data: [paste]. Examine selection bias, survivorship bias, missing channels, inaccessible research methods, historical policy effects, small samples, measurement error, stale data, geographic or language gaps, proxy discrimination, and customers forced into an ‘other’ group. For each risk, explain the possible distortion, affected people, severity, evidence, mitigation, and reviewer required. Do not declare the model fair or representative.” Bias review is not a ceremonial paragraph at the end. Suppose “high engagement” requires desktop activity while field workers rely on mobile or shared devices. The measure may classify product access as customer commitment. A clean cluster can still encode a dirty system. Do not use AI to certify fairness. Have qualified humans review sensitive decisions, and include affected people in validation when possible. 7. Stress-test whether the segments are useful “Evaluate these candidate segments against the decision and operating constraints: [paste]. Score each on observability, distinctness, reachability, stability, actionability, evidence strength, cost to maintain, privacy risk, fairness risk, and likely value. Show the reasoning and uncertainty behind each score. Identify segments that lead to the same action, depend on unavailable data, are too small to use safely, or change too quickly to operationalize.” This prompt kills decorative segmentation. If two groups need the same intervention, merge them unless another decision requires the distinction. If membership requires a data science expedition every morning, the segment may not be operationally reachable. Humans should make the final tradeoff. A segment can be measurable and profitable while still being invasive, exclusionary, or strategically foolish. 8. Draft evidence-backed segment cards “Turn these validated candidate segments into concise segment cards: [paste]. For each card include a descriptive label, decision served, membership rule, qualifying evidence, observed behaviors, directly reported needs, context, desired outcomes, exclusions, overlap, movement triggers, confidence, limitations, privacy restrictions, prohibited inferences, and source references. Add a visible ‘what we do not know’ section. Do not add names, ages, biographies, stock-photo directions, or fictional quotes.” A segment card should help someone make a decision without pretending to know a stranger’s soul. Include the rule and evidence prominently. If the card can survive after those are removed, it is probably creative writing. Review cards with researchers, analysts, frontline teams, privacy stakeholders, and—when appropriate—customers represented by the grouping. 9. Design a validation plan “Create a validation plan for these candidate segments and claims: [paste]. Include quantitative replication, holdout or later-period checks, qualitative sampling, accessibility and language coverage, contradictory cases, segment reassignment tests, actionability tests, privacy and fairness review, failure criteria, owners, timeline, and decision gates. Prioritize the assumptions most likely to change the proposed action.” Validation is more than asking customers whether a label “resonates.” People may dislike a label while the underlying behavioral distinction remains useful, or enjoy a persona while its membership rule predicts nothing. Test whether the segments reproduce, whether customers move as expected, and whether tailoring an action improves an agreed outcome without creating unacceptable harm. A segment that cannot survive new evidence should retire with dignity. 10. Produce a human-reviewed activation plan “Using only these validated segments, approved uses, operational constraints, and measures: [paste], draft an activation plan. For each segment list the permitted action, customer benefit, channel, eligibility rule, owner, implementation dependency, accessibility requirement, privacy control, fairness check, success metric, guardrail metric, review date, and rollback trigger. Include a control or comparison where appropriate. Mark every strategic choice for human approval.” Activation is where harmless-looking analysis gains consequences. A group may receive different onboarding, support, messaging, or research invitations. Document those effects. Customers should not lose access, receive worse service, or face sensitive targeting because a speculative model put them in a box. For consequential launches, use AI risk assessment prompts before deployment. The accountable owner—not the model—approves the action and rollback threshold. A simple human review workflow A defensible process can be boring. Boring is underrated when customer data and business decisions are involved. Name one decision. Do not segment the universe. Audit sources and permissions. Exclude unsafe or undefined fields. Create candidate groups. Keep rules explicit and labels descriptive. Review bias and missing people. Look for proxies and invisible populations. Validate with multiple methods. Combine behavioral and qualitative evidence. Test actionability. Confirm that groups change a legitimate decision. Approve permitted uses. Document what the segments must never be used for. Measure outcomes and harms. Track benefit, exclusion, complaints, and drift. Revisit membership. Customers and markets change. Retire weak segments. Do not preserve a bad model because the slide template is expensive. If the output is for marketing, AI marketing prompts can help draft variations after the segment and claims are validated. Human reviewers still own accuracy, accessibility, targeting ethics, and final publication. Common segmentation mistakes AI makes faster It invents motivations from behavior The model sees low usage and writes “price-sensitive casual users.” Low usage could mean poor onboarding, seasonal need, accessibility barriers, missing tracking, or a product that failed them. Keep motivation separate from observation. It treats clusters as natural species A clustering algorithm partitions rows according to chosen variables and settings. That does not reveal eternal tribes hiding in the spreadsheet. Different variables, scaling, time windows, or cluster counts can produce different groups. It uses sensitive proxies Location, device, language, purchase history, and browsing patterns can correlate with protected or sensitive characteristics. Review variables, uses, and effects—not just whether a protected field appears by name. It hides the people with missing data Customers with sparse records often land in “other,” disappear during cleaning, or get assigned using weak signals. Missingness may itself reveal exclusion, channel differences, or broken instrumentation. It confuses precision with truth “Segment 4 has a 63.7% propensity” looks scientific. Without definitions, uncertainty, validation, and a suitable test set, the decimal is just wearing a lab coat. It makes every segment sound marketable Reality includes overlapping needs, contradictory behavior, and groups that are not worth separate treatment. You do not need a slogan for every row in a table. Frequently asked questions Can ChatGPT create customer segments from a spreadsheet? It can help inspect a sanitized schema, suggest analysis questions, draft code, explain candidate patterns, and format findings. It should not receive raw identifiable customer data through an unapproved tool. An analyst must verify cleaning, definitions, methods, results, and limitations. For sensitive or consequential uses, privacy, legal, and fairness review are also required. What data should I use for customer segmentation? Use data relevant to the decision: verified behavior, transactions, directly reported needs, context, outcomes, and research evidence. Prefer variables you can define and justify. Do not include sensitive characteristics or proxies merely because they improve separation. More columns do not automatically create more useful segments. How many customer segments should we have? As few as needed to support distinct actions. Three useful groups can beat twelve impressive ones. The right number depends on evidence, operational capacity, overlap, stability, and the decision. If teams cannot explain or serve the groups consistently, simplify. Are AI-generated customer personas reliable? Not by default. AI is very good at filling gaps with plausible biography. Build segment cards from evidence and label unknowns. If you create personas for communication, trace every material detail to research and avoid turning a diverse group into one fictional mascot. Should segments be based on demographics or behavior? Behavior, needs, context, and desired outcomes are often more actionable and less stereotyped. Demographics may be relevant for some legitimate research or equity questions, but they require careful purpose, consent, governance, and review. Never assume a demographic trait explains motivation. How often should customer segments be updated? Set a review date based on how quickly the market, product, and customer behavior change. Also review after major pricing, product, policy, channel, or measurement changes. Monitor whether group sizes, outcomes, movement, and missingness drift. Retire segments that no longer reproduce or change action. Can AI remove bias from customer segmentation? No. AI can help list bias risks, inspect definitions, generate test code, and document mitigations. It can also reproduce and conceal bias from the data and prompt. Qualified humans must define fairness goals, inspect effects, include affected perspectives, and decide whether the use is acceptable. What is the biggest sign a segment is useless? Nothing changes when you know someone belongs to it. If the same product, service, research plan, and communication apply to every group, the distinction may be descriptive rather than actionable. Another warning sign is a label everyone recognizes but nobody can calculate consistently. The point is better decisions, not better stereotypes AI customer segmentation prompts are useful when they force teams to define evidence, uncertainty, membership rules, bias risks, validation, and action. They are dangerous when polished output turns guesses into customer truth. Start with one decision and the minimum approved evidence. Keep observed behavior separate from inferred motives. Include missing and contradictory cases. Test whether a grouping survives new data and improves a legitimate outcome. Then keep a human accountable for every consequence. That is the larger survival skill behind using AI at work without the nonsense: use the machine for speed and structure, while humans keep context, judgment, responsibility, and taste. For the broader argument, what AI can and cannot do is the five-minute version. If you want the blunt field guide for making that division of labor work, Don't Replace Me by Dmitry Kargaev takes it further. The machine may produce the segments. You still have to decide whether they describe reality—or merely make the meeting feel productive. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Churn Analysis Prompts: Find Why Customers Leave 10 AI churn analysis prompts to investigate why customers leave, test assumptions, protect customer data, and turn evidence into accountable action. AI promptschurn analysiscustomer retention AI Customer Persona Prompts: Build Useful Profiles 10 AI customer persona prompts to turn verified research into useful profiles, expose assumptions, reduce stereotypes, and plan human validation. AI promptscustomer personasbuyer personas AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation --- ## AI Customer Service Prompts: 10 Templates That Work URL: https://dontreplace.me/articles/ai-customer-service-prompts-10-templates-that-work AI promptscustomer serviceChatGPT at workAI at worksupport templates AI Customer Service Prompts: 10 Templates That Work 10 copy-paste AI customer service prompts for ChatGPT and Claude. Draft replies, de-escalate complaints, and summarize tickets without sounding like a robot. May 30, 2026 Updated May 30, 2026 Your inbox has 47 unread tickets. Three of them are angry. One is in all caps. You have a meeting in 20 minutes and no idea what "UNACCEPTABLE!!!" actually wants because the original issue is buried in six reply-all threads. AI customer service prompts won't fix your whole job. But they will help you get through that inbox faster, without every reply sounding like it was generated by a bored robot who just learned what empathy is. This is a practical template article. No fluff, no theory. Just the formula, the prompts, and the rules you need to not embarrass yourself. The one formula behind every good AI customer service prompt Before the templates, here's the structure that makes them work. Think of it as a recipe. Swap the ingredients, keep the shape. Role. Tell the AI what it is. "You are a customer support agent for a home goods company." Context. Give it the situation. What happened, who's involved, what the customer said. Task. Say what you want. Draft, summarize, rewrite, translate, shorten, apologize. Constraints. What it can't do. Don't promise refunds. Don't mention the specific team name. Keep it under 150 words. Format. What you want back. One paragraph, three bullet points, a subject line and body, etc. That's it. Role, context, task, constraints, format. If your AI output is garbage, one of those five things is missing. This is what Rule #13 in Don't Replace Me calls the "smart intern" model: the AI is fast, tireless, and occasionally wrong, and it needs a clear brief the same way any intern would. The difference between a prompt that produces something useful and one that produces a two-paragraph word salad is almost always the constraints. "Write a reply" gets you something generic. "Write a reply under 120 words that acknowledges the delay, explains the next step, and doesn't promise a specific delivery date" gets you something you can actually edit and send. Specificity is the whole game. A quick word on what not to paste Before anything else: don't paste real customer names, email addresses, account numbers, payment details, medical information, private employee notes, or confidential company data into a public AI tool unless your company has explicitly approved it. Anonymize everything. "Customer in Manchester" not the customer's name and postcode. "Order placed in early March" not the exact order ID. "Customer mentioned a health condition" not the condition itself. If you're using a company-approved AI tool with a proper data agreement, follow your company's specific policy. If you're using a personal ChatGPT or Claude account, assume nothing is private. Verify everything before you send. Refund policies, warranty terms, delivery timelines, legal disclaimers. AI will confidently state whatever sounds plausible. You are the one signing off on the reply. Don't send anything you wouldn't stand behind yourself. Right. On to the prompts. 10 AI customer service prompts you can use today These are copy-paste ready. Replace the bracketed parts with your actual details. 1. Summarize a messy ticket When a ticket has six replies, two departments, and a lot of "as per my previous email" energy. You are a customer support agent. Here is a messy email thread from a customer: [paste anonymized thread]. Summarize the core issue in 3 bullet points: what the customer wants, what has already been tried, and what is still unresolved. Keep it factual. Do not add opinions. 2. Draft a first response For when you understand the issue but don't want to stare at a blank compose window. You are a customer support agent for [type of company]. A customer has contacted us about [brief description of issue]. Draft a professional, warm first response that: acknowledges their frustration, confirms we've received their request, tells them the next step, and gives a realistic timeframe. Keep it under 150 words. Do not promise a specific outcome or compensation. 3. Rewrite a reply to sound more human For when your draft sounds like it was written by a compliance document. Rewrite this customer service reply to sound warmer and more natural, like a real person wrote it. Keep the same information. Don't add fluff or empty phrases like "I hope this email finds you well." Keep it under [X] words. Here is the original: [paste your draft]. 4. Apologize without overpromising One of the hardest things to write. You need to sound sorry without accidentally committing to a full refund you can't authorize. You are a customer support agent. A customer is upset because [brief description of the problem]. Write a sincere apology email that: acknowledges what went wrong, takes responsibility without blaming other teams, expresses genuine regret, and explains the next step we can take. Do not promise a refund, compensation, or specific resolution. Do not use the phrase "we apologize for any inconvenience." Keep it under 120 words. 5. Explain a policy clearly For translating internal jargon into words a human being can parse. Here is our policy on [topic, e.g., returns / shipping delays / account cancellations]: [paste anonymized policy text]. Rewrite this in plain language for a customer who is frustrated and not familiar with our terms. Keep it factual, friendly, and under 100 words. Do not add exceptions that aren't in the original policy. 6. De-escalate an angry customer The all-caps ticket. The person who has typed "DISGUSTING" and means it. You are an experienced customer support agent. A customer has sent this message: [paste anonymized message]. They are clearly very upset. Draft a de-escalation response that: validates their frustration without being defensive, avoids corporate-speak, shows we take this seriously, and gives a clear next step. Do not be dismissive. Do not make promises we haven't authorized. Keep it under 150 words. 7. Ask for missing information For when a customer has described a problem in a way that tells you almost nothing useful. A customer has reported this issue: [brief description]. To investigate, we need [list of missing information, e.g., order number, date of purchase, photo of the item, error message]. Write a polite reply asking for this information without making them feel interrogated. Keep it under 100 words. Be specific about what we need and why. 8. Turn technical notes into customer language For when a developer or internal team has explained the fix in terms no customer should have to read. Here are technical notes from our team about a customer's issue: [paste anonymized technical notes]. Rewrite this as a clear, simple explanation for the customer. They are not technical. Avoid jargon. Explain what happened and what we did or are doing about it. Keep it under 120 words. 9. Write a follow-up after resolution The closing message that most support agents either forget or phone in. You are a customer support agent. We recently resolved an issue for a customer involving [brief anonymous description of issue]. Write a short follow-up message that: confirms the issue is resolved, thanks them for their patience, invites them to reach out if anything else comes up, and leaves them with a positive impression. Keep it under 80 words. Don't be sycophantic. 10. Summarize recurring complaints for the product team For turning your ticket queue into useful intel. Here are [X] anonymized customer complaints from the past [timeframe] about [product/feature/process]: [paste anonymized summaries]. Identify the top 3 recurring themes, note any common language customers use to describe the problem, and suggest one question each theme raises for the product or operations team. Format as: Theme, Customer Language, Question for the Team. These are the same type of practical, reusable frameworks covered in the AI prompts for work guide if you want to go broader than customer service. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → How to adapt these AI customer service prompts for your specific situation The ten templates above cover the most common scenarios. But customer service work is wildly varied, and "common" only gets you so far. Here's how to modify any of them when the situation doesn't quite fit. When the customer is a business, not an individual. Add "This customer is a B2B client with an ongoing account relationship" to your context block. The tone shifts slightly: more formal, more focused on minimizing disruption, less about emotional validation and more about practical resolution. Same structure, different emphasis. When you've already replied once and it didn't land. Add "We sent a previous reply that the customer found unhelpful. Here is what we said: [paste previous reply]. Do not repeat the same approach." This forces the AI to try a different angle rather than producing a slightly fancier version of what already didn't work. When there's a language barrier. If you can tell from the writing that English isn't the customer's first language, add "Write in simple, clear English. Avoid idioms, complex sentence structures, or phrases that don't translate well literally." It won't make the reply flawless, but it makes it more accessible. When you're working from a script or template your company already uses. Add "Match the tone and structure of this approved template: [paste template]. Adapt it for the specific situation described." The AI fills in the gaps without wandering off-brand. The formula stays constant. The context block is where all the customization happens. That's the part that requires your judgment, because you're the one who actually read the ticket. Why customer service is not just "language work" Here's the thing about using AI for support tickets. It's genuinely useful for the draft-shaping part: taking a raw situation and turning it into a coherent, appropriately toned reply. That part, AI does well. But customer service is also trust work. The customer on the other end of that ticket isn't just looking for correct words. They're looking for a sense that someone actually read their problem and actually gives a damn. AI doesn't give a damn. It patterns its way toward something that sounds like giving a damn. Your job is to make sure the gap between "sounds like it" and "actually is" stays small. That means reading the output and asking: does this feel right for this specific person? Is the apology actually proportionate to what went wrong? Is there something in this ticket that a template will make worse? Research from Salesforce consistently shows that customers who feel heard are more likely to stay loyal, even after a bad experience, than customers who received a technically correct but cold response. The words matter less than the sense that a real person thought about their situation. AI can help you get to good words faster. It can't do the thinking-about-their-situation part for you. For more on the human skills that remain yours even as AI gets better at drafts, the piece on jobs AI can't replace covers the empathy, judgment, and accountability side of things plainly. What happens when you skip the quality check Most AI-related mistakes in customer service don't happen because the AI said something obviously wrong. They happen because the person sending the reply didn't read it carefully enough. A real pattern: AI drafts a reply to a shipping delay complaint. The draft is warm, apologetic, well-structured. It also includes the line "your order will arrive within 3-5 business days." The actual timeline is 10-12 business days because of a warehouse issue the AI didn't know about. The agent sends it. The customer gets the order 11 days later, furious, citing the email you sent them. That's not an AI failure. That's a verification failure. The AI did what you asked. You didn't check. The same thing happens with refund commitments, warranty terms, stock availability, and policy exceptions. AI will write something that sounds authoritative whether or not it's accurate. A 2024 study in PLOS ONE found that large language models produce confident, fluent text even when the underlying information is incorrect, which is a feature, not a bug, for most tasks, but a liability when you're making commitments to customers on behalf of a company. The fix is simple: build one minute of verification into every AI-assisted reply. Check the one factual claim in the message. That's it. One minute. It's the difference between a tool that helps and a tool that creates more tickets than it closes. The quality check before you hit send A quick mental checklist before sending any AI-drafted support reply: Did I verify the policy, timeline, or process I'm describing? Am I making any promises I'm not authorized to make? Does this reply actually address what the customer asked? Does the tone match the severity of their frustration? Would I be comfortable if my manager read this reply? Is there any personal or confidential data I shouldn't have included? If you can say yes to all of these, send it. If something feels off, rewrite it. The AI gave you a draft, not a final answer. You're still the one with your name on it. For a broader walkthrough of building AI into your actual work day, the no-BS starter guide to using AI at work covers the mindset shift without any of the usual hype. Frequently asked questions Can I use AI to reply to customer service emails? Yes, but as a drafting tool, not an autopilot. AI can generate a starting point faster than you can from scratch, but you should verify every factual claim (policies, timelines, compensation terms) and read the final reply before sending. Don't paste real customer data into a personal AI account without company approval. What's the best way to prompt AI for customer service replies? Use the five-part structure: role, context, task, constraints, format. Tell the AI what kind of company it's representing, what the customer's issue is, what kind of reply you need, what it shouldn't promise or include, and how long the reply should be. Vague prompts produce vague output. Will AI make my customer service replies sound robotic? It can, if you use the output without editing it. AI-generated customer service text often defaults to hollow phrases like "we apologize for any inconvenience" and "please don't hesitate to reach out." Prompt it specifically to avoid those, then read the output and adjust the parts that feel cold or generic. Is it safe to paste customer information into ChatGPT? Not unless your company has an approved data agreement with the AI provider. Anonymize all tickets before pasting them: no names, email addresses, account numbers, order IDs, health details, or payment information. When in doubt, assume the tool is not private. What customer service tasks is AI actually bad at? Deciding whether a situation warrants an exception. Reading subtext in a complaint. Knowing when a policy technically applies but applying it would be the wrong call. De-escalation with context the ticket doesn't include. Any situation where the customer's actual need is different from what they wrote. These are judgment calls, and they stay yours. How do I know if an AI-drafted reply is good enough to send? Run it through the five-question check: Does it answer what was asked? Are all facts verified? Does the tone match the situation? Would I be comfortable if my manager read this? Does it sound like a human I'd want to talk to? If yes to all five, send it. If something feels wrong, it probably is. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Launch Checklist Prompts: Ship Without Gaps 10 copy-paste AI launch checklist prompts for ChatGPT and Claude. Turn messy plans into structured checklists with owners, risks, and rollback steps. ai promptslaunch checklistproject management AI Meeting Follow-Up Prompts (10 Templates) 10 copy-paste AI meeting follow-up prompts for ChatGPT and Claude. Turn messy notes into recaps, action items, and exec summaries without inventing decisions. ai promptsmeeting follow-upaction items AI Meeting Notes Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into summaries, action items, and client emails. With a reusable formula and safety reminders. AI promptsmeeting notesChatGPT at work --- ## AI Customer Success Plan Prompts: Use Real Evidence URL: https://dontreplace.me/articles/ai-customer-success-plan-prompts-build-a-plan-that-uses-real-evidence AI promptscustomer success plancustomer successaccount managementcustomer adoption AI Customer Success Plan Prompts: Use Real Evidence 10 AI customer success plan prompts to verify goals, map adoption gaps, assign milestones, and build an account plan without inventing customer value. September 10, 2026 Updated September 10, 2026 A customer success plan should explain what the customer is trying to achieve, what both sides will do, and how they will know whether it worked. Too often it becomes a handsome document assembled during onboarding, admired briefly, and abandoned in a CRM drawer until renewal panic begins. AI customer success plan prompts can help you audit evidence, organize goals, expose missing information, draft milestones, and prepare better questions. They cannot know what the customer values, infer satisfaction from silence, invent ROI, approve a commitment, or maintain a relationship while you attend another internal meeting about “customer centricity.” A useful success plan is a living agreement about outcomes and work. It is not account fan fiction with a logo on the cover. These ten templates help customer success managers, account managers, founders, implementation leads, and revenue operations teams create plans that are specific enough to use and honest enough to trust. What belongs in a customer success plan? A practical plan connects current customer goals to verified evidence, concrete actions, named owners, review dates, and decisions. It should answer: Which customer outcomes are current and confirmed? How will each outcome be measured? What baseline and current evidence exist? Which product behaviors may support the outcome? What adoption, support, delivery, or relationship gaps could block progress? Who has authority to decide and approve? What will the customer do, and what will your team do? When will both sides review progress and change the plan? Which statements are facts, hypotheses, proposals, or unknowns? Keep those categories separate. A verified fact has a current source. A customer statement records what someone actually said. A hypothesis needs testing. A proposal needs approval. A commitment has an authorized owner and date. An unknown needs a question rather than a confident paragraph. This matters because language models are unusually good at turning incomplete records into complete-sounding stories. If the wider account context is scattered, use AI account planning prompts before pretending the success plan is ready. If the plan will feed a formal review, pair it with AI QBR prompts. Use this evidence-first customer success prompt formula Add this instruction to every template below: “Act as a customer success planning assistant. Use only the sanitized, approved sources I provide for [customer goals, measures, adoption, support, delivery, stakeholders, risks, milestones, and decisions]. Cite a source label and date for every goal, metric, customer statement, deadline, stakeholder role, and commitment. Separate verified facts, customer statements, hypotheses, proposals, commitments, and unknowns. Mark stale, conflicting, or unsupported claims [VERIFY]. Do not invent sentiment, business value, ROI, causation, decision authority, approval, urgency, product capabilities, or future results.” The source requirement does most of the useful work. Without it, a model may combine an onboarding objective, an old CRM note, three recent logins, and one cheerful email into a magnificent strategic outcome that exists nowhere outside the draft. Never paste customer names, emails, contracts, credentials, private CRM notes, raw usage exports, support transcripts, payment details, confidential pricing, security findings, regulated information, health data, or legally sensitive material into an unapproved AI tool. Use approved systems, minimum necessary data, de-identification, access controls, retention limits, and human privacy, security, legal, finance, product, and customer-success review. What should you collect before using AI? Build a small source pack instead of uploading the complete archaeological record of the account. InputWhat it can supportHuman owner Confirmed customer goalsOutcomes the plan should pursueCustomer and account owner Agreed measure definitionsWhat success means and excludesCustomer, analytics, or business owner Baseline and current resultsEvidence of change over timeAnalytics or process owner Dated usage summaryAdoption patterns and product activityProduct or analytics owner Implementation statusDelivered work, dependencies, and blockersImplementation owner Support and incident summaryKnown friction and unresolved issuesSupport owner Stakeholder notesStated priorities, roles, and participationAccount owner Approved commitmentsWork authorized by each sideAuthorized owners Product capability statementsWhat the product currently supportsProduct owner Review cadenceWhen evidence and priorities get updatedCustomer and account owner Data-use rulesWhat may enter the AI systemPrivacy or security owner If sources disagree, preserve the conflict. Choosing whichever number makes the dashboard greener is not data cleaning. It is optimism with administrative privileges. 10 AI customer success plan prompts Replace brackets with approved, sanitized information. Review every output before it reaches a customer, CRM, project system, presentation, or executive report. 1. Audit the source pack “Compare this sanitized customer success source pack with these planning requirements: [paste]. Create a table with required input, supplied evidence, source, date, status, conflict, missing information, owner, and next action. Do not fill gaps by inference. Mark stale or unsupported items [VERIFY].” Run this before asking for strategy. It catches goals copied from a proposal, metrics with no definition, old stakeholder assumptions, incomplete implementation notes, unresolved incidents, and commitments that mysteriously belong to “the team.” AI can identify that a field is empty. It cannot determine whether the missing evidence is harmless, politically sensitive, or the entire reason the account is struggling. Route each gap to a person who can verify it. 2. Separate outcomes from product activity “Classify each supplied item as product activity, adoption signal, measured business outcome, customer statement, internal interpretation, hypothesis, or unknown. Include the reporting period, source, measure definition, caveat, and customer-confirmation status. Do not convert activity into value or claim causation without evidence.” Logins, active seats, workflows created, reports exported, and features clicked can describe activity. They do not automatically prove that work became faster, cheaper, safer, more accurate, or less annoying. A good plan may say, “Weekly active use increased, but business impact has not yet been measured.” That is less dramatic than announcing transformational value. It is also a sentence you can defend when the customer asks how you know. 3. Clarify goals and success measures “For each customer goal in these approved notes, create a planning row with exact goal statement, source, last-confirmed date, business context, baseline, proposed measure, measure owner, target or decision rule, time period, caveat, and question to confirm with the customer. Label inferred or outdated goals [RECONFIRM].” Goals decay. A priority agreed during implementation may be irrelevant after a reorganization, budget shift, policy change, acquisition, leadership departure, or Tuesday. Ask the customer whether the outcome still matters and whether the proposed measure represents it. AI can draft a metric menu. It cannot decide that saving five minutes per task matters more than reducing errors, improving compliance, or keeping an exhausted team from quitting. 4. Map stakeholders and decision authority “Create a stakeholder map from these dated, approved records: [paste]. Include role, stated goal, known concern, plan responsibility, participation, authority evidence, relationship owner, last verified date, and unanswered question. Put inferred influence or sentiment in a separate hypothesis column. Do not profile people using demographic or unrelated personal information.” The daily user, champion, executive sponsor, technical owner, security reviewer, budget holder, procurement contact, and signer may all be different people. A senior title does not prove authority, and a friendly message does not prove sponsorship. Use the map to identify missing voices and questions. Do not use it to manufacture synthetic intimacy. The dependable method for learning what a stakeholder thinks remains embarrassingly old-fashioned: ask them and listen. 5. Diagnose adoption gaps without inventing causes “Review these sanitized usage summaries, workflow notes, training records, and customer statements: [paste]. Build an adoption-gap table with observed signal, source, date, affected goal, possible explanations clearly labeled as hypotheses, evidence needed, customer question, owner, and next checkpoint. Do not infer intent, competence, satisfaction, or root cause.” Low usage can reflect missing permissions, poor training, broken integrations, weak fit, seasonal work, a changed process, bad data, accessibility barriers, or a feature nobody needed in the first place. AI can group patterns and propose questions. Humans need to investigate. If a gap spans a broader workflow, AI workflow audit prompts can help map the process before you prescribe another webinar as medicine. 6. Organize support, delivery, and risk signals “Review these approved support summaries, incident records, implementation notes, dependencies, and commitments: [paste]. Create a risk register with observed issue, source, date, known impact, affected outcome, severity rationale, current response, owner, customer visibility, unresolved question, escalation condition, and next review. Do not assign churn probability or root cause without a validated method.” A delayed integration, repeated ticket, missing capability, data-quality problem, or disputed commitment can block the plan even when adoption charts look healthy. Surface uncomfortable evidence early rather than hiding it beneath a motivational milestone. For a deeper check, use AI risk assessment prompts. The model can organize known signals. It cannot decide whether a customer relationship is safe, whether a security issue is acceptable, or whether an executive needs to be called today. 7. Draft milestones with owners and dependencies “Turn these confirmed goals, verified constraints, and approved commitments into a draft milestone plan. For each milestone include outcome supported, deliverable, named owner, due date, dependency, evidence of completion, review point, escalation path, and approval status. Keep customer actions, vendor actions, proposals, and confirmed commitments separate. Do not invent owners or dates.” A milestone should change the state of the work. “Improve adoption” is an aspiration. “Customer operations owner confirms the pilot workflow with ten approved users by October 15” is checkable, assuming the owner and date actually agreed. Do not let generated symmetry create fake obligations. If your team has five tasks and the customer has two, that may be reality. A balanced table is not automatically a fair agreement. 8. Prepare discovery and review questions “Using these verified goals, evidence gaps, adoption signals, risks, and stakeholder unknowns: [paste], draft open, neutral questions grouped by outcomes, measures, workflow, adoption, support, priorities, authority, and next-step tradeoffs. For each question, state which unknown it tests. Avoid leading wording, sales pressure, and assumptions about satisfaction or intent.” Useful questions include: “Which outcome matters most now?” “What changed since implementation?” “Where does the current workflow still create friction?” “Which measure would your team trust?” “Who must approve this milestone?” A bad question contains its preferred answer: “Given the incredible value delivered, would you agree expansion is the obvious next step?” That is not discovery. It is a sales pitch trying to sneak into the meeting wearing a fake mustache. 9. Create a review cadence and change log “Design a review cadence for this success plan using the supplied constraints: [paste]. Include meeting purpose, frequency, participants, evidence required, decisions expected, owner, update deadline, and escalation trigger. Then create a change-log template with date, changed item, previous version, new version, reason, source, approver, and affected milestones.” A plan is useful only if someone updates it. Monthly working reviews may track milestones and blockers; quarterly reviews may revalidate goals and measures; urgent risks may need separate escalation rather than waiting for calendar etiquette. Record consequential choices with AI decision log prompts. Otherwise, the account will spend next quarter debating whether a commitment was agreed, suggested, or hallucinated by an impressively formatted recap. 10. Draft the plan recap and human sign-off check “Draft a concise customer success plan recap from these approved notes: [paste]. Include confirmed goals, measures, current evidence, milestones, named owners, review dates, open risks, unknowns, and proposed items awaiting approval. Then audit the draft for unsupported claims, invented agreement, privacy risk, vague ownership, unapproved product or commercial language, missing dependencies, and specialist-review needs. Mark every issue [VERIFY]. Do not send or approve the document.” The recap should preserve disagreement and uncertainty. If the customer questioned a baseline, say so. If product must verify a capability, do not convert that into a roadmap promise. If nobody accepted a date, leave it open and assign the follow-up. The accountable customer-success owner should review the complete plan. Product, implementation, support, analytics, finance, legal, security, and privacy owners should verify claims in their domains. The customer confirms their goals, meaning, responsibilities, and commitments. AI approves none of it. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → A simple workflow for building the plan Use the prompts in sequence rather than asking a model to “create a winning customer success strategy” from a CRM dump: Define the planning purpose. Name the outcomes and decisions the plan must support. Collect minimum necessary evidence. Use current, approved, sanitized sources. Audit the source pack. Find gaps, stale records, and conflicts. Confirm goals with the customer. Do not inherit them forever from a sales deck. Define measures carefully. Separate activity, adoption, and business outcomes. Investigate gaps and risks. Treat possible causes as hypotheses. Agree on milestones. Use authorized owners, dates, and dependencies. Document decisions and proposals separately. Formatting does not equal approval. Review the plan on a real cadence. Update goals when reality changes. Use the plan in QBRs and renewals. Do not rediscover the account from scratch. For renewal preparation, use AI customer renewal prompts without turning the success plan into evidence retrofitted around a commercial target. If retention risk appears across accounts, AI churn analysis prompts can structure the investigation while keeping correlation and causation separate. Common mistakes AI can make worse Treating the sales promise as the customer goal A proposal may describe hoped-for value, not a current and confirmed objective. Revalidate goals with the people doing and owning the work. Turning usage into ROI Product activity may contribute to an outcome, but the connection needs an agreed method and evidence. AI cannot establish causation because two columns moved in the same direction. Guessing customer health from tone A polite email is not proof of satisfaction. A delayed reply is not proof of churn. Use stated feedback, verified behavior, and direct questions instead of amateur emotional forensics. Hiding unresolved support behind milestones Known incidents, recurring tickets, missing capabilities, and trust problems belong in the plan. A fresh adoption campaign does not erase a broken integration. Letting generated text become a commitment AI cannot authorize delivery dates, pricing, staffing, product features, security exceptions, legal terms, or customer obligations. Proposals require real approval from people with authority. Freezing the plan after onboarding Goals, stakeholders, constraints, and measures change. Date material sources, keep a change log, and review the plan before it becomes an elegant description of a customer who no longer exists. For the broader boundary, read what AI can and cannot do. Fast drafting is useful. It is not judgment, authority, empathy, or trust. Frequently asked questions Can ChatGPT write a complete customer success plan? It can draft a structure from approved, sanitized evidence and flag missing fields. Humans must verify every goal, measure, stakeholder role, risk, date, capability, and commitment. Do not upload raw customer systems to an unapproved tool or send generated plans without review. What is the best AI prompt for customer success planning? The best prompt requires dated sources, explicit evidence categories, visible unknowns, and strict limits on inference. Tell the model not to invent sentiment, outcomes, causation, authority, or approval. Ask for checkable tables before polished prose. How is a customer success plan different from an account plan? A customer success plan centers on the customer's confirmed outcomes, measures, joint work, and review cadence. An account plan may include broader relationship, commercial, territory, and internal strategy. They can share verified facts, but sensitive internal strategy should not leak into a customer-facing plan. Should the customer approve the success plan? The customer should confirm their goals, success measures, responsibilities, priorities, and commitments. Your internal owners must approve statements and obligations in their domains. Approval should be recorded clearly; a meeting attendee nodding near the end of a call is not durable governance. Can AI score customer health? AI can organize supplied signals or apply a validated scoring method. It cannot decide which signals matter, infer customer intent, or guarantee a churn prediction. Document definitions, limitations, data quality, and human review before using any score for consequential action. What customer data is safe to use with an AI tool? Use only information allowed by your organization's policies, contracts, access rules, and approved system configuration. Minimize and de-identify data where possible. Keep personal information, credentials, contracts, raw private communications, payment details, confidential pricing, security findings, and regulated data out of unapproved tools. How often should a customer success plan be updated? Update it whenever goals, stakeholders, evidence, risks, commitments, or constraints materially change. Use a regular working cadence for milestones and a periodic strategic review for goals and measures. Do not wait for renewal season to discover that the plan is six months stale. Who owns the final customer success plan? A named customer-success or account owner should be accountable for maintaining the plan. Customer stakeholders confirm their side; product, implementation, support, analytics, legal, finance, security, and privacy owners verify relevant claims. AI can draft and inspect the document, but it owns nothing. The useful boundary The strongest use of AI in customer success planning is controlled preparation. It can expose missing inputs, separate activity from outcomes, organize risks, draft neutral questions, and turn confirmed decisions into a readable plan. The work customers actually experience remains human: listening, understanding context, acknowledging problems, choosing tradeoffs, making authorized commitments, and following through after the meeting ends. That is the larger point of Don’t Replace Me by Dmitry Kargaev. Use the machine for speed and structure, then keep judgment and accountability attached to people. AI can make the plan look finished. Only humans can make it true. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Customer Renewal Prompts: Prepare With Evidence 10 AI customer renewal prompts to verify outcomes, expose account risks, prepare options, and protect trust without inventing value or customer sentiment. AI promptscustomer renewalcustomer success AI Customer Retention Prompts: Keep Customers 10 AI customer retention prompts to diagnose churn signals, plan useful follow-up, protect customer data, and keep human judgment in charge. AI promptscustomer retentioncustomer success AI QBR Prompts: Run Reviews Without the Theater 10 AI quarterly business review prompts to verify outcomes, expose risks, build a useful agenda, and plan next quarter without inventing customer value. AI promptsquarterly business reviewQBR --- ## AI Cutover Plan Prompts: Switch Systems Without Chaos URL: https://dontreplace.me/articles/ai-cutover-plan-prompts-switch-systems-without-chaos ai cutover plansystem migrationcutover promptsIT change managementrollback planning AI Cutover Plan Prompts: Switch Systems Without Chaos 10 copy-paste AI cutover plan prompts for run-of-show tables, dependency checks, rollback criteria, and comms. With safety rules for what not to paste. July 29, 2026 Updated July 29, 2026 Cutovers fail at handoffs. Not at the plan. The plan looks great. Sixteen pages, color-coded, owners in every cell. Then it's 2am, the database migration is running 40 minutes behind, the person who was supposed to validate the config file is in a different timezone and not responding, and someone just noticed the rollback window closes in 90 minutes. The plan didn't help. It just documented the chaos in advance. AI cutover plan prompts won't fix that on their own. But they will help you build something better than the usual copy-pasted runbook from the last migration, the one with the wrong owner names and the step that says "confirm with ops" with no follow-up from anyone. These 10 prompts are for the people actually running cutovers: IT leads, project managers, ops coordinators, change managers, and whoever got voluntold to own the weekend rollout. Paste them, adapt them, and then do the part AI can't do. Verify everything. What AI is actually good for in a cutover plan AI is a fast sequencer and a decent drafter. Give it a blob of cutover notes and it will produce a structured run-of-show in minutes. Give it a list of systems and it will suggest dependency checks you might have missed. Give it a draft comms email and it will make it cleaner and less panicked-sounding. That's the value. Structure. Sequencing. Checklists. Communication drafts. What it's not good for: knowing whether your test evidence is real, whether your rollback window is actually achievable, or whether the person listed as approver has actually approved anything. AI will fill those gaps with confident-sounding guesses if you let it. That's the trap. Understanding exactly where AI helps and where it invents is worth five minutes before you use any of these prompts. Then come back. AI produces structure. You verify substance. That split is everything in a cutover. The reusable cutover prompt formula Every prompt in this list follows the same structure. You can adapt it to any cutover. The formula: "You are a cutover planning assistant. I'm [your role] preparing a cutover for [system/project name]. The cutover window is [date/time/duration]. Affected users are [scope]. Here are my current notes: [paste sanitized details]. Please [specific output: run-of-show / dependency list / comms draft / etc.]. Flag anything that looks incomplete or unclear." The last sentence matters. Asking AI to flag gaps is more useful than asking it to fill them. Gaps filled by AI look identical to gaps filled by real information. Gaps flagged by AI give you somewhere to start. Before you paste anything: no customer PII, no credentials, no access tokens, no production logs, no security vulnerabilities, no private tickets, no financial records, no regulated data, no HR details, no unreleased strategy, no private client conversations. Sanitize first. Always. 10 AI cutover plan prompts you can use today Prompt 1: Turn messy cutover notes into a run-of-show You are a cutover planning assistant. I'm preparing a run-of-show for a system cutover. Here are my working notes: [paste sanitized notes]. Please structure these into a time-sequenced run-of-show table with columns for: time, step, owner, dependency, completion signal. Flag any steps that are missing owners or completion signals. Do not invent information. What to check after: Every owner cell should contain a real name, not a team name. "IT team" is not an owner. If a step says "pending," it's pending in real life too. The run-of-show is a mirror, not a magic fix. Prompt 2: Map owners and approvers You are a cutover planning assistant. Here is a list of cutover tasks and the people provisionally assigned to them: [paste sanitized list]. Please identify: (1) tasks with no assigned owner, (2) tasks where the same person appears more than three times as a single point of failure, (3) approval steps with no named approver. Suggest where backup owners might be needed. Do not invent names or roles. What to check after: AI can spot structural gaps. It cannot tell you whether those people have been notified, confirmed availability, or have the access they need. That's your job. Prompt 3: Sequence pre-cutover tasks You are a cutover planning assistant. Here is an unordered list of pre-cutover tasks that need to happen before go-live: [paste sanitized task list]. Please sequence these logically, identify which tasks are blockers for others, and flag any circular dependencies or tasks that appear to require completion before they can be started. Show your reasoning. What to check after: AI is good at logical sequencing but doesn't know your actual system. A task that looks independent might share a database schema, a shared service account, or a maintenance window with something else. Check the logic against real infrastructure. Prompt 4: Check dependencies and access You are a cutover planning assistant. Here are the systems involved in this cutover: [list sanitized system names and relationships]. Please generate a dependency check list covering: upstream systems, downstream consumers, shared services, authentication dependencies, batch job timing, and access requirements for each named system. Flag anything that looks like it could block cutover progress if not verified ahead of time. What to check after: This gives you a checklist, not a verified status. Every item on that list still needs a human to confirm actual current state. "Expected to be ready" is not ready. Prompt 5: Validate test evidence You are a cutover planning assistant. Here is our test summary for the pre-cutover validation phase: [paste sanitized summary with test names, outcomes, and dates]. Please identify: (1) tests marked as passed with no date, (2) tests completed more than 30 days ago that may need re-running, (3) tests that describe partial pass or workarounds, (4) areas with no test coverage mentioned. Flag these clearly. Do not interpret ambiguous outcomes as passed. What to check after: AI will find structural gaps in your documentation. It cannot confirm whether a test actually ran or whether the "passed" status is real. Treat flagged items as mandatory follow-ups, not optional reviews. Dmitry Kargaev makes this point directly in Don't Replace Me: tidy documentation is not proof of a ready system. Prompt 6: Draft customer and internal communications You are a cutover planning assistant. I need to draft [type: customer-facing / internal] communication about a planned system cutover. Key details: [system name, expected downtime window, affected functionality, support contact]. Tone should be [professional/direct/reassuring]. Draft a communication that covers: what is happening, when, what users should expect, what they should do if they have issues, and who to contact. Keep it under 200 words. What to check after: Do not send anything AI drafted without human review. Check that the downtime window is accurate, the support contact is real and staffed, and the tone matches your organization's standards. Get approval before it goes anywhere. For the internal version, run it past whoever owns change comms. Also worth having: AI stakeholder update prompts if you need to brief leadership separately. Prompt 7: Define rollback criteria You are a cutover planning assistant. I need to define clear rollback criteria for our cutover. The cutover involves: [describe system change at a high level, sanitized]. Please generate a rollback criteria framework covering: (1) technical conditions that trigger rollback, (2) time thresholds after which rollback becomes unavailable, (3) who has authority to call rollback, (4) steps required to execute rollback, (5) customer communication required if rollback is invoked. Flag any gaps in this framework. What to check after: AI can generate a logical framework. The rollback steps it produces are illustrative, not tested. Your actual rollback plan needs to have been rehearsed, not just documented. For a full rollback template, AI rollback plan prompts covers this in detail. Prompt 8: Prepare support coverage You are a cutover planning assistant. I'm preparing support coverage for a cutover window of [duration] starting [day/time]. Affected users are [description]. Please help me build a support coverage checklist covering: (1) roles that need to be available, (2) escalation paths for technical issues, (3) escalation paths for customer-facing issues, (4) communication channel setup (status page, Slack channel, bridge line), (5) handoff process if cutover extends beyond the initial window. Flag any typical gaps in cutover support coverage. What to check after: Coverage looks different on a spreadsheet than at 3am. Check that everyone listed is actually aware they're on coverage, has the right access to respond, and knows the escalation path. "On call" means nothing if they're not set up to act. Prompt 9: Set monitoring and escalation triggers You are a cutover planning assistant. After our cutover, we need to monitor for issues. The system being cut over is: [describe at high level, sanitized]. Please suggest: (1) key metrics to monitor in the first 2 hours post-cutover, (2) threshold values that should trigger escalation, (3) who should be alerted at each threshold level, (4) monitoring dashboard or tool categories to have open. Flag any monitoring gaps that are common in cutovers of this type. What to check after: AI will suggest sensible metric categories. It doesn't know your actual dashboard URLs, your real alerting thresholds, or whether your monitoring is set up to catch what matters. For post-cutover monitoring templates, AI post-launch monitoring prompts has the follow-up covered. Prompt 10: Convert cutover gaps into action items You are a cutover planning assistant. I've completed a review of our cutover plan and identified the following gaps and open questions: [paste sanitized gap list]. Please convert these into a structured action item list with: owner (placeholder if unknown), due date (relative: e.g., "3 days before cutover"), priority (high/medium/low), and whether the item is a blocker. Sort by priority and flag any items that have no obvious owner. What to check after: Placeholder owners need to become real owners before this list means anything. A gap with no owner is still a gap. Assign, confirm, and get written acknowledgment where the stakes are high. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → The part AI can't do for you Every prompt in this list produces a draft, a structure, or a checklist. None of them produce a cutover that's actually ready to run. The human part is: verifying test evidence is real. Confirming approvals are confirmed, not assumed. Checking that rollback has been rehearsed. Making sure the person listed as owner for step 14 knows they're the owner for step 14. Getting explicit sign-off from whoever has authority to call the cutover. If you want a complete framework for the go/no-go decision that happens the day before, AI go/no-go decision prompts will save you from launching on vibes. And if you're still building out the migration plan before you get to cutover, AI migration plan prompts and AI operational readiness prompts are worth working through first. The broader point is in Rule #7 from Don't Replace Me: the human value in a cutover isn't generating the runbook. It's knowing which details actually matter. The timing that can't slip. The customer impact that wasn't in the brief. The dependency that everyone assumed someone else had checked. AI will give you a cleaner document. You still have to know whether it's right. Frequently asked questions Can AI write a cutover plan for me? AI can generate a structured cutover run-of-show, dependency checklist, and communications draft quickly. But it can't verify that owners are real, that tests actually passed, or that rollback steps are achievable. Use AI for structure and drafting, then verify every substantive detail before the window opens. What should I never paste into an AI tool when planning a cutover? Never paste customer PII, credentials, access tokens, production logs, security vulnerabilities, private tickets, financial records, regulated data, HR details, or anything under legal review. Sanitize your inputs before using any AI assistant, even internal approved tools. How do I know if my cutover plan is actually ready? A plan is ready when owners are named and confirmed, test evidence has dates and real outcomes, rollback steps have been rehearsed, monitoring dashboards are configured, support coverage is booked and briefed, and someone with authority has formally signed off. A clean document isn't the same thing. See AI operational readiness prompts for a structured review. What's the difference between a cutover plan and a migration plan? A migration plan covers the full scope of moving from one system to another, including data transfer, testing, training, and transition. A cutover plan is the specific execution document for the go-live window itself: the run-of-show, the timing, the rollback criteria, and the decision points. Both need AI assistance applied carefully with human verification. How do AI cutover plan prompts help with rollback decisions? The prompts help you define rollback criteria, time thresholds, and authority structures in advance, before you're tired and under pressure at 2am. But the rollback decision itself is always a human call. AI can structure the criteria. A named person with authority has to make the call. How early should I use these prompts before a cutover? Use the dependency, owner, and test-evidence prompts at least two weeks out. Use the run-of-show and communications prompts one week out. Use the monitoring and support coverage prompts two to three days before the window. Waiting until the day before to find gaps is a strategy, just not a good one. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Migration Plan Prompts: 10 Templates That Work 10 copy-paste AI migration plan prompts for system mapping, dependency checks, rollback criteria, cutover plans, and stakeholder comms. With safety reminders. AI promptsmigration planningsystem migration AI Rollback Plan Prompts: 10 Templates That Work 10 copy-paste AI rollback plan prompts for ChatGPT and Claude. Define scope, map owners, write comms, and verify recovery without starting from scratch. AI promptsrollback planninglaunch recovery AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation --- ## AI Data Analysis Prompts: 10 Templates That Work URL: https://dontreplace.me/articles/ai-data-analysis-prompts-turn-messy-data-into-decisions ai data analysis promptschatgpt for dataspreadsheet formulassql promptsdata cleaning AI Data Analysis Prompts: 10 Templates That Work 10 copy-paste AI data analysis prompts for ChatGPT and Claude. Write formulas, clean data, build SQL, and turn findings into decisions. June 04, 2026 Updated June 04, 2026 Your boss drops a spreadsheet in your lap. Forty columns, three years of data, inconsistent date formats, a column called "FINAL_v3_USE_THIS," and a tab named "DO NOT DELETE." You have a meeting in two hours. AI data analysis prompts are your fastest way out of this situation, and most people are using them wrong. Wrong means: pasting raw data into ChatGPT and hoping it figures out what you actually need. It won't. Not because AI is dumb, but because AI is fast, confident, and has no idea what your boss actually cares about. That gap between confident-sounding output and correct output is where decisions go to die. This article gives you 10 copy-paste prompts you can use today, a reusable formula for every analysis task, and the guardrails to stop you from trusting a hallucinated benchmark in front of your CFO. Before you touch a prompt: the one rule that matters AI analysis works exactly like hiring a smart intern who reads everything instantly, remembers nothing about your business, and will confidently answer a question you didn't ask. Your job is to give that intern three things before they touch the data: The dataset structure (columns, row count, what each row represents) The business question in plain English The output format you actually need Skip any of those and you'll get something that looks like an answer but isn't. This isn't a flaw to work around. It's the entire model. Garbage in, garbage out. The better your setup, the better your output. The reusable formula: Act as [role]. I have a dataset with [describe columns, row count, grain]. My business question is: [specific question]. Please [specific task]. Format the output as [table / bullet list / SQL / Python / plain English]. Flag any assumptions you make. That last sentence, "flag any assumptions," is the one most people leave out. It's the most important one. A word on data safety before anything else. Do not paste confidential customer records, employee salaries, health data, unreleased financial results, or regulated information into any AI tool unless your company has an explicit policy that permits it. Anonymize first. Use fake IDs. Paste structure and sample rows, not the whole damn file. This isn't paranoia. It's basic professional hygiene, and it's on you, not the AI. 10 AI data analysis prompts you can use right now Prompt 1: Understand what you're even looking at You've got the file. You don't know what it contains. Start here. Act as a data analyst. I have a dataset with the following columns: [paste column names and a few sample rows]. Each row represents [one customer / one transaction / one month of sales, etc.]. Please: (1) describe what this dataset likely tracks, (2) identify any columns that look ambiguous or misnamed, (3) flag any obvious data quality issues from the sample, (4) suggest three questions this data could answer. Flag any assumptions you make. This prompt stops you from jumping straight to analysis before you understand the shape of what you're working with. It takes two minutes and saves you from building a chart on top of a column that means something completely different than you thought. Prompt 2: Clean messy columns Inconsistent capitalization. "N/A" and "null" and blank cells all meaning the same thing. Dates in three formats. Every analyst's Tuesday. Act as a data cleaning assistant. This column contains [paste column name and 20 sample values]. Values should represent [what this column is supposed to contain]. Please: (1) identify the distinct messy patterns, (2) suggest a cleaning rule for each pattern, (3) write an Excel formula or Python/pandas one-liner to apply each fix. List any values you're unsure about rather than guessing. The "list what you're unsure about" instruction matters. Without it, AI will make a call on every edge case and not tell you it did. You want it to surface the ambiguous ones so a human (you) makes that judgment call. Prompt 3: Build spreadsheet formulas This is where AI genuinely earns its keep. Formula writing is tedious, error-prone, and the kind of task that eats 45 minutes when it shouldn't. Our AI spreadsheet prompts guide goes deeper on this, but here's the core data analysis version: I have an Excel / Google Sheets file. Column A is [description], Column B is [description], Column C is [description]. I want to calculate [exactly what you want to calculate] and put the result in Column D. The formula should [handle blanks / ignore zero values / look up from another sheet called X]. Write the formula and explain each part. Flag if there's more than one way to do it. Paste the formula into a test row first. Check it against three rows you can verify manually. Always verify row counts and filters before you trust a formula on a full dataset. Prompt 4: Write SQL for a specific question Act as a SQL analyst. I have a table called [table_name] with these columns: [list columns with data types]. I want to find [specific business question, e.g., "the top 10 customers by total revenue in Q1 2024, excluding refunded orders"]. Write a SQL query that answers this. Use standard SQL. Add comments explaining each clause. Flag any joins or filters I should double-check before running this in production. Always run AI-generated SQL on a sample or dev environment first. Check the row count. Check that your joins aren't multiplying rows. Check that your date filters are doing what you think they're doing. AI will write syntactically correct SQL that answers the wrong question with more confidence than you'd like. Prompt 5: Write Python / pandas steps Act as a data analyst using Python and pandas. I have a DataFrame called df with these columns: [list columns and dtypes if you know them]. I want to: [numbered list of transformation steps]. Write the pandas code for each step. Use comments. After the code, list any assumptions you made about the data types or nulls, and suggest one sanity check I should run after each transformation. The sanity check request is the move here. Something like "print the row count before and after the merge" or "check for duplicates on the key column" catches the silent errors that wreck downstream analysis. Prompt 6: Check summary statistics You've run your analysis. Now you want to know if the numbers smell right before you put them in a slide. I ran a summary of [dataset description]. Here are the results: [paste summary stats: mean, median, min, max, row count, null count per column]. Does anything look suspicious? What would a reasonable range be for each metric given this is [context: e.g., e-commerce orders for a mid-size retailer]? List anything worth investigating before I present these numbers. Notice what this prompt is not doing: it's not asking AI to invent benchmarks. It's asking for a smell test on your specific numbers. If the AI starts citing industry benchmarks you can't verify, don't use them. Ask it to flag the concern without making up a comparison point. Prompt 7: Find suspicious outliers Act as a data quality reviewer. Here are rows from my dataset that appear to be outliers: [paste 10-20 suspect rows]. The column in question is [column name] and it should represent [what it represents]. For each row: (1) explain why it might be an outlier, (2) suggest whether it looks like a data error, a real edge case, or something that needs manual review. Do not delete or replace values. Just flag and explain. The human decides what to do with outliers. The AI helps you see them faster and think through the options. That's the right division of labor. Prompt 8: Explain a chart in plain English You made a chart. Your boss won't read it unless there's a sentence underneath it that explains exactly what to see. This is where AI is genuinely useful and where a lot of people are already saving time. Dee covers this kind of task-level thinking in Don't Replace Me, where the point is simple: use AI for the parts of work you hate or the parts that slow you down. Writing chart captions at 6pm is exactly that. Here is a description of a chart I made: [describe chart type, axes, what data it shows, the key trend or pattern you see]. Write three versions of a plain-English caption that explains what the chart shows and why it matters. Audience: [your specific audience, e.g., non-technical marketing team]. Keep each caption under 40 words. Do not editorialize or attribute causation unless the data clearly supports it. Prompt 9: Turn findings into a decision memo This is where the work becomes useful. Raw findings aren't decisions. Someone has to translate them. AI can draft that translation fast, and you clean it up. Act as a business analyst writing for senior leadership. Here are my key findings from a data analysis: [bullet point your 3-5 main findings, including the numbers]. The original business question was: [restate the question]. Please write a one-page decision memo with: (1) context, (2) key findings, (3) what the data suggests, (4) one recommended next step with a rationale. Flag anywhere the data doesn't fully support the recommendation so I can decide how to handle it. That last instruction keeps AI from overstating your conclusions. The memo draft is a starting point, not a finished product. Read every sentence and ask: does my data actually support this? If not, cut it or qualify it. For research and synthesis tasks like this, the approach in our desk research prompts guide applies directly. Prompt 10: Check whether your analysis supports the conclusion This one's uncomfortable, which is why it's the most useful. I'm about to present the following conclusion: [state your conclusion]. Here is the analysis I used to reach it: [describe or paste your key numbers, method, and assumptions]. Please play devil's advocate. What are the three most credible ways someone could challenge this conclusion? What data or information would I need to make the conclusion airtight? What assumptions am I making that might not hold? You're not outsourcing your judgment here. You're stress-testing it before someone else does it for you in a meeting. The prompts guide for work has a similar adversarial prompting pattern for other contexts. What AI data analysis prompts can't do for you AI can write the formula. It can't tell you whether the metric you chose is the right one to care about. That distinction matters more than any prompt. What AI can and can't do gets into this properly, but the short version for analysis work is: AI is excellent at transformation, formatting, formula writing, and explaining patterns in data you provide. It's bad at knowing whether your business question is the right one, whether a correlation is causal, whether a benchmark it cites is real, or whether the tradeoff your analysis implies is actually acceptable to your organization. Those are judgment calls. They belong to you. That's not a limitation to resent. It's the reason you still have a job. The quality checklist before you trust anything AI generates: Row counts match before and after joins or filters Date ranges are correct and inclusive/exclusive as intended Denominators are what you think they are Units are consistent (are you mixing monthly and annual figures?) Null handling is explicit, not assumed Any benchmarks or external numbers come from sources you can verify Formulas have been tested on rows you can manually confirm One more time on data privacy: no customer PII, no employee records, no health data, no unreleased financials, no regulated data. Anonymize. Paste structure and sample. Keep your job. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Frequently asked questions Can I use ChatGPT or Claude to analyze a spreadsheet? Yes, and it works well if you give it the column names, a description of what each row represents, and a specific question. Don't paste the whole file. Paste the structure and a sample of rows. Both ChatGPT and Claude can write formulas, spot data quality issues, and help you build a summary. Verify everything against your source data before you use it in a report. What's the best way to write an AI prompt for data analysis? Start with a role, describe your dataset (columns, row count, what each row represents), state your business question in plain English, specify the output format, and ask it to flag any assumptions it makes. That five-part structure covers most of what AI needs to give you something useful rather than something that looks useful. Is it safe to paste data into ChatGPT for analysis? Only if your company policy explicitly permits it and the data isn't confidential, regulated, or personally identifiable. For most work data, anonymize first: replace names with IDs, remove email addresses, and use fake values for anything sensitive. When in doubt, paste only the column structure and a few fake sample rows, then describe the real data in plain text. Can AI write SQL queries for me? Yes, and it's reasonably good at it. Give it the table name, column names with data types, and the exact question you're trying to answer. Always test AI-generated SQL on a sample or non-production environment first. Check row counts, verify joins aren't multiplying records, and confirm date filters are doing what you think. Will AI make up statistics or benchmarks in my analysis? It can, and it will do so confidently. Never ask AI to supply benchmarks, industry averages, or external comparisons unless you can verify them independently. Ask it to flag where it's making assumptions rather than stating facts. Treat any number that didn't come from your own dataset as unverified until you find a primary source. What tasks is AI actually useful for in data analysis? Writing formulas, generating SQL, cleaning messy columns, explaining what a chart shows, drafting summary memos, and checking for data quality issues are all genuinely useful. Deciding which metric matters, interpreting whether a pattern is meaningful for your business, and making the final recommendation are still on you. AI speeds up the mechanical parts. The judgment parts are your job. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance --- ## AI Deal Desk Prompts: Review Complex Deals URL: https://dontreplace.me/articles/ai-deal-desk-prompts-review-complex-deals AI promptsdeal deskrevenue operationssales operationsdeal review AI Deal Desk Prompts: Review Complex Deals 10 AI deal desk prompts for checking discounts, terms, risks, approvals, and complex sales deals without inventing policy, evidence, or authority. September 06, 2026 Updated September 06, 2026 Complex deals rarely fail because nobody could produce a summary. They fail because the summary quietly mixes facts, seller optimism, old pricing rules, unsigned promises, and one heroic spreadsheet nobody wants to question. AI deal desk prompts can help organize verified opportunity data, identify missing inputs, compare commercial options, and prepare approval questions. They cannot approve a discount, interpret a contract as legal advice, confirm a buyer’s authority, promise a product capability, or decide how much risk the company should accept. AI can prepare a deal for judgment. It cannot supply the judgment, evidence, or authority. The ten templates below help deal desk teams, revenue operations leaders, sales managers, finance and legal partners, and founders review complex deals without promoting autocomplete to chief commercial officer. What a deal desk actually controls A deal desk is a decision system. It brings the right evidence and owners together before the company commits price, terms, scope, capacity, security obligations, or risk. The useful version answers: What has the buyer actually requested? Which details are verified, and where? Is the proposed price allowed under current policy? Which terms are standard, negotiable, or prohibited? What delivery, security, legal, finance, or product dependencies exist? Who has authority to approve each exception? What remains unknown before anyone signs? The bad version is an administrative obstacle course performed at quarter-end while everyone types “urgent” in increasingly creative ways. Keep these categories separate in every AI-assisted review: Verified fact: supported by a current, dated source. Customer statement: what the buyer said, not proof that it is true or approved. Seller interpretation: a reading of the situation, not the situation itself. Requested exception: a deviation awaiting an authorized decision. Scenario: a possible outcome used for comparison. Unknown: information still requiring discovery or specialist review. Decision: an explicit choice made by a named human owner. Restricted data: information that must not enter the selected AI tool. A polished deal memo is still dangerous if it turns “customer asked” into “company agreed.” Language models are unusually talented at sanding off exactly that distinction. For adjacent work, use AI sales pipeline review prompts to find stalled opportunities, AI sales forecast prompts to build a grounded number, and AI account planning prompts to prepare the wider relationship strategy. The evidence-first deal desk prompt formula Add this instruction to every prompt below: “Act as a deal desk preparation assistant. Use only the sanitized, approved sources I provide for [deal, product, region, customer segment, review stage, and decision]. Cite a source label and date for every price, discount, term, capability, deadline, commitment, dependency, and policy statement. Separate verified facts, customer statements, seller interpretations, requested exceptions, scenarios, decisions, and unknowns. Mark unsupported or stale items [VERIFY]. Do not invent approvals, policy, precedent, legal interpretations, buyer motives, product capabilities, implementation capacity, or probability. Route each open issue to the named human owner.” This makes the output less magical and considerably less likely to become evidence in an unpleasant meeting. Never paste customer names, personal contact details, contracts, credentials, payment information, private CRM notes, confidential margins, unannounced pricing, security findings, regulated data, or legally sensitive information into an unapproved AI tool. Use approved systems, minimum necessary data, de-identification, access controls, retention limits, and human legal, finance, security, privacy, product, delivery, and commercial review. What to collect before prompting Build a controlled source pack. Do not paste an entire CRM export into a chatbot and call the resulting privacy incident “revenue intelligence.” InputWhy it mattersHuman owner Current opportunity recordGrounds stage, amount, dates, and scopeAccount owner Customer request in writingSeparates buyer language from seller memoryAccount owner Approved price bookEstablishes valid products and list pricesFinance or commercial owner Discount and approval policyDefines thresholds and authorityRevenue operations Proposed order form and termsShows the actual commitmentLegal and commercial owners Product capability sourcePrevents accidental roadmap promisesProduct owner Delivery and implementation planTests whether the company can performDelivery owner Security and privacy requirementsRoutes specialist obligationsSecurity and privacy owners Margin or unit-economics modelShows the commercial trade-offFinance owner Decision deadline and reasonDistinguishes urgency from theaterDeal sponsor Source dates and versionsExposes stale evidenceDocument owners Restricted-data rulesControls what may enter the toolSecurity or privacy owner If two sources conflict, preserve the conflict. Do not ask the model to pick whichever number makes the opportunity look healthier. Named owners resolve it. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI deal desk prompts Replace brackets with sanitized, approved information. Every output is a draft until authorized humans verify and approve it. 1. Check deal intake completeness “Review this sanitized deal intake against the supplied intake requirements: [paste]. Create a table with required field, supplied value, source, source date, confidence, conflict, missing information, owner, and next action. Separate verified facts from customer statements and seller interpretations. Do not fill blanks by inference. Mark every unsupported field [VERIFY].” This is the sensible first use of AI: finding empty boxes before twelve people join a call to discover them together. Require sources for the buying entity, products, quantities, term length, start date, billing schedule, requested discount, implementation scope, security needs, legal changes, decision process, and deadline. “It’s in Salesforce” is not a source if the field has not been updated since the previous account owner left. The human deal owner should resolve missing inputs with the buyer or relevant specialist. The model can identify the hole; it cannot pour truth into it. 2. Build a verified deal summary “Using only these labeled sources: [paste], draft a one-page deal summary. Include customer request, verified business need, proposed products and quantities, list price, proposed price, term, payment schedule, delivery scope, requested exceptions, dependencies, risks, unknowns, decision deadline, and required approvers. Cite every material statement. Put seller interpretations in a separate section and mark unsupported content [VERIFY].” A useful summary lets an approver understand the decision without excavating chat threads. It should also make uncertainty visible rather than laundering it into confident prose. Do not describe a deal as strategic unless the organization defines that word and the evidence meets the definition. “Large logo” is not a complete business case. Neither is “the rep says they will expand later.” For customer-facing language after approval, use AI sales proposal prompts. The proposal must reflect the approved deal, not become a parallel negotiation invented by a text generator. 3. Review a discount request “Evaluate this requested discount against the supplied current price book, discount policy, margin guardrails, and approved precedent register: [paste]. Show list price, proposed price, effective discount, term, volume, payment timing, included services, margin impact supplied by finance, threshold triggered, required approvers, evidence offered, and missing evidence. Compare approve, revise, and decline scenarios without recommending a decision unless criteria are explicitly supplied.” Discounts are not isolated percentages. A smaller discount with expensive custom work, generous payment terms, and broad termination rights may be worse than the bigger number everyone is arguing about. Require finance to validate calculations and assumptions. Require commercial owners to decide whether the trade is worth making. The model should never invent a precedent because another deal “sounds similar.” Precedent needs a source, context, and authorized interpretation. Also avoid using AI to infer how much a buyer can afford from company size, location, demographics, or personal information. Price decisions belong to approved commercial policy and actual negotiation evidence, not algorithmic fortune-telling. 4. Map the approval path “Map this proposed deal against the supplied approval matrix: [paste]. For each price, discount, term, legal change, security requirement, data-processing obligation, implementation exception, product commitment, and payment condition, identify the rule triggered, evidence, required approver, sequence, deadline, and unresolved question. Do not treat silence, attendance, or a previous approval as approval for this deal.” This prompt turns a messy request into a routing plan. It does not grant authority. Approval matrices often contain thresholds, regional differences, and combinations. A discount may be acceptable until paired with nonstandard payment terms. A standard contract may stop being standard when a data-processing addendum changes liability. Ask owners to confirm the current matrix before trusting the output. Record decisions in a durable system. AI decision log prompts can help structure the record, but the approval evidence must come from the actual approver. 5. Compare commercial scenarios “Using these verified inputs and finance-approved calculations: [paste], compare [scenario A], [scenario B], and [scenario C]. For each, show price, discount, term, payment timing, included scope, excluded scope, margin input, delivery demand, customer concession, company concession, dependencies, risks, approval path, and unknowns. Keep calculations traceable. Do not assign probabilities or declare a winner unless explicit decision criteria are provided.” Scenario comparison helps teams negotiate bundles rather than obsess over one number. Perhaps the buyer receives a lower price in exchange for a longer term or faster payment. Perhaps custom work is removed. Perhaps the responsible answer is no. Do not let the model fabricate buyer acceptance. A scenario is an internal option until the buyer accepts it and authorized humans approve it. Clearly label assumptions and expiry dates because product availability, staffing, and pricing can change. 6. Surface contract, security, and delivery dependencies “Review these sanitized requirement summaries from legal, security, privacy, product, and delivery: [paste]. Create a dependency register with requirement, source, status, owner, due date, blocker, proposed response, approval needed, and effect on scope, timing, or price. Quote no legal conclusion beyond the supplied source. Route interpretation to the relevant specialist.” The phrase “legal is reviewing” hides a small universe. Which clause? Which entity? What decision is needed? Does security require architecture work? Does implementation have capacity on the promised date? AI can organize specialist inputs and show where one answer affects another. It cannot provide legal advice, conduct a security assessment, or commit delivery resources. Treat generated interpretations as drafts and have each specialist validate their row. For broad third-party evaluation, AI vendor evaluation prompts provide a useful mirror image: buyers need evidence too. 7. Separate facts from deal optimism “Classify every statement in this sanitized deal narrative: [paste] as verified fact, customer statement, seller interpretation, requested exception, scenario, decision, or unknown. Add source, date, owner, and verification action. Flag certainty words such as ‘will,’ ‘committed,’ ‘approved,’ ‘easy,’ ‘standard,’ and ‘guaranteed’ when the evidence does not support them. Preserve useful hypotheses but do not promote them to facts.” This is the anti-vibes prompt. A champion saying “procurement should be fine” is a customer statement. It is not procurement approval. A salesperson believing expansion is likely is a hypothesis. It is not contracted value. A product manager saying something may fit the roadmap is not a delivery commitment. Use this classification before forecast calls too. The goal is not to punish optimism; it is to keep optimism from impersonating evidence. 8. Prepare an exception memo “Draft an exception memo from these verified sources: [paste]. Include the exact policy exception requested, business rationale supplied, customer concession, financial effect supplied by finance, operational effect, precedent references, risks, mitigations, expiry or one-time boundary, alternatives considered, owner, required approvers, and unresolved questions. Mark proposed language DRAFT. Do not invent precedent, mitigation, approval, or strategic value.” A good exception memo makes the trade visible. It does not bury a permanent operational headache beneath “logo value.” If an exception is approved, state whether it creates precedent. State who owns fulfillment. State when the approval expires. One-time custom work has a charming habit of becoming standard customer expectation unless somebody labels it aggressively. Run an AI risk assessment when the exception affects security, privacy, delivery, finance, customer commitments, or multiple teams. 9. Draft the approval meeting brief “Turn this verified deal record into a meeting brief for [approvers]. Include the decision required, deadline and source, five essential facts, requested exceptions, scenario comparison, material risks, specialist positions, unresolved questions, recommendation supplied by the accountable human, and explicit decisions to record. Add a parking-lot section for issues outside scope. Do not invent consensus or mark attendance as approval.” The brief should enable a decision, not narrate the opportunity from the first cold email. Send it early enough for specialists to inspect their sections. In the meeting, record decisions, conditions, dissent, owners, and expiry dates. If facts change afterward, reopen the relevant approval rather than assuming the old decision covers the new deal. A summary generated after the meeting must be checked against the actual record. Fast notes are useful. Fictional consensus is not. 10. Run the final human sign-off checklist “Review the final proposed deal against the supplied approved sources and approval matrix: [paste]. Produce a sign-off checklist covering entity, products, quantities, price, discount, term, billing, scope, implementation capacity, product claims, legal terms, security, privacy, data processing, support, renewal, cancellation, customer commitments, internal obligations, approvals, document consistency, and restricted-data handling. Cite evidence for every completed item. Mark an item complete only when explicit proof is supplied.” This is the gate before signature, not a decorative table appended to an email. Compare the CRM record, quote, order form, contract, implementation scope, and approval record. A correct approval attached to the wrong document version is not a correct deal. Ask the named owners to sign off in their domains. Use an AI QA checklist for the document package, then let authorized humans make the final decision. The model gets no vote and, conveniently, accepts no liability. A practical deal desk operating rhythm AI-assisted deal review works best inside a visible sequence: Intake: collect the minimum required, current, sanitized sources. Classify: separate facts, statements, interpretations, exceptions, scenarios, decisions, and unknowns. Validate: ask source owners to resolve missing or contradictory information. Compare: examine commercial options using finance-approved calculations. Route: send legal, security, privacy, product, delivery, and finance issues to specialists. Decide: record explicit human approvals, conditions, and rejections. Reconcile: confirm every final document matches the approved position. Maintain: preserve the decision record and update policy from real lessons. Start with one bounded workflow, such as intake completeness or approval routing. Do not begin by giving a model access to every contract, CRM note, and pricing workbook in the company. That is not a pilot. It is a resignation letter for your security team. If your team needs the basics first, read the no-BS guide to using AI at work and what AI can and cannot do. The useful pattern is boring: controlled evidence, explicit labels, narrow outputs, and accountable humans. Frequently asked questions What are AI deal desk prompts? They are structured instructions that help an AI system organize verified deal information, identify gaps, compare scenarios, and draft review artifacts. Good prompts define approved sources, classifications, prohibited assumptions, output format, and human owners. They do not ask the model to approve a deal from vibes. Can ChatGPT approve a sales discount? No. It can compare a proposed discount with a supplied policy and show which threshold appears to be triggered. Finance, sales leadership, or another authorized owner must validate the calculations, interpret the policy, assess the trade, and approve or reject the request. Can AI review a customer contract? AI may help organize clauses or questions inside an approved legal workflow, but it should not be treated as legal advice or an authorized interpretation. Contracts can contain confidential and legally sensitive information. Qualified counsel must review the actual language and own legal decisions. What information should never go into an AI deal review? Do not put personal data, credentials, payment details, private CRM notes, contracts, confidential margins, security findings, regulated information, or legally sensitive material into an unapproved tool. Follow company data-classification, access, retention, privacy, security, and legal rules. Use minimum necessary and de-identified inputs where possible. How do we stop AI from inventing approvals or policy? Supply the current policy and approval matrix, require citations for every rule, label missing evidence [VERIFY], and forbid the model from marking approval complete without explicit proof. Then have the named human owners inspect the output. Prompting reduces risk; it does not guarantee truth. Is an AI-generated deal score reliable? Not by default. Scores can hide stale CRM fields, biased history, inconsistent definitions, and missing context. Prefer transparent evidence, decision criteria, and human review. If a score is used, document its inputs, limitations, validation, and prohibited uses. Never use it as a substitute for accountable judgment. Can AI replace a deal desk team? No. AI can reduce administrative work around intake, classification, comparison, and documentation. It cannot negotiate trust, interpret novel risk responsibly, balance company priorities, authorize exceptions, or own legal and commercial outcomes. What is the safest first deal desk use case? Start with intake completeness using sanitized data and a current checklist. It has clear sources, limited scope, and an obvious human owner. Measure whether reviews arrive more complete and whether corrections decrease before expanding into scenario comparison or exception preparation. Speed up preparation, not authority Deal desk work contains exactly the kind of structure AI handles well: repeated fields, classification, comparison, summaries, checklists, and missing-information detection. It also contains exactly the kind of judgment AI should not own: trust, negotiation, policy interpretation, legal risk, security obligations, financial trade-offs, delivery capacity, and authorization. Keep the source next to the claim. Keep the customer request separate from company acceptance. Keep scenarios separate from forecasts. Keep exceptions visible. Keep approvals explicit and human. The machine can make a complex deal easier to inspect. It cannot make the deal true, safe, profitable, or approved. For a broader field guide to staying useful while AI makes knowledge work faster and stranger, read Don’t Replace Me by Dmitry Kargaev. The durable advantage is not typing the cleverest prompt. It is knowing which decisions must remain attached to evidence, taste, and a human name. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Customer Renewal Prompts: Prepare With Evidence 10 AI customer renewal prompts to verify outcomes, expose account risks, prepare options, and protect trust without inventing value or customer sentiment. AI promptscustomer renewalcustomer success AI Sales Enablement Prompts: Build Useful Materials 10 AI sales enablement prompts for grounded messaging, training, coaching, role-play, objection practice, and seller resources without invented claims. AI promptssales enablementsales training AI Sales Forecast Prompts: Skip the Fake Certainty 10 AI sales forecast prompts to test evidence, build scenarios, challenge assumptions, protect buyer data, and produce an accountable forecast. AI promptssales forecastingrevenue operations --- ## AI Decision Log Prompts: Stop Re-litigating Work URL: https://dontreplace.me/articles/ai-decision-log-prompts-stop-re-litigating-work ai decision logdecision documentationai promptsdecision makingworkplace productivity AI Decision Log Prompts: Stop Re-litigating Work 10 copy-paste AI decision log prompts to turn messy notes, Slack threads, and meeting chaos into a clear record. Templates for every decision type. July 01, 2026 Updated July 01, 2026 Every team has a decision that keeps coming back from the dead. You made the call six months ago. You documented nothing. Now three people remember three different versions of what was decided, the Slack thread is 200 messages deep, and you're in a meeting re-litigating a choice that should have been buried and forgotten. AI decision log prompts won't fix bad judgment. But they will stop you from having that meeting again. This is a template article. The theory is short. The prompts are the point. Why decisions rot without documentation The problem isn't that teams make bad decisions. It's that they make fine decisions and then let them dissolve into institutional memory, which is another way of saying "whatever the most confident person in the room remembers." Six months later, the context is gone. The person who had the full picture left the company. The tradeoffs that made the original choice sensible look arbitrary in hindsight. So everyone fights about it again. A decision log fixes this. Not because it's bureaucratic busywork, but because it creates a single source of truth: here's what we decided, here's what we knew, here's who owns it, here's when to revisit. AI helps you build that record fast. What it doesn't do is supply the judgment you skipped the first time. What AI can and can't do is worth understanding before you treat a polished memo as a substitute for a real conversation. The reusable AI decision log prompt formula Before the templates, here's the structure every good decision log prompt follows. Use it to build your own when the templates below don't quite fit. Role: Act as a decision documentation assistant. Input: Here's the raw material: [paste your notes, thread, or bullet points]. Task: [What you want: a summary, a tradeoff list, a stakeholder update, etc.] Format: [One-pager, bullet list, table, prose paragraph, etc.] Constraints: Flag any gaps in the information I've given you. Don't invent facts, consensus, or approvals that aren't in the source material. That last line matters. AI will cheerfully fill gaps with plausible-sounding fiction. Tell it not to. One rule before you start: Don't paste customer PII, employee records, credentials, confidential strategy, unreleased product details, financial forecasts, legal disputes, security incidents, contracts, board materials, or sensitive client information into unapproved AI tools. Strip it first, or use your organisation's approved environment. AI decision log prompts: the 10 templates Prompt 1: Turn messy notes into a decision log You've got a mix of bullet points, half-sentences, and things you typed during a meeting while also pretending to pay attention. Start here. Act as a decision documentation assistant. Here are my raw notes from a decision made on [DATE] about [TOPIC]: [PASTE NOTES] Turn these into a structured decision log with the following sections: - Decision made (one sentence) - Context and background (2-3 sentences) - Options considered - Tradeoffs for each option - Final decision and rationale - Owner - Review date Flag any sections where the notes don't give you enough to complete the field. Do not fill gaps with assumptions. Prompt 2: Extract options and tradeoffs When all you have is the conclusion and you need to reconstruct the reasoning for anyone who wasn't in the room. Act as a decision documentation assistant. The decision we made was: [DECISION]. Based on the following context: [PASTE CONTEXT] List the main options that were likely on the table. For each option, write out the tradeoffs in a two-column table: Pros | Cons. Then note which option was chosen and why based only on what I've given you. Flag anything you're inferring rather than deriving from the source material. Prompt 3: Write a one-page decision brief For decisions that need sign-off from someone who wasn't in the weeds. This is the "here's what you need to know without reading 40 Slack messages" version. Act as a decision documentation assistant. Here's the background on a decision we're documenting: [PASTE NOTES OR SUMMARY] Write a one-page decision brief in plain language. Include: the problem we were solving, the options we considered, the decision we made, the key tradeoffs, the owner, and the next review date. Keep it under 400 words. No jargon. Write it for a senior stakeholder who has context on the business but not on this specific issue. Prompt 4: Document why an option was rejected The part everyone skips. Three months later someone will suggest the rejected option again without knowing why it was ruled out. Write it down now. Act as a decision documentation assistant. We decided NOT to go with [OPTION]. Here's the context on why: [PASTE YOUR NOTES] Write a short "rejection record" for this option: what the option was, what problem it would have solved, why we ruled it out, and under what conditions it might be worth reconsidering. Keep it factual. Do not editorialize or add reasons that aren't in my notes. Prompt 5: Turn a Slack thread into an action-ready decision record The Slack thread is not a decision. It's a negotiation, a fight, and three tangents about the naming convention. You need to pull the actual decision out of it. Act as a decision documentation assistant. Here's a lightly edited Slack thread from [DATE] about [TOPIC]: [PASTE THREAD - remove names, PII, and any confidential information first] Extract the following: 1. The decision that was reached (or confirm if no clear decision was made) 2. Who owns the action 3. What the next steps are and by when 4. Any unresolved questions or risks flagged in the thread 5. Any dissenting views that should be noted Flag clearly if the thread doesn't include a firm decision. Don't invent one. This is a natural companion to the AI meeting notes prompts if you're also capturing what happened in the call that preceded the thread. Prompt 6: Create an executive summary The decision is made. Now the people above you need the version that fits in a two-minute read. Act as a decision documentation assistant. Here's the full decision log for [DECISION TOPIC]: [PASTE DECISION LOG] Write a three-paragraph executive summary. Paragraph one: the problem and why it needed a decision now. Paragraph two: the options and why we chose what we chose. Paragraph three: what happens next and who's responsible. Write it for an executive audience. No bullet points. Plain prose. Under 200 words. Prompt 7: Identify missing evidence You think the decision is documented. Is it actually complete? Run this check before you call it done. Act as a decision documentation assistant. Here's a decision log I've drafted: [PASTE DRAFT] Review it and flag: any factual claims that aren't sourced, any assumptions that are treated as facts, any gaps in the options analysis, any missing owner or date, and any risks that are mentioned but not addressed. Give me a checklist of what's missing or weak. Don't suggest made-up fixes. Just tell me where the holes are. This is where Rule #13 from Don't Replace Me bites people: garbage in, garbage out. A polished document with gaps in the reasoning is still a gap. AI is very good at making those gaps look like prose. Prompt 8: Write a stakeholder update The decision affects people who weren't part of it. They need to know what happened and why, without getting the entire war room version. Act as a decision documentation assistant. We made the following decision: [DECISION] The main audience for this update is: [WHO THEY ARE AND WHAT THEY CARE ABOUT] Background: [PASTE RELEVANT CONTEXT] Write a short stakeholder update that covers: what we decided, what it means for them specifically, what (if anything) they need to do, and who to contact with questions. Keep it under 150 words. Plain language. No jargon. For decisions with broader organisational impact, pair this with the AI status report prompts to keep downstream teams updated. Prompt 9: Set review triggers Decisions aren't meant to be permanent. Most are made with incomplete information. Build in the mechanism to revisit them before reality makes the decision irrelevant. Act as a decision documentation assistant. Here's the context for a decision we've just made: Decision: [DECISION] Key assumptions we're relying on: [LIST ASSUMPTIONS] What could change: [LIST KNOWN UNCERTAINTIES] Write a review triggers section for the decision log. Include: a recommended review date, the conditions that should trigger an earlier review (if assumptions change or new data arrives), and the questions we should be asking at that review. Keep it concrete and specific to this decision. Prompt 10: Create a follow-up checklist The decision is logged. Now what? Who does what, by when, and how does anyone know it happened? Act as a decision documentation assistant. We've made the following decision and documented it: Decision summary: [SUMMARY] Owner: [NAME] Next steps identified: [LIST STEPS] Convert this into a follow-up checklist with: each action, the person responsible, the due date, and a brief note on what "done" looks like for that action. Flag any steps where an owner or date is missing. Do not assign owners yourself. For decisions that feed into a larger project, the AI project management prompts will help you plug these actions into a broader tracking structure. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What AI decision log prompts can't do This is the part people skip because they want the good feeling of a well-formatted document without the hard work of actual accountability. AI is useful for structure, summaries, tradeoff tables, stakeholder versions, and follow-up checklists. It will produce a clean, professional-looking document regardless of whether the underlying decision was sound. What it can't do: Invent consensus that didn't happen Replace the conversation with the person most affected Tell you whether the decision was actually the right one Cover for a decision that needed legal, security, HR, or financial review and didn't get it If a decision involves those areas, the log is not the review. It's the record of the review that should have happened with qualified humans first. AI risk assessment prompts can help you map the blast radius before you commit, but they don't replace a lawyer's sign-off or a security team's check. A polished decision memo is not leadership. It's a document. How to use AI at work is worth reading if you're trying to figure out where the line is between "AI made this faster" and "AI gave me the feeling of thoroughness without the substance." The teams that stay out of the endless re-litigation meetings aren't the ones with the prettiest decision logs. They're the ones that wrote down the real reason, named a real owner, and set a real date to revisit. AI helps you do that faster. You still have to do it. Frequently asked questions What should a good AI decision log include? A good decision log includes the decision made, the date, the context, the options considered, the tradeoffs, the rationale for the final choice, the owner, key assumptions, risks flagged, and a review date. Use AI to help structure these sections quickly, but make sure all facts, owners, and dates come from you, not the model. Can I use ChatGPT to document decisions from a Slack thread? Yes, but clean the thread first. Remove names, customer data, PII, and any confidential or unreleased information before pasting it into an unapproved tool. Ask the model to flag any gaps rather than fill them. A Slack thread often contains negotiation and noise, not a clear decision, so be explicit in the prompt about what you need extracted. What's the difference between a decision log and a meeting summary? A meeting summary captures what was discussed. A decision log captures what was decided, why, who owns it, and when to revisit it. You often need both, but they serve different purposes. Use AI meeting notes prompts for summaries and the templates in this article for the decision record itself. What decisions should NOT be documented using AI tools? Don't use unapproved AI tools to document decisions involving employee performance, legal disputes, security incidents, unreleased financials, board materials, personal data, or sensitive client information. The decision log itself may contain that information, and pasting it into a public LLM creates a data risk. Use your organisation's approved environment or strip sensitive details first. How do I stop teams from re-litigating decisions? Write down the decision at the time it's made, including why the rejected options were ruled out. Name an owner. Set a review date. Share the log with everyone affected. When someone raises the rejected option again, point them to the rejection record. The re-litigation problem is almost always an information problem, not a people problem. What if the AI fills in gaps I didn't give it information for? That's the main risk with decision documentation prompts. Always add a line to your prompt: "Flag any gaps in the information I've given you. Do not invent facts, consensus, or approvals that aren't in the source material." Then read the output critically. If it sounds too complete, it probably filled something in. Review every named fact against your original notes before treating the log as final. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance --- ## AI Decision-Making Prompts: 10 Templates That Work URL: https://dontreplace.me/articles/ai-decision-making-prompts-think-clearer-not-less ai decision makingdecision promptsai at workworkplace aidecision framework AI Decision-Making Prompts: 10 Templates That Work 10 copy-paste AI decision-making prompts for ChatGPT and Claude. Structure messy decisions, run pre-mortems, and draft memos without outsourcing your judgment. June 12, 2026 Updated June 12, 2026 Bad decisions have a format. They're usually made in a meeting that ran 20 minutes over, with six people who wanted different things, zero written criteria, and someone saying "let's just go with our gut" right before lunch. Then someone else sends a recap email that invents consensus. You know the one. AI decision-making prompts won't fix the politics. But they can do something genuinely useful: force the mess into a structure before it becomes a regret. If you know what you're deciding, what matters, what you're risking, and what you don't know yet, you're already ahead of most decision processes. AI can help you get there faster. It can't make the call for you, and it shouldn't. Here's how to use it without handing over your judgment. Why AI decision-making prompts work (and where they fail) AI is fast and confident. Those two qualities together are a trap if you're not careful. Throw a vague question at ChatGPT or Claude and it will hand you back a beautiful, structured, completely authoritative-sounding recommendation. The problem: it's optimizing for plausible, not correct. It doesn't know your company's real constraints. It doesn't know which executive will kill a project if it touches their budget. It doesn't know what your customers actually said last quarter versus what the last meeting slide claimed they said. Rule #5 in Don't Replace Me is blunt about this: it's not smart, it's fast. Fluency is not wisdom. AI will write you a confident decision memo whether or not the underlying logic holds up. Your job is to bring the constraints, the stakes, the context, and the final call. Used right, AI is a decision analyst with infinite patience and no ego. It will compare options, surface counterarguments, draft memos, and run structured thinking exercises without getting defensive about its preferred option. That's actually rare and useful. The catch is: garbage in, garbage fog out. Vague prompts get vague output dressed up as insight. The prompts below are built around that reality. They help you structure your thinking, not outsource it. Before you start: a quick word on what not to paste in Seriously, don't skip this part. Do not paste the following into any AI tool that hasn't been approved by your company's IT or legal team: confidential company strategy, unreleased product plans, acquisition details, customer names or private data, employee information, financial figures that aren't public, legal matters, security details, or system credentials. Ever. Anonymize everything. Replace "our Q3 acquisition of [Company Name]" with "a potential strategic acquisition in our sector." Replace customer names with "Client A." Replace specific revenue figures with ballpark ranges or just describe the scale. You can get 90% of the structured thinking value without giving AI anything it shouldn't have. Also: once you've got output, a human verifies it. All of it. Facts, numbers, timelines, risks, ethical implications, legal exposure, and stakeholder impact need human eyes before anything gets committed to paper or presented to anyone who matters. The reusable AI decision-making prompt formula Before you get to the 10 templates, here's the underlying structure they're all built on. You can use this as a starting point for any decision prompt you write: Context: [What's the situation? What's already decided or fixed?] The decision: [State it as a clear, specific question with a deadline or trigger if relevant.] Constraints: [What are the real limits? Budget, time, people, legal, political.] Options on the table: [List them, even the ones that seem unlikely.] What I need from you: [Comparison table? Risk list? Draft memo? Devil's advocate?] Audience: [Who will see or act on the output?] The clearer you are with those six inputs, the more useful the output. Skip any of them and you're back to getting plausible-sounding mush. The reason most people get bad AI output on decisions isn't that the tool is limited. It's that they paste in "should we do X or Y?" with no context and expect the response to be useful. It won't be. AI has no idea what X and Y actually cost you, who championed them internally, or what the fallback looks like if both go wrong. This formula forces you to think that through before you type anything. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 copy-paste AI decision-making prompts These are ready to use. Edit the brackets for your situation. Keep the structure. 1. Turn a messy situation into a clear decision statement I need help clarifying a decision before we analyze it further. Here's the situation: [Describe what's happening and why a decision is needed.] Help me write a clear, one-sentence decision statement in this format: "We need to decide [WHAT] by [WHEN] in order to [WHY IT MATTERS]." Then list any sub-decisions that depend on the main one. 2. Compare options against real constraints I'm comparing [NUMBER] options for [DECISION TOPIC]. Options: [List them.] Key constraints: [Budget range, timeline, team size, any hard limits.] What matters most: [List 3-5 criteria in rough priority order.] Create a comparison table scoring each option against each criterion. Use a simple scale (Low/Medium/High or 1-3). Note where you're uncertain rather than guessing. 3. Identify hidden assumptions Here is the decision we're considering: [State the decision.] Here is our preferred option: [Describe it.] List the assumptions we're making that this option depends on. For each assumption, tell me how likely it is to be wrong and what would happen to the decision if it is. 4. Build a risk register We're moving forward with [DECISION OR INITIATIVE]. Generate a risk register with these columns: Risk, Likelihood (Low/Medium/High), Impact (Low/Medium/High), Early warning signs, Possible response. Include at least one risk in each of these categories: operational, financial, people/stakeholder, and external/market. Flag anything you're uncertain about rather than filling in confident-sounding guesses. 5. Run a pre-mortem It's [DATE 12 MONTHS FROM NOW] and the decision we made to [DECISION] has failed badly. Write a short post-mortem report as if this already happened. What went wrong? What did we miss? What assumptions were wrong? What warning signs did we ignore? Be specific. Don't give me generic risks. Make it feel real. This is one of the most useful prompts on the list. It bypasses the optimism bias in most planning discussions. Nobody wants to be the pessimist in a meeting, but a fictional failure report gives everyone cover to say the uncomfortable thing. 6. Pressure-test a preferred option Our team is leaning toward [OPTION] for [DECISION]. Play the role of a skeptical senior stakeholder who has seen this kind of decision go wrong before. Ask me the five hardest questions you'd ask before approving this. Don't be gentle. Pair this one with the pre-mortem. Between the two, you'll surface more real problems in 15 minutes than most teams find in a two-hour review. 7. Write a one-page decision memo Help me write a one-page decision memo for [AUDIENCE, e.g., "a senior leadership team"]. Decision: [State it clearly.] Context: [2-3 sentences on why this needs to be decided now.] Options considered: [Brief list.] Recommended option: [What you're proposing.] Key reasons: [Top 3.] Main risks: [Top 2-3.] What we need: [Approval, resources, a decision by a specific date.] Keep it tight. Under 400 words. No unnecessary hedging. After you get the draft, read it out loud. If anything sounds like something you don't actually believe, rewrite that part yourself. AI will write confidently even when you're not. Don't let the formatting do the thinking. 8. Map stakeholder tradeoffs For the decision to [DECISION], different stakeholders will see different tradeoffs. Stakeholders involved: [List them by role or team, not name.] For each stakeholder, identify: What they gain from our recommended option, what they lose or give up, and what concern they're most likely to raise. Don't invent quotes or political dynamics you don't know. Flag where you're speculating. This prompt is good preparation for any decision that needs buy-in from people who weren't in the room when you made it. If you do proper AI-assisted research on your stakeholders beforehand, you'll get better output here. 9. Decide what you still don't know Here is our current understanding of [DECISION]: [Brief summary of what you know and what you're leaning toward.] What information would most change this decision if we had it? List the top 5 unknowns, ranked by how much they'd affect the outcome. For each, suggest one practical way to get better information before we commit. This prompt is underused and deeply useful. Most teams make decisions with information they have, never stopping to ask which missing pieces actually matter. This forces that conversation. Pair it with data analysis work when the unknowns are quantitative. 10. Turn a meeting debate into a recommendation Our team has been debating [DECISION] and hasn't reached agreement. Here are the main positions: Position A: [Describe it and who holds it.] Position B: [Describe it and who holds it.] [Add more if needed.] Help me write a recommendation memo that: acknowledges each position fairly, identifies where there's actual disagreement vs.different assumptions, and proposes a decision with a clear rationale. The goal is clarity, not a compromise that pleases no one. How to get better output from each prompt The templates above will get you somewhere. But there are a few habits that separate mediocre AI decision output from genuinely useful output. Be specific about the audience. "Write a decision memo" gets you something generic. "Write a decision memo for a CFO who approved similar spend last quarter but rejected a larger version six months ago" gets you something you can actually use. AI can adjust tone, emphasis, and framing based on audience context. Give it that context. Tell it what you've already ruled out. If two options got eliminated early, say so and say why. AI will sometimes resurrect them if you don't, which wastes your time and muddies the output. "We've already decided we're not building this in-house due to timeline constraints" saves you a loop. Ask it to flag its own uncertainty. The default mode for most AI tools is confident. Left to its own devices, it will estimate, infer, and project without flagging where it's guessing. If you explicitly prompt it to "mark anything you're uncertain about" or "note where you're speculating rather than stating a fact," you get output that's actually more honest and more useful. Iterate, don't accept. The first draft of any AI decision output is a starting point. Read it, identify the parts that feel wrong or incomplete, and prompt again. "The risk register looks thin on the people side. Add three more people or stakeholder risks, including any that might come from outside the immediate team" is a normal second prompt, not a sign the first one failed. Good output usually takes two or three passes. That's not a flaw. That's how thinking works. What AI should and shouldn't decide in this process Here's a clean line to keep in your head. AI is useful forHumans own Structuring optionsChoosing between them Writing comparison tablesWeighting what actually matters Surfacing possible risksAssessing which risks are real Drafting the memoSigning off on it Playing devil's advocateDeciding when to stop debating Listing unknownsDeciding what to do anyway Generating stakeholder concernsActually talking to stakeholders The thing AI can't do is be accountable. It doesn't have a job on the line. It doesn't know that the CFO will kill anything over $200K before Christmas. It doesn't know that two of your three options are politically toxic for reasons that never made it into any document. That context lives with you and your team, and human judgment in complex situations is exactly what makes the difference between a smart-looking memo and a decision that actually sticks. For a full breakdown of what AI genuinely can and can't do with your thinking, this plain-language explainer covers it without the hype. The rule: AI builds the structure. You fill it with the things that actually matter. Then you sign your name. Once you've got your decision memo drafted, the implementation side needs its own structure. The project management prompts pick up right where this leaves off. Frequently asked questions Can AI make decisions for me at work? No, and you shouldn't want it to. AI can organize information, surface options, draft memos, and flag risks, but it can't weigh the political, ethical, or relational factors that determine whether a decision actually works. The accountability stays with you. Use AI to think more clearly, not to outsource the call. What's the risk of using AI for decisions involving sensitive business information? It's real. Don't paste confidential strategy, customer data, employee information, financial details, legal matters, or unreleased product plans into any AI tool that hasn't been cleared by your IT or legal team. Anonymize everything. You can get 90% of the structural value without sharing anything sensitive. What's a pre-mortem and why does it help with AI prompts? A pre-mortem is an exercise where you imagine a decision has already failed and then work backward to figure out why. It bypasses the optimism that kills most planning sessions. When you use it as an AI prompt, you get a structured fictional failure report that's often more honest than anything a live team discussion produces. How do I make sure AI decision output is accurate? You verify it yourself. Check every fact, number, timeline, and assumption the AI cites. AI will fill gaps with plausible-sounding content when it doesn't have real information. Any output that will influence a real decision needs a human to confirm it before it gets shared or acted on. How is this different from just googling my problem? The structure. These prompts force you to clarify what you're actually deciding, what your constraints are, and what you need from the output. Google gives you information. A well-structured AI prompt gives you a thinking framework applied to your specific situation. The quality of the output depends entirely on how clearly you define the input. What if my team disagrees with the AI's recommendation? Good. The AI's recommendation is a starting point, not a verdict. If your team's disagreement is based on context, relationships, or judgment the AI didn't have, trust the team. If the disagreement is about assumptions you can test, test them. The point of the prompts is to make the disagreement more structured and productive, not to replace human debate with machine output. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Prompts for Work: 12 Templates You Can Use Today 12 copy-paste AI prompt templates for emails, reports, and decisions. Real examples for normal people, with what to check before you send anything. ai prompts for workchatgpt promptsworkplace ai Should I Tell My Boss I Use AI? 68% of workers hide their AI use from employers. Here's the honest answer on when to disclose, when to stay quiet, and how to handle it either way. AI at workworkplace AIChatGPT at work AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work --- ## AI Defect Triage Prompts: Sort Bugs Faster URL: https://dontreplace.me/articles/ai-defect-triage-prompts-sort-bugs-faster AI promptsdefect triagebug reportsQApost-launch AI Defect Triage Prompts: Sort Bugs Faster AI defect triage prompts for sorting bug reports, severity, duplicates, owners, workarounds, retests, and customer impact after launch. August 02, 2026 Updated August 02, 2026 The bug is not “urgent” because someone used three siren emojis. The bug is urgent when the evidence, customer impact, release context, and rollback risk say it is. That is where AI defect triage prompts help. Not because AI can magically inspect production, read private tickets, verify logs, or decide business priority from vibes. It cannot. But it can turn scattered bug reports, support notes, screenshots, reproduction steps, owner debates, severity arguments, workaround ideas, and retest evidence into a triage process that does not make everyone relive the same Slack fight twelve times. Defect triage is the boring adult part of shipping software. After launch, someone has to decide what is real, what is duplicate noise, what is customer-impacting, what needs a hotfix, what can wait, and what was actually expected behavior wearing a fake mustache. If you are still preparing the release, start with AI smoke test prompts or AI launch checklist prompts. If the launch is live and bugs are arriving sideways, this guide is for you. AI can organize defect triage. A human still owns the evidence, customer impact, security/privacy review, priority, and final decision. What is defect triage? Defect triage is the process of reviewing reported bugs, confirming what is known, grouping duplicates, assessing severity and priority, assigning owners, choosing next actions, and tracking the decision. A useful defect triage process answers these questions: What exactly happened? Can someone reproduce it? Which users, customers, workflows, versions, or environments are affected? Is this a duplicate, regression, configuration issue, expected behavior, or genuinely new defect? What is the customer impact? Who owns investigation, fix, communication, workaround, and retest? What gets fixed now, what waits, and what needs escalation? Without triage, every bug becomes theater. The loudest report wins. The newest screenshot resets the conversation. Someone says “P0?” and seven people open a meeting invite with no agenda, no repro steps, and one cursed attachment named image-2-final.png. AI defect triage prompts are useful because AI is good at structure. It can rewrite vague reports into clearer tickets, summarize sanitized support themes, create severity questions, group likely duplicates, draft owner follow-ups, and build retest checklists. That is a real productivity gain. But it is the same rule as what AI can and can't do: AI can make the mess easier to inspect. It cannot make the evidence true. The reusable AI defect triage prompt formula Use this formula when you want AI to help with defect triage: “You are a defect triage assistant. I am [your role] triaging defects for [product/release/workflow]. Here is sanitized context: [release scope, affected versions, bug reports, repro notes, customer impact summaries, known issues, logs summarized safely, support themes, owners, severity definitions, workaround options, rollback criteria]. Create [specific triage output]. Separate confirmed facts from assumptions. Flag duplicates, missing evidence, risks, owner questions, approval needs, and security/privacy concerns. Do not invent repro steps, customer impact, root cause, severity, owners, production status, commitments, or final decisions.” That last sentence is not decoration. AI loves filling gaps because filling gaps looks helpful. In defect triage, a filled gap can become a fake root cause, a fake owner, or a fake promise to a customer. Congratulations, you invented process debt with punctuation. Before using any prompt, sanitize the input. Do not paste customer PII, credentials, access tokens, raw production logs, private tickets, unreleased strategy, legal disputes, security vulnerabilities, financial records, HR issues, regulated data, medical information, private client conversations, or sensitive personal information into unapproved AI tools. If the prompt needs raw private data to work, rewrite the input. Use summaries, categories, counts, redacted excerpts, approved screenshots, and links to authorized systems instead. The robot does not need your customer’s email address to group three checkout bugs. What to bring before asking AI to triage bugs AI cannot triage what you never give it. It can only organize the ingredients in the bowl. If the ingredients are expired, you get a beautiful spreadsheet of food poisoning. Bring this before asking for defect triage help: InputWhy it mattersHuman check Release scopeKeeps triage tied to what actually changedConfirm it matches the shipped version Affected version or environmentSeparates production, staging, browser, device, and config issuesVerify with logs or system records Reproduction stepsTurns “it broke” into something testableMake sure steps are real, not guessed Expected vs. actual behaviorClarifies whether this is a defectConfirm expected behavior with product or requirements Customer impactGuides severity and communicationUse privacy-safe summaries and approved numbers Existing known issuesPrevents duplicate panicCheck the current known-issues list Severity definitionsKeeps “urgent” from meaning “loud”Use agreed criteria, not mood lighting Owners and backupsMakes work movableName real people, queues, or rotations Workaround statusHelps support customers safelyMark approved vs. proposed clearly Retest evidencePrevents zombie bugs from reopening foreverInclude environment, version, and date If you do not have these inputs, ask AI to create a missing-information checklist first. Do not ask it to declare priority from three screenshots and a feeling. For better raw reports before triage, pair this with AI bug report prompts. Good tickets make good triage. Bad tickets make performance art. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI defect triage prompts you can use today Prompt 1: Turn scattered bug reports into a triage table You are a defect triage assistant. I am triaging bugs for [release/product/workflow]. Here are sanitized bug reports, support summaries, screenshots described in text, affected versions, known issues, and owner notes: [paste details]. Create a defect triage table with columns for issue summary, confirmed facts, assumptions, affected workflow, likely duplicates, severity question, priority question, owner, next action, missing evidence, and customer communication needed. Do not invent root cause, severity, owner, or customer impact. What to check after: Make sure the table did not convert “maybe” into “confirmed.” A tidy table with fake certainty is just a lie wearing business casual. Prompt 2: Rewrite vague defects into reproducible tickets You are a QA ticket assistant. Here are vague defect notes: [paste sanitized notes]. Rewrite each into a clearer bug ticket with title, environment, affected version, steps to reproduce, expected behavior, actual behavior, evidence needed, customer impact question, suspected area if provided, and open questions. If reproduction steps are missing, do not guess them. List what the reporter must provide. What to check after: AI can polish the ticket. It cannot reproduce the bug for you unless you provide real steps or approved evidence. Prompt 3: Group likely duplicates safely You are helping with defect triage. Here are privacy-safe summaries of reported issues: [paste issue list]. Group likely duplicates by symptom, affected workflow, environment, error wording, timing, and suspected area. For each group, explain why the issues may be duplicates and what evidence would confirm or disprove the grouping. Keep uncertain matches in a separate “possible duplicate” section. What to check after: Duplicate grouping is a hypothesis, not a verdict. Do not close customer reports just because AI thought two symptoms sounded cousins. Prompt 4: Draft severity and priority questions You are a defect triage facilitator. Given these sanitized defect reports and our severity definitions: [paste details], draft the questions we need to answer before assigning severity and priority. Separate user impact, revenue impact, data risk, security/privacy risk, workaround availability, frequency, affected customer tiers, regression risk, and release blocker status. Do not assign final severity unless the evidence supports it. What to check after: Severity is about impact. Priority is about when you act. People mix those up, then wonder why the triage board looks like a raccoon organized it. Prompt 5: Separate customer impact from internal annoyance You are a customer-impact analyst for defect triage. Here are sanitized reports, support notes, and internal complaints: [paste details]. Separate confirmed customer impact, suspected customer impact, internal inconvenience, cosmetic issues, operational risk, and unknowns. For each item, list evidence needed before external communication or escalation. What to check after: Internal pain matters, but “our dashboard is annoying” is not the same as “customers cannot pay invoices.” Different animals. Different cages. Prompt 6: Check release scope and affected versions You are a release triage assistant. Here is the shipped scope, affected versions, environments, deployment notes, and defect summaries: [paste sanitized details]. Map each defect to the likely related release area if evidence exists. Mark anything that may be unrelated, pre-existing, configuration-specific, environment-specific, or missing version evidence. Do not claim causation without proof. What to check after: “Happened after release” does not always mean “caused by release.” Time is suspicious, but it is not a root-cause analysis. Prompt 7: Create workaround options without overpromising You are a support-safe workaround assistant. Here are confirmed defects, constraints, support policies, and any proposed workarounds: [paste details]. Create a workaround options table with defect, workaround, who can use it, customer-facing wording if approved, risks, owner approval needed, expiration date, and follow-up action. Mark unapproved workarounds clearly as draft only. What to check after: A workaround can create its own bug farm. Get product, engineering, support, security, legal, or customer-success approval when the workaround touches customer commitments or sensitive data. Prompt 8: Prepare a stakeholder triage update You are a stakeholder update assistant. Here is the current defect triage state: [paste sanitized triage table]. Draft a concise update for [audience]. Include what changed, confirmed high-impact defects, open investigations, decisions made, decisions needed, owner actions, customer communication status, workaround status, and next update time. Separate confirmed facts from assumptions. What to check after: Stakeholder updates should reduce panic, not launder uncertainty. If you do not know root cause, say you do not know root cause. For ongoing launch support, AI post-launch monitoring prompts can help turn recurring signals into cleaner updates. Prompt 9: Build a retest checklist You are a QA retest assistant. Here are defects marked fixed or ready for validation: [paste sanitized defect list]. Create a retest checklist with defect ID, fix summary, environment, version/build, test data requirements, steps to retest, regression checks, evidence to capture, owner, pass/fail criteria, and reopen conditions. Flag missing information. What to check after: A fix without retest evidence is a bedtime story. Pleasant, perhaps. Not production confidence. If retesting reveals something uglier, AI incident review prompts can help after the immediate fire is out. Prompt 10: Write the final triage decision log You are a defect triage documentation assistant. Here are the final decisions from triage: [paste sanitized decisions]. Create a decision log with defect, decision, evidence considered, severity, priority, owner, fix or workaround path, customer communication status, retest requirement, risks accepted, approver, date, and follow-up review. Mark anything missing before sign-off. What to check after: The decision log is not bureaucracy. It is how future-you proves the team did not just pick the loudest Slack message and hope. Where AI helps and where humans must stay in charge AI helps most when the work is messy but bounded. Defect triage has a lot of that: grouping notes, rewriting tickets, spotting missing fields, drafting questions, summarizing themes, and formatting decision logs. AI should not own judgment. It should not assign final severity, approve customer messages, decide whether to roll back, interpret legal/security exposure, publish workaround instructions, or tell support what customers are owed. It especially should not touch raw production logs, credentials, private tickets, or customer data unless your organization has approved tooling and rules for that data. Use AI like a fast triage clerk. Give it sanitized inputs. Ask it to separate facts from assumptions. Make it flag gaps. Then have the responsible humans decide. This is also why defect triage connects to AI rollback plan prompts and AI support handoff prompts. Triage decides what is happening. Rollback decides whether to undo it. Support handoff makes sure customers and frontline teams are not abandoned with a pile of half-known problems. A simple defect triage workflow with AI Here is the practical version: Collect reports in one place. Sanitize anything going into AI. Ask AI to create a triage table, not a verdict. Have humans verify repro steps, affected versions, impact, severity, and owners. Use AI to draft stakeholder updates and retest checklists. Record final decisions in a decision log. Revisit open defects on a predictable cadence until fixed, accepted, or closed. The boring cadence matters. A triage meeting without a table is a podcast. A triage table without decisions is a shrine. A decision without follow-up is how bugs become folklore. Frequently asked questions Can AI triage bugs automatically? AI can help organize bug triage, but it should not automatically make final triage decisions. It can group duplicates, summarize reports, draft severity questions, and format retest checklists. Humans still need to verify evidence, customer impact, security/privacy risk, priority, and accountability. What should I remove before pasting bug reports into AI? Remove customer PII, credentials, access tokens, raw logs, private tickets, security vulnerabilities, legal issues, financial records, medical information, regulated data, private customer conversations, and internal strategy. Use sanitized summaries, categories, counts, redacted excerpts, and links to approved systems instead. Are AI defect triage prompts good for QA teams? Yes, AI defect triage prompts are useful for QA teams when they need cleaner tickets, duplicate grouping, missing-evidence checks, retest plans, and decision logs. They are not a replacement for QA judgment, exploratory testing, environment knowledge, or release-risk decisions. Can ChatGPT decide bug severity? ChatGPT can help draft severity questions against your definitions, but it should not decide severity alone. Severity depends on verified customer impact, frequency, data risk, security/privacy implications, workaround availability, and business context. Those are human-owned decisions. How do I use AI for duplicate bug reports? Give AI sanitized issue summaries and ask it to group likely duplicates by symptom, workflow, environment, timing, error wording, and affected version. Require it to explain why issues may be duplicates and what evidence would confirm or disprove the match. Keep uncertain matches open. What is the biggest risk of using AI for defect triage? The biggest risk is false certainty. AI can make incomplete evidence look organized, which can trick teams into accepting guessed root causes, fake severity, fake customer impact, or premature closure. Use it to surface gaps, not to hide them. Final thought: AI can sort the pile, not own the call AI defect triage prompts are useful because defect work is mostly messy thinking under time pressure. AI can make the thinking visible. It can turn chaos into a table, a checklist, a decision log, or a stakeholder update. But the value is still human: judgment, taste, accountability, customer empathy, security awareness, and knowing when a “small bug” is actually the loose thread holding the sweater together. That is the larger point of Dmitry Kargaev’s book Don’t Replace Me. Use AI to move faster through the sludge. Do not outsource the part where someone has to know what matters. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management AI Change Impact Analysis Prompts: Spot Damage AI change impact analysis prompts for mapping affected systems, teams, customers, QA checks, support gaps, and rollout risks. AI promptschange impactrisk AI QA Checklist Prompts: Catch Mistakes Before They Ship 10 copy-paste AI QA checklist prompts for catching bugs, edge cases, and gaps before launch. Includes safety rules and a reusable prompt formula. AI promptsQAquality assurance --- ## AI Delegation Prompts: 10 Templates for Managers URL: https://dontreplace.me/articles/ai-delegation-prompts-hand-off-work-without-abdicating ai delegation promptstask briefdelegation at workmanager promptsai at work AI Delegation Prompts: 10 Templates for Managers 10 copy-paste AI delegation prompts for managers. Write clearer task briefs, define done, spot risks, and prep handoffs without abdicating judgment. June 23, 2026 Updated June 23, 2026 Most managers don't have a delegation problem. They have a clarity problem. The task lives in their head, half-formed. The handoff comes out vague. The person receiving it guesses wrong. Then there's a follow-up meeting to fix what the original conversation should have prevented. AI delegation prompts don't solve the people part. But they're genuinely good at the clarity part: turning the mess in your head into a brief that actually tells someone what you need, by when, and what "done" looks like. That's the job here. Not outsourcing your judgment. Not generating fake accountability. Just using AI to prep better handoffs, faster, before you open your mouth or send the Slack message. Here's how to do it without abdicating. Why AI delegation prompts work (and where they fail) The reason vague delegation is so common isn't laziness. It's that managers often know what they want without being able to articulate it. "Can you sort out the client report situation?" makes perfect sense in your head. To the person receiving it, it's noise. AI is useful here because it forces you to externalize your intent. When you describe a task to a language model, you have to write it down. That act alone catches missing context. You realize mid-sentence that you haven't defined the deadline, or that you don't actually know who owns the decision. The other thing AI does well is ask obvious questions you've stopped asking yourself. You've been close to the task for days. You know what "good" looks like. But the person you're handing it to has none of that context, and AI, when prompted correctly, will surface the gaps that familiarity makes invisible. Where AI fails is when you skip that step and just paste a vague thought in, expecting a complete brief to come back. As Rule #13 in Don't Replace Me puts it: garbage in, garbage out. A polished handoff built on invented context is worse than a rough one built on real facts, because it looks finished when it isn't. AI should never invent the owner, the deadline, the quality bar, the budget approval, or the strategic priority. You supply those. AI helps you structure them. Before you start: what not to paste into AI tools A quick word before the templates. Do not paste any of the following into an AI tool that hasn't been approved by your company for sensitive data: customer PII, employee performance records, credentials or access details, confidential product or business strategy, private financial information, legal disputes or contracts, security incidents, HR issues, or board materials. If your delegation brief requires any of that, write the sensitive parts yourself and use AI only for the structural scaffolding around it. Most of what a good task brief actually needs (scope, deadline, success criteria, dependencies, risks) doesn't require sensitive data anyway. If you're unsure whether your company has an approved AI policy, assume you don't and treat every AI tool as a public surface until told otherwise. The downside of caution here is a slightly less convenient brief. The downside of carelessness is an HR or legal incident. The delegation prompt formula Every useful AI delegation prompt has the same bones: Act as [role]. I need to delegate [task]. Here's what I know: [context]. The goal is [outcome, not activity]. The owner will be [person or role]. The deadline is [specific]. Done means [measurable criteria]. Constraints are [budget, scope, tools, people]. Risks I see: [list]. Draft a [brief / checklist / follow-up / stakeholder update]. You fill in the brackets. AI drafts the output. You review it before it goes anywhere near another human being. That last part isn't optional. A generated brief needs your eyes on it before you send it. Not because AI is untrustworthy, but because you're the one accountable for whether the work gets done right. If you want to understand the broader limits of what AI can and can't actually do in a work context, this plain-language explainer is worth five minutes. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 copy-paste AI delegation prompts Use these as starting points. Replace the bracketed sections with your actual information. Prompt 1: Should I even delegate this? I'm a [your role] considering whether to delegate the following task: [describe task]. Here are the relevant factors: my current workload, the skill level of my team, the deadline, the stakes if it goes wrong, and whether it requires my specific authority or knowledge. Help me think through whether to delegate, keep it, or break it into parts. Ask me for any missing information before you respond. Prompt 2: Write a clear task brief Write a task brief for the following handoff. Task: [describe task]. Owner: [name/role]. Goal: [specific outcome]. Deadline: [date and any milestones]. Scope: [what's in and what's out]. Resources available: [tools, budget, people]. Definition of done: [what success looks like]. Known risks: [any blockers or dependencies]. Format it as a short written brief I can share directly. Prompt 3: Define "done" before you delegate I need to delegate [task] but I'm struggling to define what a completed version looks like. The purpose of this task is [business reason]. The audience or end user is [who benefits]. Here are my rough quality expectations: [describe]. Help me turn this into a clear definition of done with 3-5 measurable acceptance criteria. Flag any areas where I still need to make a decision. Prompt 4: Find the missing context I'm about to delegate this task: [describe task]. Here's the brief I've drafted: [paste brief]. Review it and identify: what context is missing that the person receiving it would need, what assumptions I've made that should be stated explicitly, what questions they'll likely ask me right away, and what risks aren't addressed. Give me a list of gaps, not a rewritten brief. Prompt 5: Turn a messy request into a handoff I need to turn this rough request into a proper task handoff: "[paste your rough idea or Slack message]." The person receiving it is [role/seniority]. Help me rewrite it as a clear delegation with: what needs doing, why it matters, the deadline, what success looks like, and what they have authority to decide on their own versus what needs my sign-off. Prompt 6: Write acceptance criteria I'm delegating a [describe type of task: report, design, analysis, etc.]. Help me write acceptance criteria so the person doing it knows exactly when it's ready for review. The goal is [outcome]. The audience is [who will use it]. My non-negotiables are [quality, format, tone, constraints]. Format these as a numbered checklist the person can check off before submitting. Prompt 7: Spot delegation risks I'm planning to delegate the following task: [describe task]. The person I'm delegating to is [role, experience level]. The stakes if this goes wrong: [describe]. Review this delegation and identify: skill gaps that might cause problems, dependencies on other people or systems, areas where my expectations aren't clear, and situations where this person will need to escalate. Give me a risk list, not reassurance. Prompt 8: Write a coaching-style handoff I want to delegate [task] in a way that develops [person's role], not just gets the task done. Here's context about where they are: [skills they have, skills I want them to build]. Help me write a handoff that explains what I need, but also frames it as a learning opportunity, includes a question for them to think through, and tells them what support they can expect from me. Prompt 9: Prepare a follow-up check-in I delegated [task] to [person/role] on [date]. The deadline is [date]. It's now [date] and I want to check in without micromanaging. Help me write a brief check-in message that: acknowledges their progress, asks one specific question about status or blockers, and reminds them of the deadline without making it sound like distrust. Tone: [direct/warm/neutral]. Prompt 10: Summarize progress for stakeholders I need to update [stakeholder: executive/client/board] on the status of [project or delegated initiative]. Here's what I know: [brief status update in your own words]. The original goal was [outcome]. Current status: [on track / at risk / complete]. Key blockers: [list]. Next milestone: [what and when]. Write a 3-4 sentence stakeholder update I can send by email or include in a status report. For more structured templates specifically for project work, the AI project management prompts here cover briefs, risk logs, and stakeholder updates in detail. How to adapt these prompts for your context The 10 prompts above cover the most common delegation scenarios, but work isn't one-size-fits-all. A few variations worth knowing. For remote teams, add one line to any brief prompt: "This person is working asynchronously and may not be able to ask me questions in real time. Identify anything that needs clarification upfront." That single instruction forces the AI to flag ambiguity before it becomes a three-day delay waiting for a reply. For cross-functional work, where the person you're delegating to doesn't report to you, add: "This person isn't on my direct team and I have no formal authority over them. Adjust the tone and framing accordingly." Briefs written as though you're directing a direct report land badly when you're actually asking for a favor across org lines. For high-stakes handoffs, where failure has real consequences, use Prompt 7 first, then Prompt 2. Run the risk check before you write the brief. You'll catch things that change what the brief needs to say. For junior team members, pair Prompt 2 with Prompt 8. A clean brief plus a coaching frame tells them what to do and how to grow from doing it. That's more useful than either one alone. The formula stays the same. You're just adjusting the context you feed in. What AI should never decide for you You can use every prompt above and still fail at delegation if you let AI fill in the blanks that require human judgment. AI should not invent: the task owner (that's a management decision with real accountability attached), the deadline (unless it's been confirmed by the people affected), the quality bar (vague AI-generated standards aren't your quality bar, they're average), budget approval (AI doesn't know what's been signed off), or anything touching employee performance, legal matters, HR issues, or client commitments. If you're using AI to prep a brief and you find yourself accepting AI-generated answers to these questions because you don't know them yet, stop. Those gaps are the conversation you need to have first. No AI prompt fixes a decision you haven't made. There's a related issue for managers: it's easy to mistake a polished AI-generated brief for actual management work. Writing the handoff isn't the same as understanding the task, knowing your team's capacity, or making the judgment call about who should own it. For a broader look at using AI thoughtfully in a management context, the AI manager prompts guide covers 1:1s, feedback, and team updates in the same practical format. The accountability part that no prompt can replace Good delegation has one thing AI can't generate: a real conversation with the person you're delegating to. A written brief is the setup. The handoff itself, where you confirm they understand, have the capacity, agree to the deadline, and know when to escalate, happens between humans. Sending a polished AI-generated brief and considering the task delegated is just a fancier version of the same vague Slack message you sent before. The other thing that slips when managers over-rely on AI-prepped handoffs is calibration. You learn how to delegate by watching what goes wrong. If your briefs are always polished and your check-ins are always scripted, you lose the signal that tells you which team members need more context, which ones are sandbag-estimating deadlines, and which ones are quietly drowning. AI smooths out the friction that would otherwise teach you something. Use the prompts to prepare. Then have the conversation. Review the brief before you send it. Check in at the agreed point. That's still your job. If you want a smarter framework for the whole thing, including how to audit which parts of your work are worth delegating in the first place, the workflow audit prompts are a good next step. And for the prioritization piece that comes before delegation, these prioritization prompts walk through how to decide what actually needs your attention versus what can be handed off. The point isn't to generate more documentation. It's to hand off work clearly enough that you don't have to redo it. Frequently asked questions What are AI delegation prompts? AI delegation prompts are structured inputs you give to a language model like ChatGPT or Claude to help you prepare task briefs, define success criteria, identify risks, and write handoff messages. They don't replace your judgment about who owns a task or what "done" looks like. You supply the real facts; AI helps you structure them into something clear enough to actually share. Can AI write a task brief for me? Yes, with your input. AI can turn a rough description into a structured brief with scope, deadline, acceptance criteria, and risk flags, but only if you give it accurate context. If you paste in vague intent and accept whatever comes back, you'll get a polished brief built on invented assumptions. Always review before sending. See the no-BS starter guide to using AI at work for how to get useful output without the theater. What should I never paste into an AI delegation prompt? Don't paste employee performance records, customer PII, credentials, confidential financial or strategic information, legal disputes, HR issues, security incidents, contracts, or board materials into any AI tool that hasn't been approved by your organization for sensitive data. Most delegation briefs don't require this information. If yours does, write the sensitive parts yourself and use AI only for the structural scaffolding. Should AI decide who owns a delegated task? No. Owner assignment is a management decision with real accountability attached. AI doesn't know your team's capacity, strengths, political dynamics, or development goals. It can help you write the brief once you've decided on the owner. It should never invent one. How is AI delegation different from just using a brief template? A template gives you structure. AI gives you a first draft based on what you describe, which is useful when the task is complex, when you're not sure what context to include, or when you want to spot gaps before you commit to the handoff. The best use is combining both: use a prompt to generate the draft, then review it against your own judgment before it goes anywhere. How do I check in on delegated work without micromanaging? Prompt 9 above covers this. The key is agreeing on a check-in point at the time of delegation, not adding one after the fact. If you built in a milestone review when you handed off the task, following up at that point isn't micromanaging. It's what you said you'd do. AI can help you write the message. The decision to check in, and when, is yours. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Automation Prompts: Automate the Boring Stuff 10 copy-paste AI automation prompts for mapping workflows, writing SOPs, adding approval gates, and QA testing. Practical templates with safety rules included. ai automation promptsworkflow automationSOP templates AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts --- ## AI Dependency Mapping Prompts: Find What Breaks URL: https://dontreplace.me/articles/ai-dependency-mapping-prompts-find-what-breaks AI promptsdependency mappingproject managementoperationsrisk AI Dependency Mapping Prompts: Find What Breaks AI dependency mapping prompts for finding upstream work, downstream risks, hidden handoffs, owners, schedule conflicts, and launch blockers. August 06, 2026 Updated August 06, 2026 A project dependency is anything your work needs before it can move, or anything that needs your work before it can move. Simple definition. Deeply annoying reality. The launch needs legal approval. Legal needs final copy. Final copy needs pricing. Pricing needs finance. Finance is waiting for usage data. The usage dashboard broke last Tuesday, and its owner is on vacation. Your project plan still says “launch Friday” because optimism is apparently a scheduling method. AI dependency mapping prompts can help turn that mess into a reviewable map. AI is useful for organizing supplied facts, grouping upstream and downstream dependencies, drafting questions, and spotting blank owner fields. It cannot inspect your systems, discover the undocumented spreadsheet holding the company together, verify a vendor promise, or decide that a risk is acceptable. AI can structure the dependency map. Humans still own the evidence, access, sequencing, escalation, decisions, and consequences. If you need to understand the broader blast radius of a change, start with AI change impact analysis prompts. If you already have a plan and need to expose what can block it, keep reading. What is dependency mapping? Dependency mapping is the process of identifying what a project, change, launch, or decision relies on—and what relies on it. A useful dependency map shows: the dependency itself; whether it is upstream or downstream; the accountable owner; the required date or sequencing rule; the source of the information; current status and confidence; what happens if it slips; the escalation or fallback path; the next review date. A bad dependency map is a list of vague nouns: “Engineering, legal, data, launch.” A good one says: “Legal must approve the revised cancellation language by August 12 before customer emails can be scheduled. Owner: Maya. Source: LEG-184. If late, delay the email rather than sending unapproved copy.” The difference is operational usefulness. One decorates a slide. The other helps people make decisions. AI can make the first draft faster. It can extract dependencies from notes, normalize inconsistent wording, group related items, propose review questions, and format a dependency register. But as explained in what AI can and can't do, fluent output is not verified reality. The reusable AI dependency mapping prompt formula Use this formula as your starting point: “Act as a dependency-mapping assistant. I am [role] planning [project, launch, migration, process change, campaign, client delivery, system cutover, or policy update]. Using only the privacy-safe information below, create [dependency map, register, review agenda, critical-path summary, or escalation brief]. For every dependency, include type, description, owner, source, required date, predecessor, successor, status, confidence, impact if late, validation step, fallback, and next action. Separate confirmed facts from assumptions. Flag missing owners, missing dates, circular dependencies, contradictory sequencing, external constraints, single points of failure, and claims that need validation. Do not invent systems, owners, commitments, approvals, dates, or technical behavior.” That final instruction matters. Language models complete patterns. If your notes imply that “someone in platform” owns a task, AI may turn that into a suspiciously specific person-shaped hallucination. A dependency map with invented certainty is worse than an honest blank. Before using any prompt, remove credentials, access tokens, customer PII, private tickets, raw production logs, security vulnerabilities, regulated data, unreleased strategy, legal disputes, HR records, confidential client material, and sensitive financial information. Use approved tools, redacted summaries, ticket references, and links to authorized systems. What to collect before mapping dependencies Do not feed AI a one-line project description and ask for truth. You will get a handsome fan-fiction version of operations. Bring the strongest available inputs: InputWhat it revealsHuman validation Scope and success criteriaWhat the map coversConfirm exclusions and boundaries Milestones and datesSequencing pressureDistinguish fixed dates from guesses Work breakdown or backlogTasks and deliverablesCheck against the active source of truth System and data notesTechnical relationshipsValidate with system owners Team and vendor responsibilitiesHandoffs and external constraintsConfirm named accountable owners Approval requirementsDecision gatesVerify policy, legal, security, and finance rules Customer and support impactsDownstream operating workUse privacy-safe evidence Risks and assumptionsUncertainty hiding inside the planLabel confidence and source Rollback or fallback optionsRecovery pathsConfirm they are tested and available Prior incident notesKnown failure patternsCheck whether the old problem still applies When information is missing, ask AI to produce questions, not answers. “Who approves the data retention change?” is useful. “Security approves the data retention change” is fiction unless security actually said so. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI dependency mapping prompts you can use today Replace bracketed text with privacy-safe, verified context. Keep links to source documents so reviewers can challenge the map without starting an archaeological dig. 1. Extract dependencies from a messy project plan “Review this privacy-safe project plan: [paste summary]. Extract every explicit dependency and every possible dependency that requires human validation. Return a table with dependency, upstream/downstream type, affected milestone, proposed owner field, source quote or section, required date, confidence, impact if late, validation question, and next action. Do not assign owners or dates unless stated. Put possible dependencies in a separate section labeled ‘Needs validation.’” This is useful at kickoff, when plans contain tasks but rarely explain why one task waits for another. Pair it with AI project kickoff prompts before everyone leaves the meeting believing someone else owns the hard parts. 2. Map upstream and downstream work “Using this deliverable and context: [paste], create two views. View one: everything this deliverable needs before it can start or finish. View two: every task, team, system, customer communication, report, or decision that depends on this deliverable. Include owners only when confirmed, source references, timing, validation steps, and consequences of delay. Flag unknown downstream consumers.” Teams naturally see upstream blockers because blockers hurt now. Downstream consumers are easier to forget until launch day, when support discovers its scripts describe a product that no longer exists. 3. Find critical-path assumptions “Analyze this milestone sequence: [paste]. Identify dependencies that appear to control the earliest possible completion date. Separate confirmed duration and sequence facts from assumptions. For each critical-path candidate, list evidence, owner, earliest start, required finish, slack if known, failure impact, and the question needed to validate it. Do not calculate fake precision when dates or durations are missing.” Use this to find where the schedule is pretending. A task with “two days” beside it may mean two working days after access, approval, data, and the one engineer who understands the old integration all become available. 4. Clarify owners and handoffs “Turn these tasks and team notes into a handoff-focused dependency register: [paste]. For each handoff, identify producer, receiver, artifact, acceptance criteria, delivery channel, required date, acknowledgment method, and escalation path. Leave unconfirmed owner fields blank. Flag handoffs with multiple accountable owners, no receiver, no acceptance criteria, or no acknowledgment step.” “Engineering sends it to marketing” is not a handoff. What is “it”? Which engineer? Which marketer? Where does it go? How does anyone know it is usable? Dependency maps become useful when nouns become observable actions. 5. Check vendors and external constraints “Review these privacy-safe vendor commitments, contract summaries, lead times, and project dates: [paste]. Build an external dependency table with provider, commitment, evidence source, notice period, lead time, renewal or blackout constraints, internal owner, validation status, impact if missed, fallback, and escalation date. Do not interpret contract language or provide legal conclusions. Flag anything requiring procurement or legal review.” External dependencies deserve extra suspicion because your team cannot fix them with an enthusiastic Slack thread. Confirm promises against contracts and current vendor communication. Use AI risk assessment prompts when a vendor dependency could cause material harm. 6. Map system and data dependencies “Using this approved, privacy-safe architecture and workflow summary: [paste], draft a system and data dependency map. For each connection, list source system, destination system, data or event transferred, trigger, frequency, owner, environment, monitoring evidence, failure behavior, recovery path, and validation question. Mark all inferred connections as unverified. Do not claim access to systems or logs.” Never paste secrets or raw sensitive data. The goal is to create a review worksheet for engineers and system owners, not to let autocomplete perform architecture cosplay. 7. Find schedule collisions and resource conflicts “Compare these project milestones, team allocations, blackout dates, approval windows, and maintenance periods: [paste]. Identify likely schedule collisions, shared-owner conflicts, impossible sequencing, compressed review windows, and dependencies that land on weekends, holidays, freezes, or unavailable periods. Show the source for each finding and propose questions or options, not final scheduling decisions.” This prompt helps expose the classic plan where the same person is critical to four simultaneous launches. AI can highlight the collision. A human manager must decide what moves. 8. Draft a dependency review meeting “Create a 30-minute dependency review agenda from this register: [paste]. Prioritize overdue validation, missing owners, high-impact external dependencies, critical-path risks, circular dependencies, and decisions required this week. For each agenda item, include the decision or confirmation needed, evidence link, required attendees by role, time box, and owner for the follow-up. End with a concise decision-log template.” A dependency review should resolve uncertainty, not provide a guided reading of the spreadsheet. Send the register beforehand. Use meeting time on decisions, owner commitments, and escalation. 9. Create an escalation-ready dependency brief “Turn these verified dependency facts into a one-page escalation brief: [paste]. Include the blocked outcome, dependency, confirmed owner, original and current required date, evidence, impact, options, tradeoffs, decision deadline, recommended decision owner, and immediate next step. Keep confirmed facts separate from assumptions. Do not assign blame or invent urgency.” When a dependency is stuck, clarity beats drama. For a fuller escalation workflow, use AI escalation plan prompts. 10. Update the map after decisions “Compare this previous dependency register with these verified decisions and status updates: [paste both]. Produce a change log showing added, removed, resolved, delayed, reassigned, and newly risky dependencies. Preserve source links. Flag contradictions, missing approvals, changed downstream impacts, and decisions that require updates to the roadmap, QA plan, support plan, cutover plan, or stakeholder communication. Do not silently overwrite uncertainty.” Dependency maps decay fast. Update them after scope changes, vendor news, staffing changes, architecture decisions, and launch reviews. A stale map is confidence theater with conditional formatting. How to review an AI-generated dependency map Do not review only the prose. Review the evidence and the blanks. Use this checklist: Scope: Does the map say what is included and excluded? Evidence: Does every confirmed dependency point to a source? Ownership: Is one accountable owner confirmed for each active item? Sequence: Are predecessor and successor relationships understandable? Dates: Are dates real commitments, planning targets, or guesses? Confidence: Are inferred dependencies visibly labeled? Impact: Does “high” mean something defined, or merely alarming? Fallback: Is there a usable option if the dependency slips? Escalation: Is there a decision owner and deadline? Freshness: Is the next review date visible? Then put the map in front of the people who know the work. Product validates scope. Engineering validates technical relationships. Operations validates process reality. Security, legal, finance, or compliance validate their own domains. Support validates customer-facing consequences. Vendors confirm their commitments in writing. AI does none of that merely because its table looks expensive. Common dependency-mapping mistakes Treating inferred dependencies as facts AI may reasonably infer that a customer email depends on approved copy. That does not prove which approval policy applies. Keep an “unverified” lane and close it with evidence. Naming teams instead of owners “Data” is not an owner. “Operations” cannot acknowledge a handoff. Use a named accountable human or an explicitly managed role with a real queue. Mapping only the happy path Include failure behavior, rollback, monitoring, support, documentation, training, and communication. AI operational readiness prompts can help test whether the organization is ready to operate the change after the confetti settles. Hiding uncertainty to make the plan look cleaner Unknown is a legitimate status. It is much safer than invented certainty. Add confidence labels such as confirmed, likely, unclear, and contradicted. Forgetting that dependencies change A dependency register is not a project kickoff souvenir. Review it before major gates, after material decisions, and whenever dates, scope, owners, vendors, or systems change. Confusing the map with the decision A map shows relationships and constraints. It does not choose risk appetite or approve a launch. For that, use a real decision process and AI go/no-go decision prompts as structure—not as the decider. Frequently asked questions Can ChatGPT create a dependency map from project notes? Yes, if the notes are privacy-safe and you treat the result as a draft. ChatGPT can extract tasks, relationships, dates, owner references, and open questions. Humans must verify every important dependency against current plans, systems, contracts, and responsible owners. What should an AI dependency mapping prompt include? Include project scope, milestones, tasks, systems, teams, external constraints, known owners, source references, dates, risks, and the exact output you want. Require separate labels for confirmed facts, assumptions, and missing information. Explicitly forbid invented owners, dates, approvals, and technical behavior. How is dependency mapping different from risk assessment? A dependency describes a relationship: one thing relies on another. A risk describes uncertainty that could affect an objective. A dependency can create risk, especially when it has no owner, weak evidence, tight timing, or no fallback. Use both views rather than forcing everything into one giant spreadsheet swamp. Can AI identify the critical path? AI can propose critical-path candidates from the sequence and durations you provide. It cannot know whether those inputs are complete or accurate. Validate durations, calendars, resource availability, constraints, and sequencing in the real planning system before acting on the result. How often should a dependency map be updated? Update it after material scope, schedule, staffing, vendor, architecture, approval, or rollout changes. Review it at recurring project checkpoints and before major launch gates. High-change projects may need weekly or even daily review; stable work may need less. What data should not go into an AI tool? Do not paste credentials, tokens, customer PII, raw production logs, private tickets, security vulnerabilities, regulated data, confidential contracts, legal disputes, HR records, unreleased strategy, or private client communications into an unapproved tool. Follow your organization's security and data-handling rules. Who owns an AI-generated dependency map? A human project or operational owner does. AI is a drafting tool. Accountable humans validate the contents, secure domain approvals, maintain the register, escalate blockers, and make decisions. What if the source information is incomplete? Use AI to create a missing-information checklist and stakeholder questions. Do not ask it to fill gaps with likely answers. An explicit unknown creates work; an invented answer creates damage. The useful rule Use AI to make dependency work faster to inspect, not easier to fake. A solid dependency map makes hidden relationships visible, puts owners and dates beside them, links claims to evidence, and exposes uncertainty before uncertainty becomes an outage, delay, or frantic executive meeting. AI can help draft that artifact quickly. Human judgment determines whether it describes the real world. If you want the broader survival guide for using these tools without confusing speed for intelligence, Don't Replace Me by Dmitry Kargaev is the field manual. The central idea fits dependency mapping perfectly: let the machine organize the mess; keep humans responsible for knowing which parts of the mess can actually break the business. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Change Impact Analysis Prompts: Spot Damage AI change impact analysis prompts for mapping affected systems, teams, customers, QA checks, support gaps, and rollout risks. AI promptschange impactrisk AI Dependency Register Prompts: Keep Projects From Surprising You Use AI dependency register prompts to map owners, blockers, assumptions, and downstream impact before one small change becomes a three-week incident. AI dependency mappingproject managementAI prompts AI Assumption Log Prompts: Stop Planning on Guesswork Use 10 AI assumption log prompts to expose project guesswork, design evidence checks, assign review owners, and stop unverified beliefs becoming commitments. AI promptsproject managementassumption logs --- ## AI Dependency Register Prompts: Keep Projects From Surprising You URL: https://dontreplace.me/articles/ai-dependency-register-prompts-keep-projects-from-surprising-you AI dependency mappingproject managementAI promptsoperationsrisk management AI Dependency Register Prompts: Keep Projects From Surprising You Use AI dependency register prompts to map owners, blockers, assumptions, and downstream impact before one small change becomes a three-week incident. September 11, 2026 Updated September 11, 2026 A project dependency is anything that has to be true before something else can move. A vendor has to deliver access. Engineering has to expose an endpoint. Legal has to approve language. Support needs a migration script. Someone has to remember that the “temporary” spreadsheet is now a load-bearing part of the business. Most projects do not fail because nobody made a plan. They fail because the plan described the happy path and quietly omitted the web of things around it. An AI dependency register can help. It can turn scattered notes into rows, group related risks, ask annoying questions, and make a draft reviewable. It cannot see your undocumented architecture, verify an owner, inspect production, or accept blame when the launch catches fire. That remains a human job. Tragically, humans still have to attend the meeting. This guide gives you ten AI dependency register prompts, an input format, a review workflow, and guardrails for using them without producing a beautiful spreadsheet of fiction. What an AI dependency register is actually for A dependency register is a living list of relationships that can affect delivery. A useful row usually includes: the item that depends on something else; the upstream dependency; the owner for each side; the expected date or condition; the evidence or source; the current status; the consequence if it slips; confidence in the information; the next review or escalation action. AI is good at extracting these fields from messy material. Give it meeting notes, a project brief, a release plan, or a set of approved tickets and it can propose a first pass. It is also good at comparing two lists and spotting that “customer import,” “data migration,” and “historical records” might be the same conversation wearing different hats. It is bad at knowing whether the person named in a six-month-old document still owns the system. It cannot infer that a vendor’s “two business days” excludes weekends, procurement, security review, and the one person who is on holiday. Treat the output as a hypothesis, not a control plane. For a broader blast-radius review, pair this process with AI change impact analysis prompts. If the project is still vague, start with AI project kickoff prompts before asking a model to map it. The input table: give the model something real The quality of the register depends less on clever wording than on the evidence you provide. Before prompting, collect a small, sanitized input table. FieldExample Work itemCustomer self-service export Desired outcomeCustomers can download a CSV of their records OwnerProduct: named person; Engineering: named person Known datePilot target: 2026-10-15 Upstream itemsPermissions model, API endpoint, export format Downstream itemsSupport article, analytics event, customer email ConstraintsPrivacy review, rate limits, legacy accounts EvidenceApproved ticket, decision log, vendor contract ConfidenceHigh, medium, or low Do not paste raw customer records, credentials, access tokens, private tickets, confidential strategy, legal disputes, HR records, regulated data, or production logs into an unapproved tool. Replace names and identifiers with role labels. Keep source links or document IDs in your own system so a human can verify each row later. A reusable prompt formula Use this structure when you need a new prompt: You are helping prepare a dependency register for [project]. Use only the supplied evidence. Extract [specific relationship]. Return [exact columns]. For every inference, label it as a hypothesis and explain what a human must verify. Do not invent owners, dates, systems, or approvals. Separate confirmed facts from assumptions and open questions. That final instruction matters. Without it, a model will happily turn “someone from security should review this” into “Security owner: Alex, approval due Friday.” Polished nonsense is still nonsense. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI dependency register prompts you can copy Prompt 1: Extract dependencies from a project brief Read the project brief below. Extract every explicit dependency and every dependency that is strongly implied by the stated outcome. Return a table with: dependent work item, upstream dependency, owner if explicitly named, required condition, date if explicitly stated, source quote, confidence, and verification question. Do not invent missing details. Put implied items in a separate section labeled “hypotheses.” Project brief: [paste sanitized brief] This creates the first pass. The source quote is important because it lets a reviewer challenge the row instead of arguing with an AI summary. Prompt 2: Map upstream and downstream effects Build a two-direction dependency map for [change or deliverable]. List what must happen before it can proceed and what may be affected after it ships. For each item, include the likely owner, evidence, failure mode, and a question that would confirm or reject the relationship. Use three labels: confirmed, plausible, and unverified. Only call an item confirmed when the input explicitly supports it. Inputs: [sanitized plans, tickets, decisions, and system notes] This is useful when a team has mapped prerequisites but forgotten consequences. “We can deploy” is not the same as “customers can use it safely.” Prompt 3: Find missing owners and handoffs Review this dependency list for missing ownership and handoff ambiguity. Return: dependency, current owner text, ambiguity, suggested role to verify, handoff question, and escalation threshold. Do not assign a person unless the source names one. Flag rows where one team is expected to provide an input but no acceptance criteria or receiving owner exists. Dependency list: [paste sanitized register] Ownership is not a decorative column. “Team” is often a polite way of saying nobody has agreed to do it. Prompt 4: Check dates, sequencing, and critical assumptions Examine the dates and sequence below. Identify dependencies that appear impossible, circular, under-specified, or reliant on an unstated assumption. Show the claimed sequence, the concern, the evidence needed, and a question for the project owner. Do not declare a schedule impossible without explaining what fact would change that conclusion. Schedule and assumptions: [paste sanitized plan] This catches the classic project trick: a task is marked complete because it started, while three downstream tasks assume it is finished. Prompt 5: Map vendor and external dependencies Extract all dependencies on vendors, contractors, platforms, regulators, customers, or external teams. Return: external party, promised input, contract or source evidence, lead time, internal prerequisite, failure scenario, fallback, and owner who must confirm it. Flag anything based only on an email, verbal promise, or generic documentation. External dependency notes: [paste sanitized notes] External dependencies deserve their own view because “we are waiting on them” is not a mitigation plan. Add a fallback or explicitly accept the exposure. Prompt 6: Trace data and system dependencies Using only the sanitized system notes below, draft a data and system dependency register. Include source system, destination, data or event, interface, owner, sensitivity classification if stated, expected timing, failure impact, and verification task. Separate documented relationships from guesses. Never infer that two systems are connected merely because their names appear near each other. System notes: [paste sanitized architecture summary] A human engineer, architect, or security reviewer must validate this output. The model has not inspected your network, permissions, queues, schemas, or logs. It has read words about them. Prompt 7: Turn dependencies into review questions Convert this dependency register into a review agenda. Group questions by product, engineering, operations, security, legal/compliance, support, and customer impact. Prioritize questions by consequence and uncertainty. For each question, name the evidence or owner needed to close it. Do not repeat a row as a question unless the question would change a decision. Register: [paste sanitized register] This is where AI earns its lunch: not by pretending to approve the plan, but by making the next human conversation less vague. Prompt 8: Draft an escalation-ready register Reformat the register for escalation. Include dependency, impact if missed, earliest warning signal, current status, owner, decision needed, deadline for action, and escalation path. Preserve the original evidence and mark every inferred field as “to verify.” Keep the tone factual and remove blame, speculation, and invented certainty. Register and project context: [paste sanitized materials] An escalation document should make a decision easier. It should not be a courtroom brief assembled by a frightened spreadsheet. Prompt 9: Compare two register versions Compare version A and version B of this dependency register. Report added, removed, changed, and unchanged dependencies. For each change, include the old value, new value, source evidence, likely impact, and a human verification task. Pay special attention to owner changes, date changes, status changes, scope changes, and rows that disappeared without an explanation. Version A: [paste sanitized register] Version B: [paste sanitized register] Version comparison is useful after scope changes, vendor updates, or a decision that seemed small at the time. Deleted rows deserve questions too. Prompt 10: Create the next review checklist Based on this dependency register, create a review checklist for [date or milestone]. Include only actions that can be verified. For each action, provide owner role, evidence required, pass condition, failure consequence, and follow-up if incomplete. Separate “must complete before milestone” from “monitor after milestone.” Do not mark anything complete based on the register alone. Register: [paste sanitized register] Milestone: [describe milestone] This turns the register from a museum of concerns into a working control loop. Update it when decisions change reality. A small example: the export that depended on everything Imagine a team building a customer export. The project brief says the feature needs an API endpoint, a permission check, a CSV format, a support article, and a launch email. A shallow map says: “Engineering builds API; Product launches.” A better register surfaces questions: Does the permission model cover legacy accounts? Is the export rate-limited, and who owns the limit? Does the CSV contain fields that require privacy review? Can support explain partial exports and failed downloads? Does the analytics event avoid recording sensitive customer values? Does the launch email promise availability before the rollout is complete? AI can extract those questions from the brief and arrange them by owner. It cannot confirm the answers. The team must inspect the permission rules, approve the data fields, test failure cases, and sign off on customer language. That is the point: expose the work before the calendar exposes you. A simple review cadence Use the register at four moments: At kickoff: identify dependencies and unknown owners. Before a major decision: update assumptions, evidence, and blast radius. Before launch or handoff: verify must-have dependencies and explicit sign-offs. After a change: compare the old and new register, then record what reality taught you. For risk-heavy work, use AI risk assessment prompts to score uncertainty separately from impact. Before release, AI operational readiness prompts can turn open dependencies into a go/no-go checklist. If acceptance is fuzzy, use AI acceptance criteria prompts before anyone claims the work is done. What AI can and cannot do here AI can: extract candidate relationships from supplied text; normalize inconsistent labels; compare versions and highlight changes; group questions by owner or risk; draft a review agenda and checklist; make missing evidence easier to notice. AI cannot: verify that a system actually behaves as described; know whether an owner has capacity or authority; approve security, legal, compliance, financial, or customer-impacting decisions; replace architecture, operations, or subject-matter review; take accountability for a wrong dependency map. The useful posture is simple: let the model make the register easier to read and harder to ignore. Keep verification, judgment, and sign-off with people who can access the truth. Frequently asked questions Can ChatGPT create a dependency register from meeting notes? It can draft one from sanitized notes, but every row still needs a source, an owner check, and a confidence label. Meeting notes contain omissions, shorthand, and opinions. Treat the output as a starting hypothesis. What columns should a dependency register include? At minimum: dependent item, upstream dependency, owner, condition or date, status, impact, evidence, confidence, and next action. Add sensitivity, fallback, or approval fields when the work warrants them. How do I stop AI from inventing dependencies? Tell it to use only supplied evidence, separate confirmed facts from hypotheses, preserve source quotes, and never invent owners or dates. Then sample-check the rows against the source material. Should dependencies be tracked in a spreadsheet or project tool? Use the tool where the team already reviews work. A spreadsheet is fine for an initial map; a project or risk system is better when owners, dates, notifications, and audit history matter. The format matters less than keeping it current. Is a dependency register the same as a risk register? No. A dependency describes a relationship or prerequisite. A risk describes uncertainty and potential harm. They overlap: a late vendor dependency can create a schedule risk. Track the relationship and the consequence instead of forcing one row to do both jobs. Can an AI-generated register replace a project manager? No. It can reduce clerical work and improve questions. It cannot negotiate ownership, resolve conflict, understand political context, or decide what the organization is willing to accept. Those are still human responsibilities. The takeaway Use AI to turn scattered project evidence into a dependency register people can inspect. Ask it to show its sources, label its guesses, and generate the questions your plan is avoiding. Then have the actual owners verify the rows and accept the consequences. A dependency map is not a prediction machine. It is a shared memory with enough structure to catch surprises earlier. That is already useful. You do not need to pretend the robot knows your architecture to get value from a better checklist. For more practical guidance on using AI without outsourcing your brain, Don't Replace Me is the longer field guide. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Contingency Plan Prompts: 10 Templates 10 copy-paste AI contingency plan prompts for ChatGPT and Claude. Map failures, assign owners, draft comms, and verify assumptions before a crisis hits. contingency planningAI promptsbackup plans AI Dependency Mapping Prompts: Find What Breaks AI dependency mapping prompts for finding upstream work, downstream risks, hidden handoffs, owners, schedule conflicts, and launch blockers. AI promptsdependency mappingproject management AI Assumption Log Prompts: Stop Planning on Guesswork Use 10 AI assumption log prompts to expose project guesswork, design evidence checks, assign review owners, and stop unverified beliefs becoming commitments. AI promptsproject managementassumption logs --- ## AI Documentation Prompts: 10 Templates That Work URL: https://dontreplace.me/articles/ai-documentation-prompts-10-templates-that-work ai documentation promptsdocumentation templatesknowledge managementai at workSOPs AI Documentation Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn messy notes, meeting records, and expert knowledge into useful documentation. Includes safety rules and a reusable formula. June 07, 2026 Updated June 07, 2026 Your team's documentation looks like this: a Google Doc titled "onboarding v3 FINAL (2)" last edited by someone who left eight months ago, a Slack thread with the actual process buried under seventeen reaction emojis, and a sticky note on someone's monitor that says "DON'T TOUCH THE BLUE BUTTON." AI documentation prompts won't fix your culture. But they'll get you from zero to a working first draft in twenty minutes instead of never. That's the actual value here: speed on the ugly first draft, so your subject-matter experts spend their time correcting and improving instead of staring at a blank page. Here's how to use AI for documentation without ending up with confident-sounding nonsense that gets someone fired. The core formula for every AI documentation prompt Every prompt that produces useful documentation has the same structure. Mess with it and you get slop. What to include in every prompt: Role: Tell the AI who you're writing for. "A new customer support rep on their first week" is not the same as "a senior engineer handling an incident." Source material: Paste your actual notes, transcript, or rough draft. Don't ask AI to invent the content. Give it something real to work with. Output format: Specify what you want. Step-by-step SOP, decision record, FAQ, checklist, whatever. Scope limits: Tell it what NOT to do. "Don't invent steps I haven't described. Flag anything unclear as [NEEDS REVIEW]." Verification flag: Ask it to mark anything it's uncertain about so your reviewer can catch it. One more rule before the prompts: don't paste confidential data into AI tools your company hasn't approved. No customer records, employee files, credentials, unreleased strategy, security incident details, proprietary code, or regulated information. Anonymize everything. If you're not sure whether a tool is approved, ask someone before you paste. Prompt 1: Turn messy notes into an SOP You have bullet points, a voice memo transcript, or a brain dump from the person who actually knows how this works. This prompt turns that into a structured standard operating procedure. You are a technical writer creating an SOP for [ROLE/AUDIENCE]. Here are my raw notes: [PASTE NOTES] Convert these into a step-by-step SOP using this format: - Purpose (one sentence) - Who this applies to - Prerequisites (what someone needs before starting) - Steps (numbered, one action per step) - What to do if something goes wrong - Owner and review date Do not invent steps that aren't in my notes. If anything is unclear or missing, write [NEEDS REVIEW: describe the gap] instead of guessing. Review every step with the person who actually does the job before publishing. AI is fast. It is not correct by default. Prompt 2: Interview an expert to capture tacit knowledge Some knowledge lives in someone's head and nowhere else. The problem is asking them to write it down produces either nothing or a document too vague to use. This prompt turns AI into an interviewer that extracts the details. I need to document [PROCESS/SKILL] and I'm going to answer your questions so you can help me write it up. Ask me one question at a time about how this process actually works. Focus on: - What triggers this process? - What decisions get made and why? - What are the common mistakes or edge cases? - What does "done right" look like versus "done wrong"? After I've answered all your questions, summarize what I've told you into a structured document for [AUDIENCE]. Do this over a live session with the expert, or send them the questions and have them answer in writing. Either way, the output captures real judgment instead of generic steps. For more templates like this, check out the AI training prompts guide for related approaches to capturing and sharing expertise. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Prompt 3: Rewrite a confusing wiki page The old doc exists. It's just incomprehensible. Maybe it was written by someone who knew too much, or someone who knew too little, or both at different times. Here is an existing documentation page that's confusing or outdated: [PASTE DOC] Rewrite it for [AUDIENCE] who needs to [SPECIFIC TASK]. Improvements to make: - Plain language, shorter sentences - Add a "What this is for" section at the top - Break walls of text into numbered steps or bullet points - Add a "Common mistakes" section if the content suggests any - Flag anything that looks outdated or contradictory with [NEEDS REVIEW] Do not change technical details. Only improve clarity and structure. Get a subject-matter expert to compare the rewrite against the original before you archive the old version. The AI might clean up the confusion by removing the detail that explained the exception. Prompt 4: Create a troubleshooting guide Support tickets cluster. The same five problems generate 80% of the volume. This prompt builds a troubleshooting guide from those patterns. I'm going to give you a list of common problems users/customers encounter with [PRODUCT/PROCESS]. For each problem, write a troubleshooting entry with: - Problem description (as the user would describe it) - Likely causes - Steps to diagnose - Steps to fix (if known) - When to escalate and to whom Here are the problems: [PASTE LIST OR SUPPORT TICKET SUMMARIES] Do not invent fixes you don't have evidence for. Write [SOLUTION UNKNOWN - ESCALATE] if you don't have enough information for a resolution. Have someone who actually resolves these issues verify each entry before it goes live. Especially any steps that touch production systems, customer data, or billing. Prompt 5: Convert meeting notes into a decision record Decisions made in meetings vanish. Six months later nobody remembers why you made that choice or what alternatives you considered. A decision record solves this and takes about three minutes with AI. Here are notes from a meeting where we made a decision: [PASTE NOTES] Convert these into a decision record with: - Decision made (one clear sentence) - Date - Who was involved - Context (why this decision was needed) - Options considered - Why we chose this option - Risks or tradeoffs accepted - Who owns implementation - When we'll review this decision Keep it factual. Only include what's in the notes. Flag anything missing as [NEEDS CONFIRMATION]. This one's low-risk because you're restructuring information you already have. The main error mode is AI inventing a rationale that sounds plausible but wasn't what was actually said. Read it carefully. Prompt 6: Write onboarding docs for a role The onboarding doc for most roles is either non-existent or a list of system logins. This prompt helps you build something a new person could actually use. Help me write onboarding documentation for a new [JOB TITLE] joining our team. I'll give you information about the role, and I need you to create: - First-week overview (what they'll do each day) - Key tools and systems (I'll fill in the specifics) - Who to go to for what (I'll fill in names and roles) - Common tasks with links or pointers to the relevant SOPs - Things that trip people up in the first 30 days - Where to ask questions Here's my input: [PASTE YOUR NOTES, EXISTING DOCS, OR BULLET POINTS] Use [PLACEHOLDER] wherever I need to fill in specifics. Do not invent team names, tool names, or processes. The AI gets you a solid skeleton. A current employee fills in the real specifics. Someone who recently joined the role reviews it for accuracy. Prompt 7: Document a recurring workflow Every team has processes that run on autopilot in one person's head. This gets them out. Help me document a recurring workflow for [TEAM/DEPARTMENT]. Workflow name: [NAME] Trigger: [What starts this workflow? Date, event, incoming request?] Frequency: [Daily, weekly, monthly, ad hoc?] Here are my notes on how it works: [PASTE NOTES] Output format: - Workflow overview (two sentences) - Trigger and frequency - Step-by-step process (numbered, with role responsible for each step) - Tools used - Outputs or deliverables - Common errors and how to avoid them - Handoff points (where does this go next?) Flag anything I haven't explained clearly as [NEEDS DETAIL]. This is one of the templates covered in Chapter 13 of Don't Replace Me, where Dee walks through the "Garbage In, Garbage Out" rule: the quality of your output is completely capped by the quality of your input. Vague notes make vague docs. Prompt 8: Create examples and non-examples Good documentation shows what "right" looks like. Great documentation also shows what "wrong" looks like. This is the part almost everyone skips, and it's the part that actually changes behavior. I'm writing documentation for [PROCESS/POLICY/STANDARD] and I need examples. Here's the guideline or rule: [PASTE RULE OR STANDARD] Create: - 3 examples of doing this correctly, with brief explanations of why they're correct - 3 examples of doing this incorrectly (non-examples), with brief explanations of what's wrong - 1 "edge case" example that's genuinely ambiguous, and note why it's tricky Make the examples realistic for [CONTEXT/INDUSTRY/ROLE]. Use fictional names and anonymized details. Do not invent examples that require specific technical knowledge I haven't given you. This prompt works especially well for style guides, writing standards, customer communication guidelines, and anything where "judgment" is the real skill being taught. Prompt 9: Find gaps in an existing document AI is better at spotting what's missing than most people reading their own docs, because it doesn't fill in the blanks automatically the way a subject-matter expert does. I'm going to give you an existing document and I need you to audit it for gaps and problems. Here's the document: [PASTE DOC] This document is intended for: [AUDIENCE] They will use it to: [SPECIFIC TASK OR GOAL] Please identify: - Steps that are described but not explained - Assumed knowledge the audience probably doesn't have - Missing steps (if the process seems incomplete) - Contradictions or anything that could be interpreted two different ways - Anything that looks outdated - Questions a reader would likely have that the doc doesn't answer Don't rewrite the document yet. Just list the gaps so I can decide what to fix. This is the cheapest QA step you can add to any documentation process. Run it before you publish, not after someone emails you confused. Prompt 10: Turn support questions into a FAQ Your support inbox is a documentation roadmap. Every repeated question is a gap in your existing docs. This prompt closes those gaps fast. Here is a list of questions we frequently receive from [CUSTOMERS/USERS/EMPLOYEES] about [TOPIC]: [PASTE QUESTIONS] Convert these into a FAQ section with: - The question phrased clearly (as the person asking it would phrase it) - A direct answer in plain language - Any links or resources to include (I'll fill these in) - A note if the answer varies by situation Organize by theme if there are clear groupings. Do not invent answers. If I haven't given you enough information to answer a question, write [ANSWER NEEDED: describe what's missing]. For AI tools your organization uses to communicate policy through documentation, also check the AI policy prompts guide for templates specifically built around setting and communicating rules. What AI should never do in documentation This matters. Rule #5 in the book: AI is fast, not smart. Speed is not accuracy. Never let AI: Invent API endpoints, system commands, or technical steps you haven't verified Write legal, compliance, or regulatory claims Create customer-facing promises about product behavior it doesn't know Fill in security procedures or credentials Produce anything that goes to production without a subject-matter expert sign-off If your AI documentation process doesn't include a human review step, it's not a documentation process. It's a liability generator dressed up as productivity. For a broader look at what AI is actually good at versus where it confidently makes things up, the plain-language AI explainer is a useful 5-minute read before you go all-in on AI-assisted workflows. Frequently asked questions What's the best way to use AI for writing documentation? Give AI real source material to work with, not a blank slate. Paste your notes, transcript, or rough draft and ask it to restructure and clarify. The best use is turning a brain dump from a subject-matter expert into a readable first draft, then having that expert verify accuracy before publishing. Can AI create documentation from scratch without any input? Technically yes, practically no. AI can generate plausible-sounding documentation for a process it knows nothing about, which is exactly the problem. The output will be generic at best and dangerously wrong at worst. Always start with real notes, real questions, or a real rough draft. Should I let AI write SOPs for technical or safety-critical processes? No, not without significant human review. AI can help structure and clarify an SOP, but every technical step needs to be verified by someone who actually does the job. For anything involving security, safety, compliance, or production systems, treat AI-generated content as an unreviewed draft until a qualified person signs off. How do I make sure AI documentation prompts don't expose confidential information? Only paste information into AI tools your organization has approved for that type of data. Anonymize customer names, employee details, and proprietary specifics before pasting. When in doubt, check with your security or legal team. The AI prompts in this article are designed to work with anonymized or internal-only content. See your workplace AI policy for specifics. What types of documentation is AI actually good at helping with? AI is strongest at restructuring and clarifying, not inventing. It does well with: converting notes into structured formats, rewriting confusing prose into plain language, generating examples and non-examples, spotting gaps in existing docs, and turning support questions into FAQs. It's weakest anywhere accuracy depends on specific technical knowledge it doesn't have. How do I get better outputs from AI documentation prompts? Be specific about your audience, the purpose of the document, and what format you want. Include a "don't invent things" instruction in every prompt. Ask it to flag uncertainty with [NEEDS REVIEW] rather than filling gaps with guesses. And always feed it real source material, not a vague description of what the document should cover. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Automation Prompts: Automate the Boring Stuff 10 copy-paste AI automation prompts for mapping workflows, writing SOPs, adding approval gates, and QA testing. Practical templates with safety rules included. ai automation promptsworkflow automationSOP templates AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts --- ## AI Email Prompts: 12 Templates That Actually Work URL: https://dontreplace.me/articles/ai-email-prompts-write-better-work-emails-fast ai email promptschatgpt for emailwork email tipsai at workemail writing AI Email Prompts: 12 Templates That Actually Work 12 copy-paste AI email prompts for ChatGPT and Claude. Fix tone, shorten drafts, follow up without sounding desperate. Real templates for real work emails. May 24, 2026 Updated May 24, 2026 Your inbox has 47 unread emails and you're staring at a reply you've rewritten three times. It still sounds either passive-aggressive or like you're begging. This is exactly where AI email prompts earn their keep. Not by writing your emails for you. By getting you out of the rewrite loop faster. Most people use AI email tools completely wrong. They type "write a professional email" and paste whatever comes out. Then they wonder why it sounds like a terms-of-service agreement wrote back. The fix isn't a better tool. It's a better prompt. Why most AI email prompts fail The problem isn't ChatGPT. It's the same problem you'd have if you hired an intern and said "write a good email for me" while walking away. You'd get something technically correct and completely useless. AI is fast, pattern-matching software. It's genuinely good at language-shaped work: drafts, rewrites, summaries, tone adjustments. What it's not good at is guessing what you actually need when you give it nothing to work with. The "garbage in, garbage out" rule is brutal here. Feed it a vague request and you get a vague email. Feed it role, context, the specific problem, and what outcome you want, and you get something worth editing. The goal isn't to eliminate your judgment. It's to give you a decent draft in 30 seconds instead of staring at a blank screen for 20 minutes. A lot of the anxiety around AI at work comes from misunderstanding what it actually is. It's not a mind reader. It's not a replacement for knowing what you want to say. If you've ever read about what AI can and can't do, you'll know that the gap between "writes fluent sentences" and "knows your situation" is enormous. Prompts are how you close that gap. The one prompt formula that works for almost everything Before the 12 copy-paste templates, here's the structure behind all of them. Every good AI email prompt has five parts: Role: Tell it who's writing and to whom. "I'm a project manager writing to a client." Context: One sentence on the situation. "They missed a deadline and we need to reschedule." Task: What you want the email to do. "Write a brief, professional message asking to reset the timeline." Constraints: Tone, length, anything to avoid. "Keep it under 100 words. Don't sound accusatory." Format: What you want back. "Give me two options." You don't always need all five. But the more you include, the less editing you'll do afterward. Think of it like briefing a smart intern who's fast, tireless, and occasionally confident about things they shouldn't be. You still have to supervise the output. The format instruction is one most people skip. Asking for two versions gives you something to choose between rather than just accept or reject. Asking for bullet points first before the full email helps you check the logic before you check the language. Small adjustments to how you ask change the usefulness of what comes back by a lot. What makes a bad email prompt vs. a good one Let's be specific about this because the difference matters. Bad prompt: "Write a professional email following up on my proposal." What you get: A generic, lifeless message that could have been written by anyone, for any proposal, to anyone. It'll start with "I hope this email finds you well" and end with "please don't hesitate to reach out." Good prompt: "I'm a freelance designer. I sent a brand refresh proposal to a mid-size retail client two weeks ago. They haven't responded. Write a short follow-up email that assumes they're busy, not ignoring me, and makes it easy to reply. Under 80 words. No apology for following up." What you get: Something close to usable on the first try. The difference is specificity. Context about who you are, who they are, the history, and what you want the email to accomplish. That's what separates useful AI output from the kind you delete and rewrite manually anyway. It takes an extra 30 seconds to write a better prompt. That 30 seconds saves you five minutes of editing, or worse, five minutes of trying to figure out why the draft feels wrong. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → A safety reminder before you paste anything This matters and you should read it before using any of the prompts below. Do not paste confidential information into ChatGPT, Claude, or any AI tool unless your company's policy explicitly allows it. That means no customer PII, no HR conversations, no legal or medical details, no private employee information, no financial figures tied to real names. Replace names with "[Client Name]" and strip out specific numbers if you're not sure. Once the draft comes back, read every word. Verify any facts, dates, links, attachments, or promises the AI included. It will occasionally invent details with complete confidence. Never send AI output without reading it. That's how you end up cc'ing the wrong person on a message that describes them. If you want the broader thinking on using AI at work without creating problems for yourself, the no-BS starter guide covers it. 12 copy-paste AI email prompts that actually work These are designed for ChatGPT or Claude. Paste the prompt, fill in the brackets, read the output, then edit. Don't send anything without reading it first. 1. Make a messy draft clear I've written a draft email but it's unclear and jumbled. Clean it up so it's easy to read, keeps the key points, and uses plain language. Don't change the tone much. Here's the draft: [paste draft] 2. Shorten a too-long email This email is too long. Cut it to under [100/150/200] words while keeping the main message and the ask. Remove anything that's just filler or context the reader doesn't need. Here's the email: [paste email] 3. Make a message warmer This email is technically fine but sounds cold. Rewrite it so it sounds like it came from a real person, not a policy document. Keep it professional but human. Same length. Here's the original: [paste email] 4. Make a message firmer I need to send this email but it sounds like I'm apologizing for asking. Rewrite it so it's direct and confident, not aggressive. The ask should be clear. Here's the draft: [paste draft] 5. De-escalate a tense thread This email thread has gotten tense. I need to reply in a way that addresses the issue without escalating it. Keep the tone calm and solution-focused. My goal is [state your actual goal]. Here's the thread: [paste relevant parts, with names replaced] 6. Ask for a decision I need to ask [their role] to make a decision about [topic]. They've been slow to respond. Write a concise email that explains the situation in two sentences, clearly states what I need them to decide, and gives a deadline of [date]. Keep it under 100 words. Respectful but direct. 7. Follow up without sounding needy I sent an email [X days] ago about [topic] and haven't heard back. Write a short follow-up that's polite, references the original, and makes it easy for them to respond. Don't make me sound desperate. Under 60 words. 8. Summarize a long email thread Summarize this email thread in 3-5 bullet points. Include: what was decided, what's still open, and any action items with names attached. Here's the thread: [paste thread, names replaced if needed] 9. Write a client recap email Write a short email recapping what was discussed in our meeting with [client role] on [date]. Include: what we agreed to do, any open questions, and next steps with rough timelines. Tone should be professional and organized. Here are my notes: [paste notes] 10. Turn bullet points into an email Turn these bullet points into a well-written email. Keep it clear and concise. The recipient is [their role]. The purpose is [state purpose]. Tone should be [professional/friendly/formal]. Here are the bullets: [paste bullets] 11. Rewrite jargon into plain English This email is full of jargon and buzzwords. Rewrite it in plain English without losing the meaning. The reader is [describe their background, e. g., "not a technical person"]. Keep the same length. Here's the original: [paste email] 12. Check an email before sending Read this email and tell me: Does it clearly state what I need? Does anything sound rude, confusing, or ambiguous? Is there anything missing (attachments mentioned, questions unanswered, unclear deadlines)? Here's the email: [paste email] That last one is underused. Running your email through a pre-send check takes 15 seconds and has saved at least a few people from career-limiting moments. For more templates across different work situations, the 12 prompts for work collection has formats for reports, decisions, and presentations too. How to adapt these prompts when they don't quite work The first output isn't always right. That's fine. The prompts above are starting points, not guarantees. Here's what to do when the draft comes back wrong. If the tone is off, don't rewrite from scratch. Add a sentence to your prompt: "That's close but sounds too formal. Rewrite it to be warmer, like I'm talking to someone I know." Or: "Less formal. I want this to sound more like a message from a person, less like a memo." You can tune the output without starting over. If it's too long, paste it back and say "Cut this by a third. Keep the ask and the deadline. Cut everything else." It'll do it in seconds. If it invents details you didn't give it, just delete them. Don't try to fact-check AI hallucinations by asking AI to check itself. Read the output yourself and remove anything you didn't put in. If you asked for two versions and both are bad, the problem is usually the prompt. Go back, add more context, and try again. The fix is almost always more specificity, not a different tool. One pattern that works well: ask for the email structure first, then ask for the full draft. "Before you write the email, list the three things it needs to accomplish." If the list is right, tell it to write the email. If the list is wrong, correct it there. You're checking the logic before you check the language, which saves time. What to do with the output You asked for a draft. You got one. Now the actual work begins, and it takes about two minutes. Read every sentence. Does it sound like you? If not, change it. Does it say what you actually mean? If not, fix it. Did it invent a detail, a date, or a promise you didn't make? Delete it. AI doesn't know your relationship with this person. It doesn't know the history, the politics, the unspoken thing from the last meeting. You do. That's the part that actually matters. The prompts above handle the scaffolding. You handle the judgment. This is the distinction that matters. AI handles the language work. You handle the thinking. The people who get replaced aren't the ones who use these tools. They're the ones who either refuse to use them at all, or use them without reading the output. Both are equally avoidable. Dee covers the full thinking behind this in Don't Replace Me, specifically the "smart intern" framing: AI is like a new hire who's fast, has read everything, and occasionally makes things up. You wouldn't send an intern's work to a client without reading it. Same rule applies here. The emails where you shouldn't use AI prompts at all This doesn't come up in most articles about AI email tools, but it should. There are emails where using a template is the wrong move, not because AI can't write them, but because the relationship deserves better. When someone close to you is going through something hard and they've reached out. When you're genuinely apologizing for something serious, not a missed deadline, but something that matters. When you're having a real disagreement with someone you work with every day and the relationship is on the line. AI can write something that sounds sympathetic. It can't be sympathetic. The difference is small on the screen and huge on the receiving end. People who know you can usually tell when something sounds like you and when it doesn't. Use AI for the operational stuff. The follow-ups, the recaps, the status updates, the "can you please just decide" emails. Write the human stuff yourself. If you're not sure which category an email falls into, it's probably the second one. There's also a practical reason: AI-written empathy tends to land wrong. It hits all the keywords but misses the specifics. It says the thing without knowing the thing. You end up with something that reads like a condolence card from a company you don't have a relationship with. Write those ones yourself. Should you tell anyone you're doing this? Short version: probably yes, in most situations. The longer answer depends on your company, your role, and what kind of email you're writing. The full breakdown of whether to disclose AI use to your boss is worth reading if this is on your mind. The 68% of workers who hide their AI use aren't necessarily doing something wrong. But it's worth knowing where your company actually stands before you're in an awkward conversation about it. What you're doing with these prompts isn't fundamentally different from asking a colleague to read a draft before you send it. You're getting a second pass on your own thinking. The output is yours. You read it, edited it, decided it was right, and sent it. That's how writing has always worked. One tool is enough There are about 40 AI email assistants on the market right now with names like "Mailmuse" and "InboxGenie" and similar. Most of them are wrappers around the same models you already have access to for free or at $20/month. You don't need any of them. ChatGPT or Claude and the prompts above will handle everything in this article. The guide to using ChatGPT specifically at work goes deeper if you want more context on how to get the most out of one tool instead of chasing 10. Stop collecting tools. Start using the one you have. Frequently asked questions What's the best AI prompt for writing a professional email? Start with role, context, task, constraints, and desired format. Something like: "I'm a [role] writing to [recipient role] about [situation]. Write a [tone] email that [does X]. Keep it under [word count]." The more specific you are, the less editing you'll have to do afterward. Vague prompts produce vague emails. Is it safe to paste work emails into ChatGPT? It depends on your company's policy, and you should check it. Don't paste anything with customer PII, HR details, legal content, medical information, or financial data tied to real names. Replace specific names and numbers with placeholders when in doubt. When it comes to confidential or sensitive threads, use your own judgment and write those yourself. Will people be able to tell my email was written by AI? If you paste the output without editing, yes, sometimes. AI has patterns: it over-explains, adds phrases like "I hope this finds you well," and can sound weirdly formal. The fix is to read the draft, cut the filler, and make it sound like you. Use AI for the structure, then rewrite the parts that don't sound human. What's the difference between AI writing my emails and AI helping me write emails? The difference is your judgment on the output. AI writing your emails means you paste whatever comes out and hit send. AI helping means you use it to get a draft faster, then read it, fix what's wrong, and send something you've actually approved. The second version is what these prompts are designed for. Do I need a special AI email tool or will ChatGPT work fine? ChatGPT or Claude works fine. Most dedicated AI email tools are just wrappers around the same underlying models. Save the money, learn to write better prompts, and use one general tool well. That's it. How do I follow up on an email without sounding desperate? Keep it short, reference the original, make it easy to respond, and skip the apology. "Following up on my email from Tuesday about [topic]. Happy to answer any questions, or let me know if [decision] isn't the right path." Under 60 words. Direct. Done. The prompt in section four above handles this with one paste. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Automation Prompts: Automate the Boring Stuff 10 copy-paste AI automation prompts for mapping workflows, writing SOPs, adding approval gates, and QA testing. Practical templates with safety rules included. ai automation promptsworkflow automationSOP templates AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts --- ## AI Escalation Plan Prompts: Stop Letting Problems Rot URL: https://dontreplace.me/articles/ai-escalation-plan-prompts-stop-letting-problems-rot ai escalation planescalation promptsai at workescalation email templateworkplace escalation AI Escalation Plan Prompts: Stop Letting Problems Rot 10 copy-paste AI escalation plan prompts for managers and teams. Turn messy notes into clear briefs, draft stakeholder updates, and stop problems from festering. July 05, 2026 Updated July 05, 2026 There's a problem sitting somewhere in your organization right now. Someone knows about it. They wrote a Slack message that trailed off. They added it to a doc nobody opened. They told themselves they'd escalate "once they had more information." That was three weeks ago. AI escalation plan prompts won't fix avoidance. But if you're actually ready to escalate and just can't get the damn thing written, they'll save you an hour of staring at a blank doc. That's what this article is. Ten copy-paste prompts for the most common escalation scenarios, a reusable formula, and a short list of things you should never paste into an AI tool. No theory. No $997 framework. Just prompts that work. What a good AI escalation prompt actually does The job of an escalation plan isn't to make things sound calm. It's to make the problem legible: what happened, when, what it affects, who owns it, what options exist, and what decision you're asking for. AI is genuinely useful for that. It's fast at organizing messy notes into a coherent structure. It'll suggest stakeholders you forgot. It'll draft three versions of a message for three different audiences in two minutes. It won't panic, it won't soften things to protect feelings, and it doesn't care that you're the one who missed the signal three weeks ago. What it can't do: verify facts, judge severity accurately, decide who politically needs to be looped in first, or know when a polite update has quietly become dangerous avoidance. That judgment is yours. The prompts below assume you're supplying the real situation. Garbage in, polished corporate fog out. The reusable escalation prompt formula Before the 10 prompts, here's the underlying structure. Every good escalation prompt you write should hand AI: The situation (what happened, when, with dates) The impact (customers, revenue, people, operations, legal exposure) What's already been tried Open questions and assumptions (what you know vs. what you're inferring) The audience (who's receiving this escalation) The ask (decision, resources, escalation path, or just awareness) A prompt that includes all six of these will produce something usable. A prompt that says "help me write an escalation plan" will produce something that sounds like an escalation plan but says nothing. AI escalation plan prompts: the 10 templates Prompt 1: Should this even be escalated? Sometimes the first problem is knowing whether you're overreacting. I have a situation I'm not sure warrants escalation. Help me think through it. Situation: [describe the issue in plain language] Current status: [what's happening right now] Who knows about it: [list names/roles] What I've tried: [what steps have already been taken] Timeline: [when it started, key events with dates] Potential impact if it gets worse: [your honest assessment] Ask me 5 clarifying questions that will help me decide whether this needs to be escalated now, monitored for 48 hours, or handled at my level without escalation. Prompt 2: Turn messy notes into an escalation brief You've got a trail of Slack messages, a half-written doc, and a forwarded email from a panicking client. You need a structured brief in 15 minutes. I have messy notes about an issue I need to escalate. Help me turn them into a structured escalation brief. Raw notes: [paste your notes here, with any relevant dates, names, and events] Audience: [for example, my VP, a cross-functional leadership team] Time available: [how long until I need to send this] Format the output as: - Issue summary (3 sentences max) - Timeline of key events - Current status - Impact (confirmed vs. assumed, flag which is which) - What's been tried - Open questions - Options - Decision ask For longer, more complex situations, pair this with AI status report prompts to handle the ongoing updates after the initial escalation lands. Prompt 3: Escalate to your manager The most common scenario. You've hit a wall, a risk is growing, and you need your manager involved without making it sound like you've lost control. Help me write a manager escalation message for the following situation. Issue: [what the problem is] When it started: [date] What I've done: [steps you've taken] Where I'm blocked: [specific obstacle] What I need from my manager: [decision, resource, connection, awareness] Tone: direct and calm, not panicked Keep it under 200 words. Bullet points for the key facts. One clear ask at the end. Prompt 4: Customer-impacting issue When customers are affected, the escalation needs to be faster and more precise. It also needs to separate what you've confirmed from what you're still investigating. I need to escalate a customer-impacting issue to [role, for example, head of customer success / VP of operations]. What happened: [describe the issue] Customers affected: [number, segment, or named accounts, anonymize or aggregate as needed] When it started: [date and time if known] Current status: [is it ongoing, resolved, partially resolved?] Business impact: [revenue risk, SLA breach, contractual exposure, reputational risk] What we know for certain: [confirmed facts only] What we're still investigating: [open questions] Who's working on it: [named owners and their current actions] Decision needed: [what you need the escalation recipient to decide or approve] Write a concise escalation update in plain language. Flag clearly what is confirmed vs. assumed. If the issue involves client communication, AI client communication prompts has templates for the external messages you'll need in parallel. Prompt 5: Blocked project The project isn't burning. It's just slowly strangling because nobody will make a call. A project I'm managing is blocked. I need to escalate this to get a decision made. Project: [name and brief description] What we're trying to deliver: [goal] The block: [describe exactly what's preventing progress] Who owns the block: [person, team, or dependency] How long it's been blocked: [duration, with start date] Impact of continued delay: [deadline risk, downstream effects, cost] What I've already tried to resolve it: [steps taken] What I need: [a decision, a resource, a conversation between specific people] Write a 150-word escalation note and a separate one-paragraph executive summary of the same issue. Prompt 6: Executive update on a live escalation You're already mid-escalation. Leadership wants a crisp update. You have 24 Slack notifications and a spreadsheet from 2 different people that contradict each other. I need to write an executive update on an active escalation. Issue: [original problem] What's happened since the last update: [key developments, with dates] Current status: [where things stand right now] What's resolved: [anything you can close out] What's still open: [active risks and unknowns] Owner for each open item: [name and role] Estimated resolution: [date or honest confidence interval] Decision still needed from leadership: [if any] Format as a 4-bullet status block followed by one clear decision or action ask. Flag any assumptions with "(assumed, not confirmed)". Prompt 7: Vendor problem The vendor is late, wrong, or both. You need to escalate internally and possibly to the vendor's leadership. I need to escalate a vendor performance issue. Vendor: [name, only if this is going into an internal system; anonymize for external AI tools if the vendor relationship is sensitive] What they were supposed to deliver: [scope and deadline] What actually happened: [the gap] Attempts to resolve at working level: [conversations, emails, tickets, with dates] Business impact: [what's affected by their failure] Contract or SLA terms relevant here: [briefly, not the full contract, do not paste contract text] Help me write: 1. An internal escalation note to my procurement / vendor management lead 2. A professional escalation email to the vendor's account manager or leadership contact Prompt 8: Separate facts from assumptions This one's not a message generator. It's a thinking tool. Use it before you write anything. Help me separate confirmed facts from assumptions in this escalation situation. Everything I know or think I know: [paste your raw understanding of the situation] Sort this into three columns: 1. Confirmed (source or evidence exists) 2. Assumed (plausible but unverified) 3. Unknown (we don't have this yet) Then flag which assumptions would change the severity rating if they turned out to be wrong. This prompt is worth running before any high-stakes escalation. It forces you to notice where you're confusing "probably true" with "definitely true." For deeper risk framing, AI risk assessment prompts has a full set of templates for the same discipline. Prompt 9: Calm Slack or email escalation Not everything warrants a brief. Sometimes you just need to escalate a specific thing, in one message, without starting a fire. Help me write a calm, professional Slack message (or short email) escalating the following issue. Issue: [what needs to be escalated] Recipient: [their role and relationship to you] What I want them to do: [specific action, decision, or awareness] Tone I want: [direct but not alarming / urgent but not panicked / matter-of-fact] Max length: [Slack: 100 words / email: 150 words] Avoid drama. One clear ask. End with a specific time by which I need a response. Prompt 10: Post-escalation follow-up plan The escalation happened. Now you need to track it through to resolution without losing it in a backlog. An escalation I raised is now in motion. Help me build a simple follow-up plan. What was escalated: [brief summary] Who's now responsible for what: [named owners and commitments they made] Decisions that were made: [record them explicitly] What's still open: [unresolved items] Next check-in date: [when you'll review status] Create: 1. A decision log entry I can save 2. A follow-up calendar reminder template with a one-sentence context note 3. A list of open questions to resolve before the next update For keeping that decision log clean over time, AI decision log prompts has the full toolkit. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What not to paste into AI tools when escalating This matters. When you're stressed and in a hurry, it's easy to dump everything into the prompt. Don't do that with: Customer PII (names, emails, account numbers, contact records) Employee performance records or HR complaints Security incidents or vulnerability details Legal disputes, lawyer communications, contracts Financial forecasts or unreleased results Unreleased product roadmaps or M&A activity Board materials or investor communications Credentials, tokens, system access information Confidential client strategy Use your company's approved AI tools if you need to handle any of the above. If your company hasn't told you which tools are approved, that's worth escalating on its own. AI escalation plan prompts won't replace the hard part The hard part isn't writing the brief. It's deciding that the problem is real, that waiting is now a choice with consequences, and that you're the person who needs to move it. Dee covers exactly this dynamic in Don't Replace Me, specifically the gap between using AI to produce something that looks like ownership and actually owning the outcome. A polished escalation doc written with AI is still just a document. Someone has to read it, believe it, act on it, and follow through. That's you. AI helps you get the message right. The judgment about whether to send it, when, to whom, and at what level of urgency? Still entirely yours. If the escalation involves legal exposure, a security incident, HR matters, financial risk, customer safety, or anything regulated, loop in the relevant function directly. Don't let a well-formatted AI document stand in for a conversation with legal, security, or HR. For the broader picture of using AI at work without outsourcing your judgment, the no-BS AI at work guide is the right starting point. Frequently asked questions What should an AI escalation plan include? A good escalation plan includes a concise issue summary, a timeline of key events with dates, confirmed impact (separated from assumptions), named owners, options that have been considered, and a specific decision ask. The prompts above give you a fill-in-the-blank structure for each of these. The key thing AI can't add is your actual knowledge of the situation, so bring real details. Can I use ChatGPT or Claude to write an escalation email? Yes, they're both good at structuring messy information into clear messages. Use Prompt 3 or Prompt 9 from this article. The risk is pasting sensitive information into tools that aren't approved by your organization. For anything involving customer data, HR issues, legal matters, or security incidents, use your company's approved AI environment or write it without AI assistance. How do I know when something needs to be escalated vs. handled at my level? Use Prompt 1 from this article. The general rule: escalate when you're blocked and the block will have consequences, when the risk is growing faster than you can manage it, or when a decision is needed that you don't have authority to make. If you're not sure, err toward escalation with a clear framing of what you're asking for. What's the difference between an escalation and a status update? A status update says what's happening. An escalation says what's happening, why it's a problem, what decision or resource is needed, and who needs to make that call. If you're writing a status update that has no ask, it might be an update. If there's a risk you can't resolve at your level, it's an escalation. AI status report prompts handles the former. How do I escalate without sounding like I've lost control? Be direct about what you've tried before bringing it up. Phrase the ask as "I need a decision on X" or "I need your help unblocking Y," not "I don't know what to do." One specific ask, not a list of worries. Prompt 3 and Prompt 9 are built around this. Should I use AI for security or HR escalations? No, not in unapproved tools. Security incidents need to go through your incident response process. HR matters involve employee records and legal exposure. Financial or legal escalations should involve the relevant function directly. AI can help you draft a message, but only in an environment your organization has cleared for that kind of data. When in doubt, write it yourself or call the relevant function. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Automation Prompts: Automate the Boring Stuff 10 copy-paste AI automation prompts for mapping workflows, writing SOPs, adding approval gates, and QA testing. Practical templates with safety rules included. ai automation promptsworkflow automationSOP templates AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts --- ## AI Executive Summary Prompts: 10 Templates URL: https://dontreplace.me/articles/ai-executive-summary-prompts-10-templates-that-work ai promptsexecutive summaryworkplace aiproductivityleadership communication AI Executive Summary Prompts: 10 Templates 10 copy-paste AI executive summary prompts for project updates, board briefs, risk summaries, and more. Includes the 5-part prompt formula. July 02, 2026 Updated July 02, 2026 Your project update is 900 words. Your exec has 90 seconds. That gap is your problem, and AI executive summary prompts are one practical way to close it. The bad news: most people use AI for this wrong. They paste in a wall of meeting notes, ask for "an executive summary," and ship whatever comes back. Then they wonder why leadership keeps asking follow-up questions that were already answered in the original doc. The summary was polished. It just wasn't accurate. There's a right way to do this, and it starts with giving the AI something real to work with. What AI actually does when it summarizes AI compresses. That's it. It finds patterns, trims repetition, and restructures sentences for a different audience. Rule #5 in Don't Replace Me puts it plainly: it's not smart, it's fast. Compression is not judgment. The AI doesn't know what your CFO actually cares about, what political landmine is sitting in paragraph four, or which number will immediately raise a red flag in a board meeting. You do. That's the division of labor that works. You bring context, stakes, and taste. The AI does the first-pass restructuring so you're not staring at a blank doc at 11pm. What it's useful for: first-pass structure, audience translation, decision framing, risk callouts, explicit asks, follow-up checklists. What it's risky for: inventing confidence where there isn't any, hiding caveats to sound cleaner, and making political calls it has zero context for. Know the difference before you hit send. The reusable AI executive summary prompt formula Before the 10 templates, here's the skeleton. Every good executive summary prompt has these five parts: 1. Role and audience. "You are helping a [role] write a summary for [audience]." 2. The raw material. Paste the actual notes, draft, or data. The cleaner the input, the cleaner the output. Garbage in, confident sludge out. 3. Format requirements. Length, section headers, bullet points vs. prose, what goes first. 4. What to include. Key decisions, risks, asks, owners, dates, assumptions. 5. What to flag. Gaps, missing info, things that need human review before this goes anywhere. Run any of the 10 prompts below through that frame and they'll work better. 10 copy-paste AI executive summary prompts 1. Summarize a long project update I need to turn this project update into a 3-paragraph executive summary for senior leadership. Focus on: current status, key risks or blockers, and the one decision they need to make this week. Flag anything that looks vague or needs a specific number I should add before sending. [Paste project update here] Pair this with the AI project management prompts guide if you're doing this regularly. 2. Turn meeting notes into an exec brief These are raw meeting notes. Turn them into a 200-word executive brief. Structure it as: what was decided, what's still open, who owns what, and by when. If anything is ambiguous about ownership or deadline, flag it rather than guess. [Paste meeting notes here] For fuller meeting documentation, see the AI meeting notes prompts collection. 3. Compress research into key takeaways I have research on [topic]. Write a 5-bullet executive summary for [audience]. Each bullet should state the finding, its relevance to our situation, and confidence level (high/medium/low). If a finding comes from a single source, note that. Don't invent statistics. [Paste research here] 4. Write a board-style one-pager Turn this into a one-page board brief. Use four sections: Situation (2-3 sentences), Options Considered (bullet list), Recommendation (1 clear sentence), and Next Steps (3 bullets with named owners and dates). Keep it factual. Flag any section where you'd need me to add specific numbers or approvals. [Paste background material here] 5. Summarize risks and blockers From the information below, extract a risk summary for leadership. List each risk with: description, likelihood (high/medium/low), impact if unaddressed, current mitigation in place, and who owns resolution. Don't soften risks to sound better. [Paste project or status info here] For deeper work on this, the AI risk assessment prompts guide goes further. 6. Convert a status report into leadership language This is a technical/operational status report. Rewrite it for a non-technical leadership audience. Remove jargon. Keep all dates, percentages, and named owners. Lead with the so-what, not the what. If something looks like it could turn into a problem, surface it clearly rather than burying it. [Paste status report here] The AI status report prompts collection has complementary templates for the report itself. 7. Create a client-ready summary I need a client-facing version of this internal update. Tone should be confident but honest. Remove internal jargon, internal team names, and anything marked as confidential. Include: current status, what we've delivered, what's coming next, and any decisions we need from the client this week. 150-200 words maximum. [Paste internal update here] 8. Rewrite a rambling draft This draft summary is too long and unclear. Rewrite it so it: opens with the key decision or finding, supports it with the three most important facts, and closes with a specific ask or next step. Cut anything that's background context a senior leader doesn't need. Target 150 words. [Paste draft here] 9. Surface missing context before sending Before I send this executive summary, review it and tell me: what information is assumed but not stated, what questions a senior leader is likely to ask that aren't answered, what numbers or claims look vague or unsupported, and what I should verify before this goes to leadership. [Paste draft summary here] This one's underused and worth running before anything goes to a C-suite audience. Think of it as a pre-send checklist you don't have to write yourself. 10. Summary plus recommended next steps Based on this update, write a 200-word executive summary followed by a "Recommended Next Steps" section with 3-5 actions. Each action should have: the action, who owns it, a suggested deadline, and what decision or dependency it relies on. Flag any action where ownership isn't clear from the information I've given you. [Paste update or notes here] For decisions that need to be logged and trackable, the AI decision log prompts guide connects well here. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → How to adapt these prompts for different audiences The prompts above work as written. But a summary for your CFO and a summary for your board chair are not the same document, even if the underlying facts are identical. The audience shapes everything: what goes first, what gets cut, how much risk you surface, and how many numbers you include. Here's a quick translation guide: AudienceLead withRisk framingIdeal lengthWhat to cut CEOThe decisionHigh-level, strategic150-200 wordsOperational detail CFOThe numberFinancial exposure200-250 wordsNarrative context COOThe blockerOperational impact200-300 wordsStrategic preamble BoardThe recommendationReputational and financialOne page maxEverything tactical ClientWhat we deliveredFramed as mitigation150 wordsInternal team detail RegulatorThe compliance statusExplicit and unhedgedVariesInformal language Add one line to any prompt above to activate this: "The audience is [X]. Adjust the framing, lead, and risk language accordingly." That single instruction changes the output more than almost anything else you can do. A summary that reaches the wrong audience in the wrong frame is worse than no summary. The CFO who gets a CEO-style narrative summary asks five follow-up questions. The CEO who gets the CFO's spreadsheet-forward version glazes over. Know who's reading it before you paste anything in. AI executive summary prompts only work with clean inputs This is Rule #13 in action. Vague notes create confident sludge. If you paste in "we had some issues with the timeline and we're working on it," the AI will rewrite that into fluent ambiguity. It won't manufacture a date, a name, or a cause. It'll just make the vagueness sound professional. Before you run any of these prompts, do a 60-second check on your input: Do you have actual numbers, not approximations? Is every open item assigned to a specific person? Are there real dates, not "soon" or "next quarter"? Have you noted which information is assumption vs. confirmed fact? If the answer to any of those is no, fix the input first. The AI will structure whatever you give it. It won't fix it. For the research-heavy summaries especially, make sure any statistics you include came from a source you can name. The AI research prompts guide has a good framework for checking this before anything goes upstream. What not to paste into AI tools This one matters. A lot of people are using consumer AI tools (ChatGPT, Claude, Gemini) for work summaries without thinking about what's in the document they're pasting. Don't paste these into unapproved AI tools: Customer PII or account data Employee records, performance reviews, or HR information Financial forecasts or unreleased earnings data Legal disputes, contracts, or privileged communications Security incidents, vulnerabilities, or credentials Board materials or pre-approval strategy documents Confidential client information Unreleased product details If your company has an approved AI tool with a data processing agreement, use that for sensitive material. If you're not sure what's approved, ask IT or legal before you paste. A polished summary isn't worth a data breach or a compliance violation. When the summary involves legal, security, HR, financial, safety, or customer-impacting decisions, human review isn't optional. It's the job. Common mistakes that make AI summaries land wrong Even people using these prompts correctly run into the same handful of problems. Here's what goes wrong most often and how to fix it. Leading with background instead of the decision. AI defaults to chronological order because that's how most source material is structured. Meeting notes start with who attended. Project updates start with what happened last week. An executive summary should start with what needs to happen next. If the output leads with history, add this line to your prompt: "Open with the key decision or recommendation, not the background." Burying the ask. A summary without a clear ask is just a report. Leadership reads it, nods, and moves on. Nothing gets decided. Add an explicit section to your prompt: "End with one specific ask: what you need from the reader, by when." If you don't know what you're asking for, figure that out before you write the summary at all. Too many bullet points. AI loves bullets. Leadership audiences vary. Some executives read dense prose. Some want a single paragraph. Some want three bullets and a recommendation. Know your reader, then add format instructions to the prompt. "This executive reads prose, not bullets" is a legitimate and useful instruction. Softened risks. AI has a tendency to hedge bad news. "There are some challenges with the timeline" is not the same as "we are eight days behind and at risk of missing the client deadline." Add this line to any prompt where risk matters: "Do not soften risks. State them directly." Missing owners. A summary that says "the team will resolve this" is worth nothing. If you want your summary to drive action, every open item needs a named person attached to it. Check your input for named ownership before you paste. If it's not in the source material, add it yourself. The thing AI can't replace in an executive summary Here's the honest version. AI can get you to a clean structure faster. It cannot replace taste, and taste is the actual skill that makes someone a trusted executive communicator. Taste is knowing that the third bullet will land wrong in this room. Taste is knowing your CFO responds to specificity and your COO responds to risk framing. Taste is deciding what to omit because including it will raise questions you can't answer yet. Taste is the sentence you add that wasn't in the source material because you understand the context and the AI doesn't. That's the moat. The prompts above are useful. But the person who reads the AI output, edits it with real judgment, and sends the version that actually moves a decision? That's still you. Frequently asked questions What should a good AI executive summary prompt include? A good prompt needs five things: the role and audience you're writing for, the raw material pasted in, the format you want (length, headers, bullet points vs. prose), what to include (decisions, risks, asks, owners, dates), and an explicit instruction to flag gaps rather than fill them with guesses. Vague prompts get vague summaries. Specific prompts get drafts you can actually edit. Can AI write an accurate executive summary? AI can write a structurally clean summary, but accuracy is your responsibility. It compresses what you give it. If your input has wrong numbers, missing owners, or vague timelines, the output will too, just in more polished sentences. Always verify facts, add source links, and confirm any numbers before sending to leadership. Is it safe to paste work documents into ChatGPT for summarizing? It depends on the document and your company's AI policy. Consumer ChatGPT should never receive customer PII, employee records, legal or financial documents, security information, or confidential strategy. If your company has an enterprise AI tool with a data processing agreement, use that for sensitive material. When in doubt, ask IT or legal before pasting. What's the difference between an executive summary and a status report? A status report tracks what happened over a period. An executive summary frames what decision-makers need to know right now: the situation, the key risk, the recommendation, and the ask. Most status reports can be converted into executive summaries by cutting the chronology and leading with the so-what. The AI status report prompts collection handles the former; the prompts above handle the latter. How do I stop AI from inventing things in a summary? Two ways. First, give it real data: actual numbers, named owners, confirmed dates, explicit confidence levels. Second, add a line at the end of every prompt: "Flag anything that looks assumed or unverified rather than filling in the gap." That instruction dramatically reduces confident confabulation. Still read the output critically before anything goes to leadership. Should AI executive summaries go to leadership without human review? No. Not once. A summary that reaches senior leadership without a human read is a liability. The AI doesn't know the politics, the prior conversations, the financial sensitivity, or the stakeholder relationships. You do. Run the prompts for a faster first draft. Then read it, edit it, and send it as yourself. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Briefing Document Prompts: 10 Templates That Work 10 copy-paste AI briefing document prompts for meetings, exec decisions, client calls, and launch readiness. With a reusable formula and safety rules. ai promptsbriefing documentmeeting preparation AI Meeting Follow-Up Prompts (10 Templates) 10 copy-paste AI meeting follow-up prompts for ChatGPT and Claude. Turn messy notes into recaps, action items, and exec summaries without inventing decisions. ai promptsmeeting follow-upaction items AI Meeting Notes Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into summaries, action items, and client emails. With a reusable formula and safety reminders. AI promptsmeeting notesChatGPT at work --- ## AI Exploratory Testing Prompts: Find Hidden Bugs URL: https://dontreplace.me/articles/ai-exploratory-testing-prompts-find-the-bugs-scripts-miss AI promptsexploratory testingquality assurancesoftware testingbug testing AI Exploratory Testing Prompts: Find Hidden Bugs AI exploratory testing prompts for focused charters, unusual user journeys, sharper observations, reproducible bugs, and human-owned release decisions. August 11, 2026 Updated August 11, 2026 A scripted test confirms that the checkout button works when a polite user follows the expected path. An exploratory tester asks what happens after a coupon expires, the network drops, the browser restores an old tab, and the user double-clicks because the spinner looks dead. That is often where the interesting bugs live. AI exploratory testing prompts can help you plan those investigations. They can turn known risks into session charters, suggest alternate paths, organize observations, and expose questions your requirements forgot to answer. They cannot use the product for you, observe actual behavior, know undocumented rules, reproduce a defect, or decide whether a release is safe. AI can suggest where to look. A human still has to look, notice, test, and prove what happened. If you need formal checks from requirements, start with AI test case prompts. If you need the wider strategy, use AI test plan prompts. Exploratory testing complements both. It is structured investigation, not random clicking with a dramatic soundtrack. What exploratory testing actually is Exploratory testing combines learning, test design, and execution. The tester begins with a mission, uses the product, follows evidence, changes direction when behavior raises a useful question, and records enough context for another person to understand the result. A useful exploratory session usually has: a time-boxed charter; a named feature, journey, or risk; a specific build and approved environment; relevant personas, roles, and starting states; safe test data; known requirements and unresolved assumptions; observations, timestamps, screenshots, logs, or recordings in approved tools; reproducible defect notes; coverage notes and follow-up charters; a human owner for decisions. The goal is not to generate the longest list of things that could theoretically go wrong. The goal is to learn something valuable about the actual product. A focused 45-minute session that uncovers one damaging state transition beats a 90-row AI checklist nobody executes. Exploratory testing is especially useful when behavior is complex, requirements are incomplete, a feature has many paths, integrations can fail, or the cost of a strange edge case is high. It is also useful after scripted checks pass. Scripts protect known expectations. Exploration investigates what the team did not predict. The reusable AI exploratory testing prompt formula Use this foundation before the specialized templates: “Act as an exploratory testing planning assistant. I am [role] testing [feature or journey] in [approved environment and build]. Use only the verified requirements, known risks, personas, supported platforms, and constraints below. Draft a time-boxed charter with mission, scope, starting states, test ideas, oracles, evidence to capture, stop conditions, and follow-up questions. Separate confirmed facts from suggestions. Do not invent product behavior, claim execution, assign severity, reproduce private data, or approve a release. Label unsupported assumptions ‘needs human confirmation.’” An oracle is how you judge whether observed behavior may be wrong: a requirement, design, comparable workflow, business rule, accessibility standard, security policy, user expectation, or domain expert. AI may help list candidate oracles. Humans confirm which ones apply. Never paste customer PII, credentials, access tokens, production exports, private logs, payment details, health data, HR records, private tickets, vulnerability details, regulated data, or confidential client material into an unapproved AI tool. Use approved test environments, fictional records, redacted evidence, and your organization’s privacy and security process. What to collect before prompting A feature name is not enough context. Gather the safest useful inputs first. InputWhy it mattersHuman validation Feature goal and user journeyDefines what success should enableConfirm current scope Requirements and acceptance criteriaSupplies known expectationsMark conflicts and gaps Build and environmentMakes observations reproducibleConfirm exact version Personas, roles, and permissionsReveals access-dependent pathsVerify the real matrix Known risks and recent changesFocuses limited session timeRank by impact and uncertainty Supported devices and platformsPrevents imaginary coverageConfirm supported combinations Safe test-data rulesEnables realistic states without leakagePrivacy and security review Integrations and dependenciesIdentifies failure boundariesConfirm ownership and availability Evidence rulesMakes findings usefulUse approved storage and redaction Stop and escalation conditionsProtects systems and peopleName decision owners If an input is unknown, keep it unknown. Ask the model to draft a question. “The timeout behavior needs confirmation” is honest and actionable. “The system probably retries three times” is fan fiction wearing a lanyard. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI exploratory testing prompts Replace bracketed text with verified, non-sensitive context. Treat every output as a draft plan, never evidence that testing occurred. 1. Draft a focused session charter “Using this feature goal, current build, requirements, known changes, and ranked risks: [paste], draft a [30/45/60]-minute exploratory testing charter. Include mission, in-scope behavior, exclusions, starting state, personas, safe data needs, five to eight test ideas, applicable oracles, evidence to capture, stop conditions, and unanswered questions. Do not invent requirements or claim coverage.” A good charter is narrow enough to guide a real session. “Explore checkout” is a continent. “Explore how checkout preserves cart and payment state during interruption and recovery” is a mission. Keep exclusions visible. A session focused on recovery may notice an accessibility issue, but it should not quietly become the entire accessibility audit. Record the finding, route it, and preserve the original mission. 2. Map alternate user journeys “Review this verified primary journey, supported user roles, entry points, and business rules: [paste]. Create an exploration map of alternate paths, including backtracking, skipping optional steps, reopening old links, repeated actions, cancellation, resumption, and switching roles or devices where supported. Trace every idea to supplied context or label it as a hypothesis.” Happy paths are designed. Alternate paths are where users behave like users. They refresh, hesitate, open three tabs, arrive through an old bookmark, and change their minds halfway through. Do not test unsupported behavior as though the product promised it. You can explore it, but distinguish “documented expectation failed” from “we discovered an unanswered product question.” That distinction saves defect triage from becoming a courtroom drama. 3. Probe boundaries and weird inputs “Using these verified field rules, limits, formats, and business thresholds: [paste], propose exploratory test ideas around empty values, minimums, maximums, values immediately around boundaries, long text, supported characters, repeated submissions, unusual ordering, and conflicting combinations. Include the supplied rule, test purpose, safe input pattern, observation target, and cleanup need. Do not invent limits or generate attack payloads.” Boundary exploration should follow actual rules and risks. If nobody knows the maximum attachment size, the useful output is a question for the owner, not a model-generated number. Keep security work controlled. Do not feed private vulnerability details to a general model or ask it for exploit strings. Qualified security owners define scope, tools, environments, and evidence. 4. Explore state transitions “From this approved lifecycle, state diagram, and transition rules: [paste], create exploratory charters for valid transitions, blocked transitions, repeated actions, stale views, concurrent changes, partial completion, cancellation, rollback, and recovery. For each, include starting state, action sequence, expected rule if documented, observation points, and data reset. Flag missing transitions.” State bugs often hide behind a screen that looks fine. The visible status says “cancelled,” but a background job still sends the item. A role changes, yet an old tab retains an action. A retry creates two records. The model can organize possibilities. The tester must establish the starting state, execute the sequence, inspect all relevant surfaces, and capture evidence from the actual build. 5. Test roles and permissions “Using this verified role-permission matrix, ownership model, and approved test accounts: [paste], draft an exploratory session for allowed actions, denied actions, role changes, resource ownership, shared resources, tenant boundaries, stale sessions, and direct navigation. Include setup, expected permission source, UI and service observation points, evidence, and cleanup. Never generate credentials or assume hidden controls prove authorization.” A hidden button is not access control. Exploratory permission testing compares visible behavior, direct routes, service responses, state changes, audit trails, and cross-account effects where the approved scope allows it. Use fictional accounts in a test environment. Never paste passwords, tokens, production account IDs, or private authorization logs into the prompt. 6. Explore interruption and recovery “Given this documented workflow, persistence behavior, retry policy, and integration map: [paste], draft test ideas for refresh, back navigation, tab closure, session expiry, network loss, timeout, duplicate submission, dependency failure, restart, and resumption. Include the verified expectation, observation points, evidence, stop condition, and safe recovery step. Mark undocumented behavior as a question.” Recovery separates a mildly annoying product from one that loses work or creates costly duplicates. Test both what the user sees and what the system does, using only observability your role is authorized to access. Do not casually disconnect shared environments or trigger expensive downstream actions. Environment owners define safe failure simulation and stop conditions. 7. Compare devices, browsers, and environments “Using this supported platform matrix, responsive requirements, input methods, and environment constraints: [paste], create a risk-based exploratory matrix. Prioritize combinations by user impact, recent change, technical difference, and known uncertainty. Include viewport or device, browser, interaction method, journey, visual and functional observations, evidence, and excluded combinations. Do not invent support commitments.” “Test every browser” is not a strategy. Focus on supported combinations and meaningful differences: touch versus pointer, mobile keyboard behavior, narrow layouts, browser storage, camera permissions, file pickers, and interrupted connectivity. AI can help prioritize. Humans verify the support matrix and use real devices or approved services. A generated table does not prove cross-browser coverage. 8. Challenge accessibility and localization assumptions “Using these approved accessibility requirements, supported locales, content rules, and user journeys: [paste], propose exploratory checks for keyboard use, focus order, zoom, error recovery, labels, status announcements, text expansion, supported scripts, date and number formats, right-to-left layout if supported, and meaningful empty states. Cite supplied requirements and route specialist questions to human review.” This prompt helps identify areas to investigate; it does not replace accessibility testing, assistive-technology expertise, localization review, or users with relevant lived experience. Avoid stereotypes and invented language rules. Confirm supported locales, real content constraints, and expected assistive behavior with qualified reviewers. 9. Turn observations into reproducible bug reports “Using these redacted session notes and approved evidence references: [paste], draft candidate bug reports with concise title, exact build and environment, preconditions, numbered steps actually performed, observed result, expected source if documented, frequency, impact evidence, attachments to include, uncertainty, and follow-up question. Do not invent missing steps, logs, severity, root cause, or customer impact.” Messy notes are normal during exploration. AI is useful for structuring them afterward, provided it is forbidden from completing the story with plausible fiction. Before filing, reproduce the issue where safe, verify every step, inspect the evidence, and choose severity through the team’s human-owned process. AI bug report prompts can help with deeper defect documentation. 10. Review session notes for gaps and follow-ups “Review this charter, timestamped observation log, executed paths, defects, questions, and evidence index: [paste]. Separate tested from not tested, summarize learning, map observations to the charter, identify weak evidence, list unresolved risks, and propose focused follow-up charters. Do not claim complete coverage, infer unobserved behavior, or approve release readiness.” The debrief is where exploration becomes reusable knowledge. Record what changed in your understanding, not merely how many screens received clicks. Compare follow-up ideas with AI regression testing prompts and the broader AI QA checklist prompts. Some discoveries belong in repeatable automated or scripted checks; others need another investigation. A practical human-owned workflow Use the prompts in a controlled loop: Confirm the build, environment, permissions, and safe-data rules. Gather current requirements, changes, risks, and known unknowns. Pick one narrow mission and a realistic time box. Draft the charter with AI, then remove irrelevant theater. Prepare approved accounts, data, tools, and evidence storage. Execute the session manually and follow meaningful observations. Record actions, timestamps, behavior, and questions as they happen. Stop when safety conditions, time, or scope require it. Reproduce candidate defects and verify evidence. Debrief, file useful findings, and assign follow-up charters or scripted checks. For data-heavy sessions, use AI test data prompts to design fictional records from approved schemas. Never pull production records into a chatbot because the test account needs to feel “realistic.” Common mistakes with AI exploratory testing Generating a giant checklist A 200-item list feels comprehensive and destroys focus. Ask for a charter tied to one risk, journey, or change. Exploration should adapt to evidence, not serve a spreadsheet generated before anyone opened the product. Claiming coverage from ideas Suggested tests are not executed tests. Maintain a clear line between planned, attempted, observed, reproduced, and not tested. Release decisions need real evidence. Letting AI invent expected behavior When requirements are silent, label the behavior as a product question. Do not convert common conventions into promises the team never made. Humans resolve expectation conflicts. Losing build and starting-state details An observation without the exact build, environment, role, data state, and sequence may be impossible to reproduce. Capture context before the session wanders somewhere interesting. Assigning severity without evidence AI can draft an impact question. It cannot know frequency, affected users, financial exposure, contractual importance, or operational workarounds unless verified humans provide that evidence. Follow the real triage process. Pasting sensitive evidence into the prompt Screenshots, logs, tickets, and recordings can contain names, tokens, internal URLs, customer content, or vulnerabilities. Store evidence in approved systems, redact appropriately, and give the model only what policy permits. Treating exploration as random clicking Good exploration has a mission, models, notes, oracles, and a debrief. Curiosity is disciplined. Random activity is merely hard to invoice. Frequently asked questions Can ChatGPT do exploratory testing? It can help plan charters, propose test angles from supplied context, organize notes, and draft follow-up questions. It cannot interact with or observe your product unless connected through an approved testing system, and even then humans must validate execution, evidence, impact, and decisions. What should an exploratory testing prompt include? Include the exact build, approved environment, feature goal, known risks, verified requirements, user roles, supported platforms, safe-data constraints, session time box, evidence rules, and stop conditions. Require the model to separate facts from hypotheses. Is exploratory testing the same as ad hoc testing? Not necessarily. Strong exploratory testing is intentional and traceable: the tester uses a charter, learns from behavior, adapts the investigation, records evidence, and debriefs. Unstructured ad hoc testing may have no mission or durable notes. Does exploratory testing replace test cases? No. Scripted tests are useful for repeatable known expectations. Exploratory sessions investigate uncertainty, interactions, unusual paths, and new risks. Valuable discoveries often become future scripted or automated checks. How long should an exploratory testing session last? Choose a time box that protects focus—often 30 to 90 minutes depending on the mission and team practice. The exact duration matters less than having a narrow charter, recording learning, and stopping for a debrief. How do I measure exploratory testing? Track useful learning: risks investigated, paths exercised, questions resolved, defects reproduced, evidence quality, new charters, and checks added to the regression suite. Raw click counts and AI-generated idea counts are mostly decorative. Can AI decide bug severity? No. It can organize supplied impact evidence against a verified rubric, but humans must confirm reproducibility, affected users, business impact, safety, security, compliance, workarounds, and release consequences. What data is safe to give an AI testing assistant? Only data approved by your organization and necessary for the task. Prefer public requirements, approved schemas, fictional records, and redacted notes. Keep credentials, real customer data, private logs, vulnerabilities, regulated information, and confidential material out of unapproved tools. The useful boundary AI is good at producing options quickly. Exploratory testing needs options, but it also needs perception, skepticism, domain judgment, and accountability. The model can suggest that you interrupt a workflow. It cannot feel the confusion when recovery puts the user in an impossible state, inspect the real side effects, or decide whether that behavior matters enough to block release. Use AI to draft charters, vary perspectives, clean up notes, and expose unanswered questions. Keep humans on environment access, execution, evidence, sensitive context, reproduction, impact, specialist review, and release judgment. For a broader field guide to using AI without surrendering the parts of work that require judgment, Don’t Replace Me by Dmitry Kargaev follows the same rule: use the machine for speed, and keep responsibility attached to people. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance AI API Testing Prompts: Catch Broken Integrations AI API testing prompts for contracts, edge cases, authorization, retries, error responses, and human-verified integration evidence. AI promptsAPI testingquality assurance AI Compatibility Testing Prompts: Catch Breakage 10 AI compatibility testing prompts for browser, device, OS, responsive, locale, network, defect triage, and evidence-based release decisions. AI promptscompatibility testingcross-browser testing --- ## AI Feedback Prompts: Give Clearer Feedback at Work URL: https://dontreplace.me/articles/ai-feedback-prompts-give-clearer-feedback-at-work ai feedback promptsmanager feedbackgiving feedback at workai at workfeedback templates AI Feedback Prompts: Give Clearer Feedback at Work 10 copy-paste AI feedback prompts for managers. Turn messy notes into specific feedback, prep hard conversations, and know when to escalate. June 30, 2026 Updated June 30, 2026 Good feedback is specific, timely, and honest. Most managers give feedback that's vague, late, and hedged into uselessness. AI feedback prompts can help you get from "I have some concerns about Jake" to "Jake missed three consecutive deadlines on the Hendrix account, which pushed the client handoff back by two weeks." That's not magic. That's just having a thinking partner who doesn't get tired of helping you find the right words. But let's be honest about what AI can and can't do here. It can sharpen language. It can separate your frustration from the actual facts. It can help you prepare for a hard conversation. It cannot give you courage. It cannot replace the relationship. And it absolutely cannot handle anything involving HR investigations, discrimination claims, harassment, or legal exposure. More on that at the end. Here are 10 prompts you can use today. The formula every AI feedback prompt needs Before we get to the templates, you need to understand why most AI feedback attempts produce polished garbage. You paste in "my employee is always late and it's affecting morale" and get back a lovely paragraph about "punctuality expectations in a team context." Still useless. Just dressed up. The fix is giving the AI actual inputs. Every prompt in this list follows the same structure: Role + Specific behavior + Observable impact + What good looks like + Tone you want Skip any of those and you get vague back. Feed them in and the output is actually usable. Think of it as Rule #13 from Don't Replace Me: garbage in, garbage out. Vague resentment becomes polished vague resentment. Specific examples become specific feedback. Prompt 1: Turning messy notes into specific feedback You've been scribbling notes for three weeks. You have: "missed thing again," "tone weird in that meeting," and "good job Friday." Not usable. The prompt: I'm a manager preparing written feedback for a direct report. Here are my raw notes from the past month: [paste your notes]. The employee's role is [job title]. Their main responsibilities include [list 2-3]. Please help me organize these into specific, observable feedback points. For each point, flag where I need to fill in a concrete example before this is ready to use. The "flag where I need a concrete example" instruction is the key part. Don't skip it. You want AI to show you the gaps, not paper over them. Prompt 2: Separating facts from interpretations You think Jake doesn't care about the project. But "doesn't care" is an interpretation. What you actually observed was that he submitted two deliverables without spellchecking them and showed up to the client call without reviewing the brief. Those are facts. The prompt: I'm preparing feedback for an employee. Here's what I wrote: [paste your draft]. Please separate this into two columns: observable facts and behaviors, and my interpretations or inferences about motivation or attitude. Keep my exact wording for the facts column. Flag any interpretations that I'd need to rephrase before sharing them directly. This one saves you from the conversation where the employee spends 20 minutes defending their attitude instead of acknowledging the behavior. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Prompt 3: Writing constructive criticism The standard advice is "be direct but kind." Genuinely unhelpful. Direct and kind looks different depending on the person, the issue, and what actually needs to change. The prompt: I need to give an employee constructive feedback about [specific behavior or pattern]. The context is [2-3 sentences of background]. The impact of this behavior has been [what actually happened as a result]. The expectation for this role is [what good looks like]. Please draft feedback that is direct, specific, and focused on behavior and impact. Avoid softening language that dilutes the message. Keep it under 150 words. Use this for written feedback or as a starting point for a verbal conversation. Always read it aloud before you use it. If it doesn't sound like you, edit it until it does. Prompt 4: Writing specific praise Generic praise is almost as useless as no praise. "Great work this quarter" tells someone nothing about what to repeat. Specific praise is a retention tool. The prompt: I want to recognize an employee for something they did well. Here's what happened: [describe the specific situation and what they did]. The result was [observable outcome]. Please write a short piece of specific positive feedback that names the behavior, explains why it mattered, and connects it to their role or team goals. Keep it genuine, not corporate. Specific praise is also the part of feedback most AI handles well, since there's no conflict to manage. Still, read it. "I was genuinely impressed by your proactive stakeholder communication" sounds like a compliance form. Rewrite it in words you'd actually say. Prompt 5: Preparing a feedback conversation Most feedback goes sideways not because the words were wrong, but because the manager wasn't ready for the employee's reaction. Preparation is the job. The prompt: I'm preparing for a feedback conversation with an employee. The main issue I need to address is [describe the behavior and impact]. I've seen this pattern [number] times, most recently on [date or project]. The employee's role is [job title]. Please help me: (1) structure the conversation with an opening, the core feedback, and a path forward, (2) list 3 questions I should ask the employee to understand their perspective, and (3) identify one thing I might be missing that could change how I deliver this. That third request is the valuable one. It forces the output to consider what you don't know. Prompt 6: Role-playing the employee response You say "the client presentation felt underprepared" and they say "I worked 60 hours on that." Now what? The prompt: I'm about to give feedback to [describe the employee in neutral terms: tenure, performance history if relevant, general disposition]. The feedback I'm planning to deliver is: [paste your draft]. Please play the role of this employee and give me 3 likely responses to this feedback, ranging from receptive to defensive to dismissive. For each response, suggest how I should continue the conversation. This is one of the better uses of AI in feedback prep. You're not outsourcing judgment. You're war-gaming the conversation before it happens. Prompt 7: Drafting a follow-up note Feedback without follow-up disappears. A written note after the conversation isn't bureaucratic. It's a record of what was said and agreed. The prompt: I just had a feedback conversation with a direct report. Here's a summary of what we discussed: [key points, any commitments made, timeline]. Please draft a short follow-up note I can send within 24 hours that summarizes the conversation, confirms any agreed next steps, and keeps a constructive tone. Keep it under 200 words. Don't make it sound like an HR document. That last instruction matters. A follow-up that reads like a disciplinary memo poisons the relationship even when the conversation went well. For more on drafting follow-up communications, the AI manager prompts guide has templates specifically for post-meeting notes. Prompt 8: Turning feedback into a growth plan The feedback conversation happened. Now what changes? A growth plan turns a single conversation into a direction. The prompt: Based on the following feedback I gave an employee: [paste feedback], and their role responsibilities [list them], please draft a simple 90-day growth plan. Include: 3 specific behaviors to develop, 2 ways we'll measure progress, and a checkpoint schedule. Keep it practical and achievable, not aspirational. The checkpoint schedule is the part most growth plans skip. Without it, you check in at the end of 90 days and nobody remembers what the goal was. If you're working through a formal performance review process, the AI performance review prompts article has templates that connect directly to this kind of follow-through. Prompt 9: Checking tone for clarity and respect You wrote the feedback. You read it twice. You still can't tell if it sounds harsh or if you're just being precious about it. The prompt: Please review the following feedback draft for tone: [paste feedback]. Tell me: (1) Does any language come across as personal, dismissive, or attacking character rather than behavior? (2) Is anything unclear or open to misinterpretation? (3) Is there anything that softens the core message to the point where the employee might not grasp how serious this is? Please be direct. I'd rather know now than during the conversation. That third check is the one most tone-review prompts miss. Feedback gets polished into meaninglessness as often as it stays too harsh. Prompt 10: Deciding whether AI is the right tool at all Some situations don't need better wording. They need HR, legal, or leadership. The prompt: I'm dealing with a workplace situation and I'm unsure whether to handle it myself, use AI to help me prepare, or escalate. Here's what happened: [describe the situation in general terms without names or identifiable information]. Please help me identify: (1) whether this situation has characteristics that typically require HR, legal, or leadership involvement, (2) what I might be overlooking, and (3) if this is appropriate for manager-level handling, what the first step should be. This prompt is for you, not for AI to resolve the situation. It's a thinking framework. Do not use AI tools to process, draft, or analyze situations involving: harassment claims, discrimination complaints, retaliation concerns, threats of violence, active HR investigations, protected employee health information, compensation data, or formal performance improvement plans. These require your HR team, employment counsel, or both. The AI conflict resolution prompts article has more guidance on where that line sits. What AI can't do in feedback situations AI can make your language more precise. It can't make you more honest. It can't make you show up to the conversation instead of sending an email. It can't make an employee feel safe enough to actually hear you. The managers who give the best feedback aren't the ones with the most polished language. They're the ones the employee believes. That's what Rule #4 in Don't Replace Me calls the trust currency. Feedback only lands when the person trusts your intent and trusts that your facts are real. If you've used AI to replace the actual thinking, you'll notice. The employee will also notice. Something will feel slightly off about the specificity, or the phrasing won't sound like you, or the conversation will go sideways at the first pushback because you hadn't actually done the work of deciding what you believe. Use these prompts to prepare. Then put the output away and talk to the person. If you're just getting started with AI tools at work, the no-BS guide to using AI at work covers how to build a sustainable habit without the hype. And if you want to understand where AI genuinely breaks down before you rely on it for anything important, what AI can and can't do is worth five minutes. The field guide for staying human when everyone else mistakes smoother wording for actual leadership is Don't Replace Me. The feedback chapter is the one most managers bookmark. Frequently asked questions Can AI write feedback for me? AI can draft, structure, and refine feedback based on examples you provide, but it can't verify facts, understand your relationship with the employee, or substitute for your judgment. Always treat AI output as a first draft that needs your review before any conversation or documentation. Is it okay to paste employee information into ChatGPT for feedback prep? Only use general descriptions without names, identifiable details, compensation data, health information, HR investigation details, harassment or discrimination claims, or anything from formal performance management processes. When in doubt, ask your IT or HR team whether the tool is approved for this use. What's the difference between feedback and a performance improvement plan? Feedback is a normal part of managing: specific, timely, two-way. A performance improvement plan (PIP) is a formal HR process with legal implications and usually involves documentation, HR oversight, and defined consequences. Don't use AI to draft a PIP without HR involvement. How do I give feedback when I don't have specific examples? You don't. Vague feedback is worse than no feedback because it creates defensiveness without giving the person anything to act on. If you can't point to a specific behavior and its impact, you need to spend more time observing before having the conversation. What should I do if the employee gets upset during the feedback conversation? Pause. Acknowledge what you heard. Don't keep talking to fill the silence. Most feedback conversations go wrong because the manager keeps adding caveats when the employee gets emotional. Give them time to process. You can always continue the conversation at another time if the current one breaks down entirely. When should I involve HR instead of handling feedback myself? Involve HR when the situation involves harassment, discrimination, retaliation, a protected class, threats, formal disciplinary action, or anything that could have legal implications. Also involve HR if the employee has previously raised a complaint against you. When in doubt, a quick conversation with HR before the meeting is always better than a conversation with HR after things go wrong. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Automation Prompts: Automate the Boring Stuff 10 copy-paste AI automation prompts for mapping workflows, writing SOPs, adding approval gates, and QA testing. Practical templates with safety rules included. ai automation promptsworkflow automationSOP templates AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts --- ## AI Go/No-Go Decision Prompts (10 Templates) URL: https://dontreplace.me/articles/ai-go-no-go-decision-prompts-stop-launching-on-vibes go/no-go decisionlaunch readinessAI promptsdecision memolaunch criteria AI Go/No-Go Decision Prompts (10 Templates) 10 copy-paste AI go/no-go decision prompts for launch readiness. Structure evidence, find gaps, draft memos. AI helps you think. Humans make the call. July 26, 2026 Updated July 26, 2026 Someone on your team says "I think we're ready." Someone else says "I feel like there are still some issues." A third person says nothing but their face is doing something concerning. And then you launch anyway, because the meeting ran long and the deadline was Tuesday. AI go/no-go decision prompts won't fix your organization's relationship with accountability. But they will stop you from dressing up "we have no idea" as "we've reviewed the evidence." That's not nothing. The problem with most launch decisions isn't that people are lazy. It's that readiness evidence is scattered across Jira tickets, Slack threads, a half-finished QA doc, and someone's memory of a conversation from three weeks ago. AI is genuinely useful here: it can pull structure out of chaos, turn your messy notes into a coherent decision memo, and ask you the questions you forgot to ask yourself. What it can't do is tell you whether you're actually ready. That's still your call. What AI go/no-go decision prompts are actually good for Used well, AI is a thinking partner for launch decisions, not a decision-maker. You paste in your evidence, your blockers, your QA results, and your stakeholder concerns. It helps you organize them into something a human can actually evaluate. The trap is when people treat the output as the answer. A clean, confident-sounding decision memo doesn't mean you're ready to ship. It means your uncertainty is now formatted nicely. Rule #5 in Don't Replace Me puts it plainly: it's not smart, it's fast. Organized uncertainty is still uncertainty. Use these prompts to structure your thinking, surface what's missing, and draft communications. Keep a human in the loop for every actual decision. And before you paste anything into an AI tool, make sure you're not including customer PII, credentials, access tokens, private financial records, HR data, legal disputes, security vulnerabilities, or anything regulated. Use only what's approved for your AI tool. The reusable go/no-go prompt formula Every prompt in this set follows the same structure. Customize the bracketed sections for your specific launch. You are a [launch readiness analyst / decision memo writer / risk reviewer]. Context: [What we're launching. Who it affects. When it's supposed to go live.] Here is the evidence: [Paste your notes, QA summary, blocker list, stakeholder feedback, no PII, no credentials, no sensitive data.] Your task: [Specific output you need, criteria checklist, decision memo, objection list, etc.] Format: [Bullet list / table / short memo / numbered summary.] Flag anything that looks uncertain, missing, or unverified. That last line matters. It's the difference between an AI that writes you a confident-sounding launch summary and one that tells you what you don't actually know yet. 10 copy-paste AI go/no-go decision prompts Prompt 1: Define your launch criteria Use this before you've gathered any evidence. Forces you to agree on what "ready" actually means. You are a launch readiness analyst. We are preparing to launch [DESCRIBE LAUNCH]. Help me define the go/no-go criteria we should evaluate before making a launch decision. Include: technical readiness checks, QA status, stakeholder approvals, rollback readiness, support readiness, customer impact assessment, and any timing or dependency constraints. Format as a checklist with a Yes/No/Partial column for each item. Flag any criteria that require a named human owner. Prompt 2: Summarize launch evidence into a decision memo Use this when you have evidence scattered across documents, Slack, tickets, and meeting notes. Paste in the sanitized version. You are a technical writer preparing a launch decision memo. Here is the evidence from our readiness review: [PASTE SANITIZED NOTES, no PII, credentials, or sensitive data]. Summarize this into a structured go/no-go memo with sections for: current status, confirmed readiness items, outstanding blockers, unverified assumptions, and a recommended next step. Do not fabricate any facts. If something is unclear or missing from the evidence, say so explicitly. Prompt 3: Find missing readiness facts This is one of the most useful prompts in the set. It doesn't tell you you're ready. It tells you what you can't prove yet. Review the following launch readiness summary: [PASTE SUMMARY]. Identify all claims that are not supported by specific, verifiable evidence. For each gap, write: what information is missing, who would likely own it, and how critical it is to the go/no-go decision. Format as a table: Missing Evidence | Likely Owner | Criticality (High/Medium/Low). Prompt 4: Turn QA results into a decision input QA reports are usually a mess of pass/fail rows with no clear recommendation. This prompt fixes that. For structuring your QA process before you get to this stage, these QA checklist prompts can help. You are a QA analyst summarizing test results for a launch decision. Here are the test results: [PASTE SANITIZED QA SUMMARY, no system credentials, no customer data]. Produce a decision-ready QA summary with: total tests run, pass rate, open defects by severity (Critical/High/Medium/Low), any blockers that prevent launch, and a recommended QA verdict (Pass / Conditional Pass / Fail). Include your confidence level and flag anything that requires human review before sign-off. Prompt 5: Compare launch now vs. delay tradeoffs When the team is split, this prompt forces both sides of the argument into writing before someone just wins by being louder. We are deciding whether to launch [DESCRIBE LAUNCH] on [DATE] or delay to [ALTERNATIVE DATE]. Help me map the tradeoffs. For Launch Now: list likely benefits, known risks, and which risks we've accepted explicitly. For Delay: list what we gain, what we lose, and what would need to be true for the delay to be worth it. Format as a side-by-side comparison. Do not recommend a final decision, that requires human judgment and accountability. Prompt 6: Check rollback readiness Launches fail. The question is whether you can undo it. If you want a full rollback plan, the rollback plan prompts cover the full process. You are reviewing launch rollback readiness. Here is what we know about our rollback plan: [PASTE SANITIZED ROLLBACK NOTES]. Identify: whether a rollback plan exists, who owns it, what triggers would activate it, how long rollback would take, what data or customer state would be affected, and whether the rollback has been tested. Flag any gaps as High/Medium/Low risk. Do not invent facts about our systems. Prompt 7: Map customer impact and support load This one gets skipped constantly. Someone ships at 9am on a Monday and the support team finds out on Slack. We are launching [DESCRIBE LAUNCH]. Help me assess customer impact and support readiness. Based on this context: [PASTE RELEVANT PRODUCT/LAUNCH INFO, no customer PII, no regulated data]. Produce a customer impact summary covering: who is affected and how, what errors or confusion are likely in the first 24 hours, what the support team needs to know before launch, what monitoring should be in place, and what the escalation path looks like if something goes wrong. Prompt 8: Draft stakeholder objections and responses Someone in the room is going to object. Better to prepare for it than get blindsided. For broader stakeholder communication, the stakeholder update prompts are worth bookmarking. We are about to present a go/no-go recommendation for [DESCRIBE LAUNCH] to [DESCRIBE STAKEHOLDERS: exec team / board / client / ops team]. Based on this context: [PASTE SANITIZED SUMMARY]. Draft the three to five most likely objections we'll face, and prepare a factual, honest response to each. Do not make up data or fabricate readiness. If an objection is valid and we don't have a good answer, say that clearly so we can address it before the meeting. Prompt 9: Create a final go/no-go recommendation This prompt produces a structured recommendation memo. It is a communication tool, not a substitute for human sign-off. You are writing a final go/no-go recommendation memo for [DESCRIBE LAUNCH]. Based on the following readiness summary: [PASTE SANITIZED SUMMARY]. Write a one-page memo with: the launch being evaluated, the decision date, the evidence reviewed, confirmed ready items, outstanding risks, the recommendation (Go / No-Go / Conditional Go with named conditions), and the required sign-off owners. Include a line at the bottom: "This memo summarizes available evidence. Final decision and accountability rest with [NAMED DECISION-MAKER]." Do not fabricate any facts. Prompt 10: Write the post-decision communication Whether you launch or delay, someone needs to tell people. This prompt drafts that message. We have made a [GO / NO-GO / DELAY] decision for [DESCRIBE LAUNCH]. The reason is: [SUMMARIZE THE REASON, no sensitive business strategy, no confidential data]. Draft an internal communication to [TEAM / STAKEHOLDERS / SUPPORT] that explains: what was decided, why, what happens next, who owns the next steps, and when the next update will come. Keep it honest and clear. Do not make the delay sound like a success or the launch sound more certain than it is. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What AI genuinely can't do in a go/no-go decision It's worth being blunt about the limits. AI cannot verify that your systems are actually ready. It cannot confirm that your security review happened. It cannot tell you whether an unresolved ticket is a deal-breaker for your specific customer base. It cannot substitute for compliance review, legal review, or executive accountability. If you paste in vague notes and ask for a readiness assessment, you'll get a well-formatted readiness assessment. It won't say "this is based on nothing." It'll sound confident. That's the failure mode. The risk assessment prompts are a good companion here, specifically for pressure-testing whether the blast radius of a bad launch is actually understood. And if you want the broader argument for why AI is a useful thinking tool but a dangerous authority, what AI can and can't do covers that without the jargon. The human value in launch decisions isn't filling out the checklist. It's knowing which risks are acceptable, which objections actually matter, and when a clean document is hiding a bad decision. That's taste. AI doesn't have it. Before you paste anything: the data safety checklist Every time you use these prompts, check against this list before you hit send. Do not paste: Customer names, emails, IDs, or any personal data Passwords, API keys, access tokens, or system credentials Raw production logs or error traces containing user data Unreleased product strategy or confidential roadmap details Legal disputes, ongoing litigation, or regulatory notices Security vulnerabilities or penetration test results Financial records, revenue figures, or budget details under NDA HR issues, performance records, or employee complaints Medical or health information of any kind Private client conversations or contract terms Any data category regulated in your industry (HIPAA, GDPR, SOC 2, etc.) If you're not sure whether something is safe to paste, don't paste it. Describe it in general terms instead. Frequently asked questions What is an AI go/no-go decision prompt? An AI go/no-go decision prompt is a structured instruction you give an AI tool to help organize launch readiness evidence, identify missing facts, draft decision memos, or map stakeholder objections. It's a thinking aid, not a decision-maker. The final call always requires a named human with actual accountability. Can AI actually tell me if my launch is ready? No. AI can organize the evidence you give it and flag gaps in what you've shared. It can't verify whether your systems are working, whether your security review actually happened, or whether that P1 ticket is actually a dealbreaker for your users. If you want real readiness, you need real evidence from real humans, structured clearly so everyone can evaluate it. What's the biggest mistake people make using AI for launch decisions? Treating a well-formatted output as confirmation of readiness. AI is fast and organized. A polished go/no-go memo doesn't mean you're ready to ship, it means your uncertainty is now formatted nicely. The prompt set above is specifically designed to surface what's missing, not just summarize what you think you know. What data should I never paste into an AI tool for launch reviews? Customer PII, credentials, access tokens, raw production logs, legal disputes, security vulnerabilities, financial records under NDA, HR data, medical information, regulated data of any kind, and unreleased strategy. When in doubt, describe the situation in general terms rather than pasting the source document. How is a go/no-go prompt different from a regular launch checklist? A checklist confirms whether specific items exist. A go/no-go prompt helps you evaluate whether the evidence is strong enough to make a decision, find what's missing, structure the tradeoffs, draft stakeholder communication, and prepare for objections. They work well together. The launch checklist prompts handle the checklist side; this set handles the decision layer. Who should have final sign-off on a go/no-go decision? A named human with actual authority and accountability for the outcome. Not "the team." Not the AI memo. A specific person whose name goes on the decision. Every prompt in this set includes a placeholder for that. If your organization doesn't know who that person is before the launch meeting, that's the first thing to fix. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Operational Readiness Prompts: 10 Templates 10 copy-paste AI operational readiness prompts for go-live checklists, dependency checks, owner gaps, and stakeholder updates. Includes the core formula. operational readinessAI promptslaunch readiness AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management --- ## AI Goal Setting Prompts: 10 Templates That Work URL: https://dontreplace.me/articles/ai-goal-setting-prompts-10-templates-that-work ai goal settinggoal setting promptschatgpt for goalssmart goals aiwork goals templates AI Goal Setting Prompts: 10 Templates That Work 10 AI goal setting prompts for ChatGPT and Claude. Turn vague ambitions into concrete goals with metrics, owners, and leading indicators. June 29, 2026 Updated June 29, 2026 Most goal-setting advice starts with "dream big." Then you write down something like "become a top performer" and feel great about it for 48 hours before doing nothing. AI can make that process faster. It'll take your vague ambition and dress it up in bullet points and action verbs until it looks like a real plan. That's the trap. AI goal setting prompts are genuinely useful. But only if you know what you're asking for, and what you're not. Used well, AI is a drafting partner that catches the stuff you glossed over: missing metrics, hidden assumptions, capacity problems you didn't want to think about. Used badly, it creates confident nonsense that sounds like leadership and gets you nowhere. These 10 prompts are designed to do the former. What makes AI goal setting actually useful? The problem isn't that your goals are vague. The problem is that vague goals feel fine until something goes wrong and nobody knows who was supposed to fix it or how you'd even know it was fixed. AI is good at making the invisible visible. You give it your ambition ("grow the product," "improve team performance," "reduce churn") and it can ask you the questions you avoided. What does "grow" mean in numbers? By when? Who owns the decision when the approach isn't working? What are you trading off to get there? The catch is Rule #13 from Don't Replace Me: garbage in, garbage out. If you start with a fuzzy ambition and ask AI to "turn it into a SMART goal," you'll get SMART formatting around the same fuzzy thinking. The prompts below are designed to force better input before you get the polished output. One more thing before the templates: don't paste confidential strategy documents, employee performance records, unreleased financial data, customer PII, or anything sensitive into AI tools your company hasn't cleared for that purpose. Tidy goal language isn't worth a compliance incident. The reusable AI goal-setting formula Before the 10 prompts, here's the underlying structure. Every good goal-setting prompt does three things: Gives AI the real context (not the sanitized version) Asks for the uncomfortable questions, not just the polished output Keeps you in the decision seat The formula: [Context] + [Your draft goal or ambition] + [What you want AI to do] + [Constraints or known limits] Most people skip the context and the constraints. That's why they get output that looks good and means nothing. Think of it like briefing a new hire. If you hand them a task with no background, they'll complete it confidently and get it completely wrong. Same thing happens here. There's also a sequencing issue worth naming. Most people go straight to "write me a plan" when they should be asking "what am I actually trying to change?" The prompts below are ordered to match how goal thinking actually works, not how it gets presented in slide decks. Prompt 1: Turn a vague ambition into a concrete goal You have a feeling, not a goal. Something like "we need to be better at retention" or "I want to make more of an impact this year." Here's how to get past that. "I have a vague ambition: [paste it here]. Help me turn this into a concrete goal by asking me 5 clarifying questions. Focus on: what success looks like in measurable terms, the timeframe, who it affects, and what I might be assuming. Don't write the goal yet. Just ask the questions." The "don't write the goal yet" part matters. If you skip it, AI will write you something that sounds finished before you've actually done the thinking. Most people paste that output straight into their quarterly plan and wonder why nothing happens by month three. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Prompt 2: Convert a draft goal into SMART language Once you have something concrete, SMART formatting is a useful sanity check. Not the goal itself, just a translation. "Here's my draft goal: [paste it]. Rewrite it in SMART format: Specific, Measurable, Achievable, Relevant, Time-bound. Where you can't fill in a letter because I haven't given you the information, flag it as [MISSING] and explain what I need to clarify. Don't invent numbers or assumptions." That [MISSING] instruction is the key line. Without it, AI will fill in the blanks with something plausible-sounding. Then you'll use those invented numbers like they mean something. "Increase NPS by 12 points by Q4" looks authoritative. But if AI made up 12 because you didn't give it a baseline, you're measuring yourself against a fiction. Prompt 3: Find the right metric Picking the wrong metric is how you hit a goal and still lose. Revenue goes up, margin goes down. Ticket volume goes down, customer satisfaction doesn't move. This prompt pressure-tests your measurement choice before you commit to it. "My goal is: [paste it]. My current plan is to measure success by [metric]. What are the risks with that metric? What might it miss? Suggest 2-3 alternative metrics and explain the tradeoffs. Be specific about what each metric would and wouldn't capture." For data that involves real customers, revenue, or performance figures, keep it anonymized. "Our current churn rate is roughly X%" is fine. Actual customer names or contract values are not. This step matters more than most people think. The metric you pick shapes what the whole team optimizes for. Pick the wrong one and you get Goodhart's Law in action: the measure becomes the target, and the target stops measuring what you cared about. Prompt 4: Separate outcomes from activities This is the most common goal-setting mistake. "Host monthly training sessions" is an activity. "Reduce average onboarding time from 6 weeks to 4 weeks by Q3" is an outcome. They look similar until review time, when you've done all the activities and nothing changed. "Here's my goal as written: [paste it]. Is this an outcome (a result you're achieving) or an activity (a thing you're doing)? If it's an activity, help me rewrite it as an outcome. If it's already an outcome, confirm what activity would be the main driver." This prompt works well alongside the AI prioritization prompts when you're trying to sequence a stack of goals against available capacity. Prompt 5: Pressure-test your assumptions Every goal has buried assumptions. "We'll hit this if the market holds," "if we get budget approval," "if that hire comes through." AI is decent at surfacing these, because it's not emotionally attached to your plan. "Here's my goal: [paste it]. Here's the context: [brief situation description, no confidential data]. List the top 5 assumptions I'm making for this goal to be achievable. For each one, rate the risk if the assumption is wrong (low / medium / high). Don't validate the goal. Look for what could break it." Review this output carefully. AI doesn't know your actual organization. It'll find generic risks, and some of them will be wrong for your situation. That's fine. Use it as a starting checklist, not a final answer. The value isn't a perfect risk register. It's the 10 minutes of uncomfortable reading that makes you think "oh, we never actually confirmed that dependency." Prompt 6: Identify constraints and dependencies A goal with no constraints listed is a fantasy. You have a budget, headcount, a Q3 freeze, a dependency on another team that hasn't agreed to anything. Write those down before you finalize anything. "My goal is: [paste it]. I'll list what I know about constraints: [budget range if known, team size, key dependencies, timeline pressures]. Based on this, identify what's most likely to cause a delay or failure. Then suggest what I should confirm or negotiate before committing to this goal." If you're working on quarterly planning and need to connect these constraints to a real plan, the AI quarterly planning prompts walk through this in more detail. The dependency conversation is the one most people skip because it involves awkward cross-functional negotiations. But a goal that requires another team's cooperation, without that team knowing they've been written into your plan, is just a wish dressed up as a goal. Prompt 7: Connect a goal to weekly work Big goals die because nobody knows what to do on Tuesday. The translation from a 90-day goal to a weekly task list is where most execution falls apart. "My goal is: [paste it]. My timeframe is [X weeks/months]. Break this down into weekly milestones for the first 4 weeks. For each week, suggest 1-2 specific actions and 1 thing to check or measure. Flag anything that requires a decision or approval from someone else." This is a starting point. You'll need to adjust for what's actually on your plate. The AI doesn't know you're also running a product launch and have two direct reports on leave. That part is on you. What this prompt is really doing is exposing the gap between your goal's ambition and your actual calendar. Most goal failures aren't strategic. They're scheduling. The work didn't happen because something more urgent always filled the week. Seeing the week-by-week breakdown forces a reckoning with that before you commit publicly. Prompt 8: Write owner and accountability language "The team will handle this" means nobody will handle this. Goals need a named owner, a decision rights statement, and clarity on who gets escalated to when something's off-track. "Here's a goal I'm finalizing: [paste it]. Write an ownership statement that includes: the named owner (I'll fill in the name), what decisions they can make independently, what they need to escalate, and a review cadence. Keep it short and plain. One paragraph." Don't finalize accountability structures without the actual people involved. This prompt gives you a draft. The real conversation still needs to happen. McKinsey has a useful plain-English piece on getting clear about who is accountable for what. The practical point is simple: a goal where ownership is genuinely named, not vaguely assigned to a function or a team, is easier to review, escalate, and fix before it quietly dies. Prompt 9: Create leading indicators Lagging indicators tell you after you've failed. Leading indicators give you a warning before the quarter ends. Revenue is lagging. Pipeline growth rate is leading. The difference is whether you have time to respond. "My goal is: [paste it]. The lagging indicator I'm using to measure success is [metric]. Suggest 3 leading indicators I could track weekly or bi-weekly to predict whether I'm on track. For each one, explain what it would signal and how early in the process it would show up." Pair this with your status reporting process. The AI status report prompts can help you build the update format around these indicators once you've chosen them. Prompt 10: Build a monthly goal review Goals set in January get forgotten by March. A monthly review cadence is the simplest fix, and it works better when the review questions are already written before the meeting. "My goal is: [paste it]. The key metric is [metric]. Write a monthly review template with 5 questions that a goal owner should answer in writing before each review meeting. Include questions about: progress against the metric, assumptions that need updating, blockers, decisions needed, and whether the goal itself still makes sense." That last question is the one most people skip. Circumstances change. A goal that made sense in Q1 can be the wrong goal in Q3. Build in permission to say so. A written pre-read also changes the meeting. Instead of spending 20 minutes on status updates that could have been a spreadsheet, you spend 20 minutes on the actual decision that's blocking progress. That's the goal of the goal review. When these prompts work best together You don't have to use all 10 every time. Here's a rough guide by situation: SituationPrompts to use Starting from scratch1, 2, 3, 5 Cleaning up existing goals2, 3, 4 Quarterly planning cycle1, 5, 6, 7, 10 Adding a new team goal4, 8, 9 Goal already off-track5, 6, 9, 10 The sequence matters more than completeness. Trying to write accountability language (Prompt 8) before you've confirmed the metric (Prompt 3) means you're assigning ownership of something you haven't fully defined yet. That's how you get review meetings where everyone agrees the goal is owned but nobody can explain what success looks like. What these prompts won't do for you They won't tell you what to prioritize when two goals conflict and you can't do both. They won't make the political conversation with another department happen. They won't replace the judgment call about which metric actually reflects what your customers care about. Rule #7 in Don't Replace Me is about taste as a moat: knowing which outcome matters, what quality looks like, what not to chase. AI can draft options. It cannot have opinions about your business. The human in the loop isn't a compliance formality. It's the part where the actual thinking happens. The AI decision-making prompts are useful when you've used these templates to get clear on what you're aiming for but still need to work through a hard tradeoff. And if you're connecting goals to budget, the AI budget planning prompts cover that translation. A tidy goal statement is not a plan. And a plan with no owner, no metric, and no constraints isn't a plan either. It's a document waiting to be ignored. Use these prompts to build something real. Then close the tab and go talk to the people who need to actually execute it. Frequently asked questions Can AI write my goals for me? AI can draft goal language quickly, but the underlying thinking still needs to come from you. If you paste in a vague ambition and ask for a SMART goal, you'll get polished formatting around fuzzy thinking. Use AI to pressure-test your draft, not to create the goal from scratch. What's the difference between an outcome and an activity in goal setting? An outcome is a result you're trying to achieve (reduce churn by 15%). An activity is something you're doing (run monthly customer check-ins). Activities can sound like goals but they're not. You can complete every activity and still miss the outcome. Always write goals as outcomes. Is it safe to paste my goals into ChatGPT or Claude? It depends on what's in them. Vague ambitions and draft frameworks are generally fine. Don't paste in confidential strategy documents, employee names with performance details, customer data, unreleased financials, or anything your company's AI use policy prohibits. When in doubt, anonymize and generalize before pasting. What's a leading indicator and why does it matter for goals? A leading indicator is something you can measure before the final outcome, that predicts whether you're on track. Revenue is a lagging indicator. Sales pipeline growth is leading. If you only track lagging indicators, you find out you missed the goal after it's too late to fix anything. Build leading indicators into your goal from day one. How often should I review goals? Monthly is the minimum that actually works. Weekly is better for goals tied to fast-moving work. The point of a review cadence isn't to report numbers, it's to catch when an assumption has broken or the goal itself needs updating. Set the review questions before the quarter starts so the meeting has structure, not just a status update. Do I need to be technical to use these AI prompts? No. These prompts are plain English. You don't need to understand how AI works, just how to describe your goal and its constraints clearly. The better your input, the better the output. That's true for any tool. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance --- ## AI Handoff Document Prompts: 10 Templates That Work URL: https://dontreplace.me/articles/ai-handoff-document-prompts-10-templates-that-work ai handoff documenthandoff promptsproject handoffwork transitionai at work AI Handoff Document Prompts: 10 Templates That Work 10 copy-paste AI handoff document prompts for project transfers, vacation coverage, role transitions, and more. Includes a reusable formula and safety rules. July 04, 2026 Updated July 04, 2026 Most handoffs fail before the person leaving even logs out. They fail because the outgoing person writes three bullet points in a Slack message, says "lmk if you have questions," and disappears. The incoming person then spends two weeks reverse-engineering what happened, asking the wrong questions, and breaking things that seemed fine. AI handoff document prompts won't fix a person who doesn't actually want to document their work. But for everyone else, they're a genuinely useful way to turn "I'll write something up before I go" into a real artifact that covers what the next person actually needs. That's what this article is: 10 copy-paste prompts for the most common handoff situations, plus a reusable formula you can adapt for anything. No inflated productivity claims. No screenshots of pretty docs. Just prompts that work if you put real information in. The one rule that makes or breaks AI handoff documents AI is fast at structure. You dump your context in, it organizes it into sections, flags gaps, asks clarifying questions. That's the value. That's it. The mistake people make is treating the output like a finished document. It's not. A well-formatted handoff with wrong information is worse than a messy one that's accurate. The next person will trust the clean version. They'll act on the wrong details. Something breaks. Rule #13 in Don't Replace Me makes this point bluntly: garbage in, garbage out. AI will organize your vague notes into confident-sounding prose. That confidence is cosmetic. If your input is incomplete or approximate, the output is confidently wrong. So before any prompt: verify your facts. Check that the owner names are current, the dates are right, the status is real, the links work. Then use AI to structure it. Not the other way around. What not to put into AI handoff document prompts This matters enough to say before the templates. Do not paste into an unapproved AI tool: customer PII or personal data, employee performance records, compensation details, login credentials or API keys, unreleased product plans, financial forecasts or board materials, legal disputes or contracts, security incidents, or confidential client strategy. If your handoff involves legal, security, HR, financial, safety, or customer-impacting risk, get a human in the loop. A well-structured document doesn't replace a conversation with your manager or legal team. It supplements it. Check your company's AI policy before you paste anything. If there isn't one, assume the conservative position and leave sensitive material out. The reusable formula for any AI handoff prompt Every handoff prompt in this article follows the same structure. You can use this as a starting point for anything not covered below. I'm creating a handoff document for [role/project/coverage period]. Here is my raw context: [paste your notes, status, and details] Please organize this into: 1. What this is and why it matters 2. Current status (verified as of [date]) 3. Upcoming deadlines and key dates 4. Named owners and their responsibilities 5. Open questions and blockers 6. Risks and assumptions 7. Where to find things (links, tools, files) 8. Escalation path if something goes wrong 9. What the incoming person should do first Flag anything that seems vague, missing, or uncertain. Do not invent details. That last line is the most important one you can add to any AI prompt. It changes the output from "confident and wrong" to "honest about what it doesn't know." This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 copy-paste AI handoff document prompts 1. Vacation coverage handoff For when you're out for a week and someone needs to keep the lights on. I'm going on vacation from [date] to [date] and [name] is covering for me. Here are my notes: [paste]. Please create a coverage handoff that includes: what to monitor daily, what can wait until I'm back, who to contact for what, standing meeting commitments, and anything that could break while I'm gone. Include a "do not touch" list. Flag anything where my notes are unclear. 2. Handing off a project to a new owner For when a project outlives your involvement in it. I'm handing off [project name] to [name] as of [date]. Here's my current context: [paste]. Please create a project handoff document that covers: what this project is trying to achieve, where it stands right now, what decisions have already been made and why, what's still open, who the stakeholders are and how they like to communicate, what the incoming owner should do in their first week, and what I'd be watching if I were staying on. Flag any gaps in my notes. This pairs well with building a decision log before you hand off, so the new owner isn't re-litigating choices you already made. 3. Client context transfer Probably the highest-stakes handoff most people face. The relationship is the asset, and it doesn't transfer through a document alone. But the document helps. I'm transitioning client [name] to [name] as of [date]. Here's what I know about this client: [paste notes on history, preferences, sensitivities, open items, contract context]. Please create a client context brief that includes: relationship history, active work and status, known preferences and communication style, risks and sensitivities, open commitments we've made, upcoming touchpoints, and what the incoming person needs to do before their first call. Do not invent relationship details. Flag anything vague. Note: leave out PII, contract financials, or anything legally sensitive. That context needs to transfer through a direct conversation and appropriate secure channels. 4. Role transition documentation For when you're leaving a role entirely, not just a project. I'm transitioning out of [role title] as of [date]. Here are my notes on my responsibilities: [paste]. Please help me create a role transition document that covers: core responsibilities and how I currently handle them, recurring processes and where they're documented, relationships and institutional knowledge that aren't obvious from the job description, tools and access I manage, things I know that aren't written down anywhere, and what the incoming person will face in their first 30 days. Flag anything that should probably be documented more formally. If you've got recurring processes worth capturing, the SOP prompts are built for exactly this. 5. Support coverage handoff For support, ops, or on-call work where someone needs to pick up exactly where you left off. I'm handing off [support/ops/on-call coverage] to [name] for [time period]. Here's current status: [paste]. Please create a coverage brief covering: open tickets and their current status, priority order for the queue, known issues and workarounds in play, customer situations that need careful handling, escalation contacts by issue type, and what "normal" looks like so the incoming person can spot when something is off. Flag anything I haven't specified clearly. 6. Paused project summary For when a project is going on ice and someone needs to be able to pick it up later, possibly much later. Project [name] is being paused as of [date]. Here's where we got to: [paste]. Please create a re-entry summary that includes: what we set out to do and why, what we completed, what we decided along the way and why, what we explicitly didn't do and why, what would need to be re-evaluated before restarting, open questions we never resolved, and what the first three things are that someone picking this up would need to do. Assume the person reading this will have no memory of the project. 7. Launch handoff For handing off something that's just gone live to the team running it long-term. We just launched [product/feature/campaign] on [date]. I'm handing ongoing ownership to [name/team]. Here's context: [paste]. Please create a post-launch handoff that covers: what we launched and what it's supposed to do, known issues and current status, monitoring responsibilities and what to watch, who to contact if something breaks, decisions made during launch that affect how this should be run, and what the first 30 days of ownership look like. Flag anything that looks like a risk I haven't addressed. For the risk side of launches, the risk assessment prompts are worth running before you hand off. 8. Vendor management handoff Vendor relationships have more hidden context than almost anything else. Pricing history, relationship quirks, renewal dates, past disputes. This one is worth being thorough. I'm handing off vendor management for [vendor name] to [name] as of [date]. Here's what I know: [paste]. Please create a vendor handoff that covers: what we use this vendor for, current contract status and upcoming renewal dates, pricing context and any negotiated terms, relationship history and how to work well with the contact, open issues and SLAs to monitor, and what the incoming owner needs to do before the first vendor call. Leave out contract financials until they can be transferred through appropriate channels. 9. Extracting open questions before you leave Sometimes you don't even know what you don't know. This prompt helps you find the gaps before someone else finds them the hard way. I'm handing off [project/role/coverage] and I want to make sure I haven't missed anything important. Here are my notes: [paste]. Please review this and generate a list of open questions the incoming person is likely to have, things I haven't specified that could become blockers, assumptions I seem to be making that aren't stated explicitly, and information that looks incomplete or out of date. I'll answer each one before finalizing the handoff document. This is underused and it's probably the highest-value prompt in this list. Running it before you think you're done will almost always surface something. 10. Tailoring one handoff for two audiences The same project means different things to an executive and to the person actually running it day to day. Here is a project handoff document: [paste]. Please create two versions of this: one for [executive/senior leader] that focuses on status, risk, decisions needed, and business impact (no operational detail), and one for [operator/team member] that focuses on what to do, how to do it, what tools to use, and who to ask when something goes wrong. Keep both factually consistent. This is one of those places where AI genuinely saves time. Writing two versions of the same thing manually takes forever. Getting a draft of both and editing from there is much faster. What AI still can't do in a handoff Structure is not continuity. A well-formatted document doesn't transfer a relationship, institutional knowledge, or judgment. It gives the incoming person a faster on-ramp. The conversation still has to happen. For any handoff involving real stakes, the document is the prep work. You still need to talk. The incoming person needs to ask questions you didn't anticipate. The outgoing person needs to answer them. A document that's never discussed is a document that sits in a folder until something goes wrong. The delegation prompts cover the handoff conversation itself if you need help structuring that side of it. And if you're handing off something complex, your project management prompts can help you get the underlying documentation in shape before you write the handoff. One more thing: AI will smooth over gaps in your notes without necessarily telling you. Ask it to flag what's missing. Then actually go fill those gaps before you finalize anything. The handoff document is not the handoff. It's what makes the handoff possible. Frequently asked questions What should an AI handoff document include? A solid handoff document should include: current status verified as of a specific date, named owners for each area, upcoming deadlines, open questions and blockers, where to find relevant files and tools, known risks and assumptions, decision history, and an escalation path if something breaks. AI can help you structure all of this, but the factual accuracy is on you. Can I use ChatGPT or Claude to write a handoff document? Yes, both are useful for structuring handoff notes. Paste your raw context in, ask it to organize by category, and ask it to flag what's missing. The critical rule: don't paste customer PII, credentials, confidential financials, or anything your company's data policy restricts. And always verify the output against your actual notes before sending it to anyone. What's the biggest mistake people make with handoff documents? Treating AI-generated structure as factual accuracy. AI will take vague or incomplete notes and produce a clean, confident-sounding document. That doesn't mean the information is right. The most common failure mode is a handoff that looks thorough but contains assumed details, wrong dates, or missing context. Verify everything before the document leaves your hands. How long should a handoff document be? Long enough to cover what the incoming person needs to get to speed without a safety net, short enough that they'll actually read it. For a week's vacation coverage, a one-pager is fine. For a role transition, expect several pages across multiple sections. The test: could someone who knows nothing about this work pick it up from your document alone? If not, keep writing. When does a handoff document need legal or HR review? Any handoff that touches employee data, client contracts, financial commitments, security systems, legal disputes, or compliance obligations should be reviewed by the appropriate team before it's distributed. A well-structured document doesn't change the legal or HR requirements around information handling. If you're unsure, ask before you finalize. See how to use AI at work for a broader guide to AI and workplace data. How is an AI handoff prompt different from an AI status report prompt? A status report tells stakeholders where something stands right now. A handoff document tells the next person everything they need to keep the work moving without you. Status reports are periodic updates. Handoffs are one-time transfers of ownership and context. The status report prompts are built for the former. These are built for the latter. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Automation Prompts: Automate the Boring Stuff 10 copy-paste AI automation prompts for mapping workflows, writing SOPs, adding approval gates, and QA testing. Practical templates with safety rules included. ai automation promptsworkflow automationSOP templates AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts --- ## AI Hypercare Plan Prompts: Support Launches After Go-Live URL: https://dontreplace.me/articles/ai-hypercare-plan-prompts-support-launches-after-go-live AI hypercarepost go-live supportlaunch supportcustomer supportimplementation AI Hypercare Plan Prompts: Support Launches After Go-Live 10 copy-paste AI hypercare plan prompts for post go-live support, ticket triage, stakeholder updates, and launch cleanup. July 31, 2026 Updated July 31, 2026 The most dangerous part of a launch is the part after everyone posts the confetti emoji. The deploy worked. The checklist went green. The kickoff call sounded calm. Then real users show up with real browsers, real data, real habits, and a remarkable talent for finding the one workflow your team tested for seventeen seconds. That is where AI hypercare plan prompts help. Not because AI can babysit your launch for you. It cannot. But it can turn go-live notes, ticket patterns, monitoring links, defect lists, owner rosters, and customer feedback into a support plan that does not depend on one exhausted project manager remembering everything while pretending Slack is fine. If you are still before launch, start with AI launch checklist prompts or AI smoke test prompts. If the thing is already live and your job is to keep it from quietly catching fire, this article is for the hypercare window: the first days or weeks after go-live when support, product, engineering, customer success, and operations need one shared view of what is happening. AI can structure hypercare work. A human still owns the facts, the customer impact, the tradeoffs, and the final decisions. What a hypercare plan actually is A hypercare plan is the temporary operating system for a launch after it goes live. It answers boring questions that become extremely exciting when nobody knows the answer: Who watches which dashboard? Who triages incoming tickets? What counts as urgent? Which customers need proactive communication? What bugs are acceptable for now, and what bugs trigger escalation? Who can approve a rollback, workaround, patch, or customer notice? When does hypercare end? Without a plan, hypercare becomes vibes with a spreadsheet. Support sees patterns before product does. Engineering gets vague bug reports. Customer success hears about angry customers through interpretive dance. Leadership asks for updates that require someone to scrape five systems manually. Everyone works hard. Nobody has the same picture. AI can help build that shared picture faster. It can summarize sanitized ticket themes, draft triage rules, turn daily notes into stakeholder updates, create owner matrices, and flag missing escalation paths. Useful. Very useful. What it cannot do is verify production health, inspect systems it cannot access, understand confidential customer nuance by default, approve customer-impacting decisions, or magically know whether a weird ticket is an isolated annoyance or the first pebble before the avalanche. That split matters. It is the same basic idea behind what AI can and can't do: use AI for structure, speed, and pattern-finding. Do not use it as a substitute for judgment. The reusable AI hypercare prompt formula Use this formula whenever you are asking AI to help with post go-live support: "You are a hypercare planning assistant. I am [your role] supporting [launch/change/project] after go-live. Here is the sanitized context: [scope, dates, users, systems, known issues, ticket themes, dashboards, owners, escalation paths]. Create [specific output]. Flag missing information, assumptions, risks, and items requiring human approval. Do not invent facts, customers, severity, owners, production status, or decisions." That last sentence is not decorative. AI will often make a messy support situation sound more certain than it is. Your job is to force it to separate evidence from assumptions. Before you paste anything, sanitize it. Do not paste customer PII, credentials, access tokens, raw production logs, private tickets, unreleased strategy, legal disputes, security vulnerabilities, financial records, HR details, regulated data, medical information, private client conversations, or sensitive personal information into unapproved AI tools. Use approved systems and privacy-safe summaries. If the prompt requires raw customer data to work, the prompt is badly designed. Fix the input instead of hoping the robot pinky-promises confidentiality. What to bring before asking AI for a hypercare plan AI is not a mind reader. It is a very fast organizer of whatever you provide. Garbage in, board-ready garbage out. Bring the strongest sanitized inputs you can: InputWhy it mattersHuman check Launch scopeKeeps support focused on what changedConfirm it matches the actual release Go-live date and hypercare windowDefines urgency and coverageSet start, end, and review checkpoints Affected users or customersShows who may feel impactSeparate internal, external, VIP, and regulated groups Monitoring links or dashboard namesAnchors the plan in evidenceUse approved links and real owners Ticket themesHelps spot patterns earlySummarize safely; do not paste raw private tickets Known defectsPrevents duplicate panicLabel severity, owner, workaround, and status Support scheduleAvoids coverage gapsInclude time zones and backup owners Escalation pathsSpeeds decisionsConfirm who can approve fixes, comms, and rollback Success metricsDefines when hypercare can endUse measurable signals, not "seems stable" Communication approvalsPrevents rogue customer updatesKnow who reviews sensitive messages If you do not have these, ask AI to make a missing-information checklist first. Do not ask it to produce a fake hypercare plan from air. Air has terrible incident response. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI hypercare plan prompts you can use today Prompt 1: Turn launch notes into a hypercare plan You are a hypercare planning assistant. I am supporting this launch after go-live: [sanitized launch summary]. Here are the launch notes, known risks, affected users, systems, support channels, owners, and monitoring links: [paste details]. Create a hypercare plan for the first [time period]. Include workstreams, owners, daily routines, check-in cadence, dashboards to watch, ticket triage rules, escalation triggers, and open questions. Flag anything that needs human approval or verified evidence. What to check after: Replace every placeholder with a real owner, dashboard, channel, and date. A hypercare plan with "TBD" everywhere is just a haunted template. Prompt 2: Map owners and escalation paths You are a hypercare planning assistant. Here are the teams, roles, support channels, and known issue categories for this post go-live period: [paste sanitized list]. Create an owner and escalation matrix. For each issue type, include primary owner, backup owner, response target, escalation trigger, approval needed, and communication channel. Mark missing owners as gaps instead of guessing. What to check after: Team names are not owners. "Engineering" cannot answer a phone, approve a workaround, or join a customer call. Put actual names or rotations where your organization allows it. Prompt 3: Summarize ticket themes safely You are a hypercare support analyst. Here are privacy-safe summaries of tickets from the first [time period] after go-live. Do not infer private customer details. Group the tickets into themes, likely affected workflows, severity levels, repeated symptoms, potential duplicates, and follow-up questions. Separate confirmed facts from assumptions. Do not invent root causes. Ticket summaries: [paste sanitized summaries] What to check after: AI is good at grouping repeated symptoms. It is not allowed to decide root cause just because three tickets use the same sad adjective. Confirm with support and engineering. For cleaner bug documentation, pair this with AI bug report prompts. Prompt 4: Define severity and triage rules You are a hypercare triage assistant. We need clear severity rules for post go-live issues. Context: [sanitized launch context]. Known risks: [risk list]. Affected users: [user groups]. Draft severity levels from Sev1 to Sev4. For each level, include definition, examples, response target, escalation path, customer communication needs, and examples of what does not qualify. Flag where leadership, legal, security, support, or customer success approval may be required. What to check after: Severity rules are political, not just operational. A bug affecting one strategic customer may need more attention than a bug affecting many low-risk internal users. Context wins. Prompt 5: Create the daily stakeholder update You are a hypercare communications assistant. Turn these sanitized post go-live notes into a concise daily stakeholder update. Include: overall status, key metrics, new issues, resolved issues, open risks, customer impact, decisions needed, owner list, next 24-hour plan, and any blocked items. Use plain language. Do not overstate certainty. Separate confirmed facts from assumptions. Notes: [paste sanitized notes] What to check after: Remove panic theater and success theater. Stakeholders need the truth: what changed, what matters, who owns it, and where decisions are needed. If you want a broader format, AI stakeholder update prompts can help. Prompt 6: Check monitoring and support coverage You are a hypercare readiness reviewer. Here is our post go-live coverage plan: [paste sanitized plan]. Review it for gaps in monitoring, support channels, time zones, owner backups, escalation paths, customer communication, dashboard access, defect logging, rollback readiness, and end-of-day reporting. Return a table with gap, why it matters, owner to confirm, suggested fix, and urgency. What to check after: AI can find obvious missing rows. It cannot know whether the named dashboard is actually working or whether the backup owner is on vacation in a cabin with one bar of service. This pairs well with AI operational readiness prompts when the launch still feels like a group project held together by calendar invites. Prompt 7: Convert customer feedback into follow-up actions You are a hypercare action-planning assistant. Here are privacy-safe summaries of customer feedback after go-live: [paste sanitized summaries]. Convert them into follow-up actions grouped by bug, training need, documentation gap, product improvement, communication issue, and unresolved question. For each action, include suggested owner, evidence needed, customer-impact level, urgency, and recommended next step. Do not include private customer details. What to check after: Do not let AI flatten customer pain into generic product chores. A confused user, a blocked user, and an angry enterprise buyer are different situations. The spreadsheet may look similar. The consequences are not. Prompt 8: Compare expected vs. actual launch behavior You are a hypercare analysis assistant. Here is what we expected after go-live: [expected behavior, metrics, ticket volume, support flow]. Here is what actually happened: [sanitized observations]. Compare expected vs. actual behavior. Group findings into: on track, watch closely, investigate, escalate, and needs decision. For each item, include evidence, uncertainty, owner, and next action. Do not invent missing metrics. What to check after: This prompt is useful because launches often fail sideways. The main feature works, but onboarding slows down. Tickets are low, but only because users gave up. Metrics are green, but support is hearing smoke. Treat AI's comparison as a draft hypothesis map, not a verdict. Prompt 9: Decide whether to extend or sunset hypercare You are a hypercare planning assistant. We need to decide whether to end, extend, or change the hypercare period for [launch/change]. Here are the sanitized metrics, ticket themes, unresolved defects, customer-impact notes, support load, monitoring status, and stakeholder concerns: [paste details]. Create a decision brief with options: end hypercare, extend hypercare, narrow hypercare, or escalate. Include criteria, evidence, risks, owners, and decision questions. Do not make the decision for us. What to check after: The best hypercare plans have an exit ramp. Otherwise the temporary process becomes permanent duct tape. If unresolved launch risk is serious, connect this to AI go/no-go decision prompts or AI rollback plan prompts. Prompt 10: Write the lessons-learned handoff You are a post-launch handoff assistant. Turn these sanitized hypercare notes into a lessons-learned handoff for the team that will own normal operations. Include: launch summary, issues found, fixes shipped, open defects, customer-impact notes, monitoring changes, support playbook updates, documentation gaps, process improvements, owners, and follow-up dates. Keep blame out of it. Flag missing evidence and unresolved decisions. What to check after: A good handoff prevents the same launch from failing twice. Do not bury known defects in narrative prose. Put them where owners, dates, and follow-ups can survive someone closing the project folder. For deeper review after something went wrong, use AI incident review prompts. Hypercare is not a substitute for incident review when users were materially affected. A simple daily hypercare rhythm If you do nothing else, run hypercare on a repeatable daily rhythm. AI can help draft each part, but humans need to verify every signal. Morning scan: Review dashboards, tickets, support channels, customer success notes, and known defects. Triage: Group issues by severity, customer impact, owner, and next action. Decision check: Identify anything that needs escalation, customer communication, workaround approval, or rollback review. Stakeholder update: Send one concise status note with facts, risks, decisions, and next steps. End-of-day cleanup: Update the defect list, action log, owners, unresolved questions, and tomorrow's focus. That rhythm sounds basic because basic is what keeps launches alive. Most post-go-live chaos is not caused by a lack of advanced frameworks. It is caused by missing owners, unclear severity, duplicated tickets, dashboards nobody checks, and the ancient corporate ritual of assuming silence means stability. Silence does not mean stability. It may mean users are confused, support is overloaded, or your alerting is decorative. Where AI gets hypercare wrong AI-generated hypercare plans often look confident in exactly the places they should be cautious. Watch for these failure modes: Invented certainty: "No customer impact identified" when nobody checked. Fake owners: Team names, roles, or placeholders presented as accountability. Generic severity: Every issue gets classified by template instead of context. Missing privacy boundaries: Raw tickets, logs, or customer details get pasted into unsafe tools. Decision laundering: AI writes a recommendation that people treat as approval. Support theater: Pretty updates hide unresolved defects and unclear next steps. Premature sunset: Hypercare ends because the calendar says so, not because evidence says the launch is stable. Your defense is simple: require evidence, owners, and human approval. Ask AI to flag uncertainty instead of smoothing it over. A useful hypercare prompt should produce phrases like "missing evidence," "owner needed," "severity requires confirmation," and "customer impact unclear." Annoying? Yes. Better than discovering the gap in front of a customer? Also yes. The human work AI cannot replace AI can help you move faster through the paperwork of hypercare. It can structure updates, summarize patterns, clean up action logs, and make repeated support routines less miserable. But the actual value is still human: knowing which customer signal matters; deciding whether a workaround is acceptable; protecting private data; weighing support burden against product risk; communicating honestly without causing unnecessary panic; telling leadership when the launch is not as stable as the dashboard says; deciding when to extend hypercare, escalate, rollback, or move to normal operations. That is the work. Not typing the prettiest update. Not generating the longest checklist. The work is judgment under incomplete information. Use AI to make the boring parts faster so you have more attention left for the parts where being an adult is unfortunately still required. Frequently asked questions What is an AI hypercare plan prompt? An AI hypercare plan prompt is a reusable instruction that helps an AI tool organize post go-live support work. You give it sanitized launch scope, known issues, ticket themes, owners, support coverage, dashboards, and escalation rules. It turns that into a plan, checklist, update, or decision brief. The prompt does not replace the people responsible for the launch. It just makes the messy support layer easier to see. How long should hypercare last after go-live? Long enough to prove the launch is stable with evidence. For a small internal workflow, that might be a few days. For a customer-facing migration, it might be several weeks. The better question is not "How long did the calendar say?" It is "Have support load, defect volume, customer impact, monitoring, and unresolved risks returned to a normal operating level?" If not, ending hypercare is optimism in a costume. What should I include in a ChatGPT hypercare prompt? Include sanitized launch scope, affected users, go-live date, known defects, privacy-safe ticket themes, dashboards or metric names, support channels, owners, severity rules, escalation paths, rollback criteria, communication approvals, and the output you want. Also tell the AI to flag missing information and not invent facts, owners, customer impact, production status, or decisions. What should I never paste into AI during hypercare? Do not paste customer PII, credentials, access tokens, raw production logs, private tickets, security vulnerabilities, legal issues, financial records, HR details, regulated data, medical information, private client conversations, or sensitive personal information into unapproved AI tools. Use summaries, masked data, approved tools, and your company's actual data policy. The launch being stressful does not make privacy optional. Rude, but true. Can AI decide whether to end hypercare? No. AI can draft a decision brief that compares evidence, open defects, support load, customer impact, and risks. It can show options like end, extend, narrow, or escalate. A human owner still has to make the decision because ending hypercare is an accountability call, not a formatting task. Is hypercare the same as post-launch monitoring? Not exactly. Post-launch monitoring focuses on signals: dashboards, alerts, tickets, metrics, and unusual behavior. Hypercare is the broader operating plan around those signals: who watches them, who triages, who communicates, who fixes, who escalates, and when the team returns to normal support. Use AI post-launch monitoring prompts for the signal layer and this hypercare plan for the operating layer. Final rule: hypercare is not a victory lap Hypercare is where launch promises meet reality. Use these ChatGPT hypercare prompts to create structure: owners, triage rules, daily updates, ticket themes, support coverage, and exit criteria. Then verify everything. Check the dashboards. Read the tickets. Talk to support. Ask customer success what they are hearing. Confirm engineering has the right evidence. Make the hard calls yourself. If you want the bigger survival guide for staying useful while everyone else mistakes tidy AI output for competence, grab Don't Replace Me by Dmitry Kargaev. It is about exactly this kind of work: using AI as leverage without outsourcing the judgment that makes you worth keeping. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance --- ## AI Incident Review Prompts: 10 Templates URL: https://dontreplace.me/articles/ai-incident-review-prompts-learn-without-blame ai incident reviewincident review promptspostmortem promptsAI at workincident management AI Incident Review Prompts: 10 Templates 10 copy-paste AI incident review prompts for timelines, contributing factors, executive summaries, and follow-up actions. Includes safety warnings. July 06, 2026 Updated July 06, 2026 The incident happened. Now you need to explain it to six stakeholders, find out what actually went wrong, make sure it doesn't happen again, and do all of this without starting a war inside your own team. That's the job of an AI incident review, and it's one of the messiest things people try to use AI prompts for. AI can help. It can also turn a real learning opportunity into a polished document that explains nothing, reassigns blame to "process failures," and makes everyone feel like they did a great retrospective when they actually just laundered the problem. This guide gives you 10 copy-paste AI incident review prompts, a reusable formula, and the warnings you need to not make things worse. Why AI incident review prompts are actually useful An incident review has a few distinct jobs: build an accurate timeline, separate what you know from what you're guessing, figure out what actually caused the problem, assign clear follow-up actions, and communicate clearly to different audiences. Each of those is a different document task, and AI is genuinely good at document tasks. What AI is not good at is deciding which detail matters, knowing when someone's explanation is self-serving, or telling you that the clean narrative in your notes is hiding a real risk. That part is yours. Rule #7 in Don't Replace Me calls it the taste moat: knowing which lesson is real and which one is convenient. AI doesn't have taste. You do. Use the prompts below to handle the structure and drafting. Keep the judgment. The reusable AI incident review prompt formula Before the 10 prompts, here's the formula behind all of them. Every good incident review prompt gives AI four things: Your role and context. What kind of incident was this? What's your role in the review? The raw material. Notes, timelines, Slack threads, meeting summaries. The more specific, the better. The output format. What do you actually need: a table, a narrative, a summary, a draft? The constraints. What should it avoid: assigning blame, speculating beyond the facts, inventing causes? Bad input gives you confident nonsense. That isn't just a data science cliche, it's the entire story of bad AI incident reviews. If you want a plain-language breakdown of what AI is actually good at versus where it confidently makes things up, what AI can and can't do covers that clearly. A critical safety note before you paste anything Do not paste the following into unapproved AI tools: Customer names, emails, account numbers, or any personal data Employee records, HR conversations, or performance data Security incidents, credentials, tokens, logs with secrets Legal disputes, contracts, board materials Unreleased product details or financial forecasts Medical or safety information Confidential client information If your incident involves any of these, stop and involve the right people first. Legal, security, HR, finance, customer leadership, and incident response teams exist for exactly this situation. AI prompts are for organizing information you're already allowed to share, not for bypassing the conversations that protect you and your company. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI incident review prompts you can use today Prompt 1: Build an incident timeline You are helping me structure an incident review. I'm going to paste my raw notes from [incident name/date]. Your job: extract a chronological timeline of events. For each entry, include the timestamp (if available), what happened, who was involved, and the source of the information (for example, "noted by [role], Slack message, monitoring alert"). Mark any entry where the timing is unclear or the source is a single person's recollection. Do not infer causes or assign responsibility. Only organize what's explicitly stated in the notes. My notes: [paste your notes here] Prompt 2: Separate facts from assumptions Review the incident notes I'm pasting below. Sort every statement into one of three categories: - CONFIRMED FACT: directly observed, logged, or documented - REASONABLE ASSUMPTION: plausible given the evidence but not confirmed - SPECULATION: not supported by evidence in these notes For each assumption or speculation, flag what additional information would confirm or rule it out. Notes: [paste your notes here] Prompt 3: Write a neutral incident summary Write a neutral, factual summary of the following incident for an internal review document. Tone: clear and professional, no blame assigned to individuals. Describe what happened, when, what was affected, what the impact was, and what initial response steps were taken. Do not speculate about root cause in this summary. Do not use passive voice to obscure who made decisions. If a decision was made, say a decision was made and by which team (not individual names unless relevant and confirmed). Source material: [paste your notes here] Prompt 4: Identify contributing factors Based on the incident summary below, help me identify contributing factors. These are conditions that made the incident possible or worse, not a single root cause. For each factor, note: - What category it falls into (process, tooling, communication, environment, knowledge gap, etc.) - Whether this factor is confirmed by evidence or inferred - What evidence supports it Avoid framing this as "who caused what." Frame it as "what conditions existed." Flag any factor that requires more investigation before it can be confirmed. Summary: [paste your clean summary here] Prompt 5: Turn messy notes into a review brief I have messy notes from an incident that happened on [date]. I need to turn them into a structured review brief for [audience: your team / leadership / cross-functional group]. The brief should include: - Incident title and date - Severity (I'll fill this in) - What happened (2-3 sentences) - Who was affected and how - Timeline (condensed, major events only) - Contributing factors (preliminary) - What was done to resolve it - Open questions - Proposed next steps Flag anything that needs to be verified before the brief goes out. Do not invent details I haven't provided. My notes: [paste here] For help structuring what goes into a brief, the AI status report prompts guide has a useful set of templates for turning incident updates into clear stakeholder communication. Prompt 6: Draft follow-up actions Based on the incident review below, help me draft a follow-up action table. For each action item, include: - The specific action (concrete, not vague) - The team or role responsible (not vague ownership like "everyone" or "the team") - The deadline - How we'll know it's done (success criteria) - Whether this is a short-term fix or a longer-term systemic change If an action is vague in the source material, rewrite it as a specific task or flag it as needing more definition before it can be owned. Source material: [paste your notes or initial action list here] Prompt 7: Prepare a customer-safe update I need to write an external update for affected customers about an incident that occurred on [date]. Write a clear, honest, professional message that: - Acknowledges what happened (without technical jargon) - States the impact on customers - Explains what was done to resolve it - Describes what we're doing to prevent recurrence - Avoids: blaming individuals, overpromising, speculation about causes, and language that minimizes the impact Keep it under 200 words. Flag any claim that I need to verify before this goes out. Incident summary: [paste your neutral summary here] For drafting difficult customer communications around incidents, the AI client communication prompts guide covers how to write the harder messages clearly. Prompt 8: Create an executive summary Write a short executive summary of the following incident for senior leadership. Length: under 150 words. Include: - What happened and when - Business or customer impact (quantified if the data is in my notes; do not estimate) - Root causes (preliminary, flagged as unconfirmed if not yet verified) - What was done - Key decisions made and by whom (team level) - What needs a decision or resource from leadership Do not include speculative causes. Do not soften language about impact. Write for people who don't have context on the technical details. Source material: [paste brief or notes here] Prompt 9: Check whether action items are real Review the following list of action items from our incident review. For each item, evaluate: - Is this action specific enough to be owned and completed, or is it a vague intention? - Does it have a clear owner? - Does it have a deadline? - Is it addressing a root cause or contributing factor, or is it theater (an action that sounds useful but won't change the conditions that caused the incident)? Flag any item that needs to be rewritten or challenged before the review closes. Action items: [paste your current action list here] This is the prompt most teams skip. It's also the most valuable one, because bad incident reviews close with action items that are theatrical. "Improve documentation" assigned to no one with no deadline is not an action item. It's a way to end the meeting. If you want to track follow-through systematically, the AI risk assessment prompts guide covers how to map what happens if the actions don't get done. Prompt 10: Plan a 30-day follow-through review I want to schedule a 30-day follow-through check on the action items from an incident that occurred on [date]. Draft a short agenda for a 30-minute follow-through review meeting. Include: - A check-in on each action item (complete, in progress, blocked, not started) - A review of any recurrences or related issues since the incident - An evaluation of whether the contributing factors have changed - Time to confirm or close open questions from the original review - A decision point: is this incident fully closed, or does it need to stay open? Also draft a brief reminder message I can send to action owners two weeks before the review. Action items from original review: [paste here] Original incident date: [date] What AI incident review prompts can't do Here's the short version: AI can organize your information. It can't tell you whether the information is complete, honest, or the right frame for what actually happened. It will confidently produce a contributing factors list from vague notes. That list will feel authoritative. It probably won't be. Speed is not the same as truth, and the gap between the two is where bad incident reviews live. The actual work of an incident review is human. Someone needs to decide whether the action items are real or ceremonial. Someone needs to push back when the narrative is too clean. Someone needs to say "we don't actually know that yet" when the document implies certainty. That's not AI's job. It's yours. For escalation decisions, especially when an incident involves legal risk, security exposure, customer safety, or significant business impact, the AI escalation plan prompts guide covers how to think about who needs to know and when. If the broader question of what AI is actually good at and what it quietly gets wrong keeps coming up in your work, what AI can and can't do is a useful reference for keeping the right level of skepticism without becoming paralyzed by it. Frequently asked questions Can I use ChatGPT or Claude to write an incident review? Yes, with the right input and appropriate skepticism. AI can organize timelines, draft summaries, and structure action items. It can't verify facts, assess blame fairly, or replace the human judgment needed to determine whether a review is honest. Always verify AI output against source documentation before sharing it. What should I never paste into AI when doing an incident review? Never paste customer PII, employee records, security credentials, legal documents, financial forecasts, medical or safety data, confidential client information, or anything from a board or regulatory process into unapproved AI tools. When in doubt, involve legal, security, HR, or your incident response team before using AI. How do I stop an incident review from becoming a blame exercise? Separate facts from assumptions early (Prompt 2 above), use contributing factors instead of root cause framing (Prompt 4), and write summaries in neutral language focused on conditions rather than individuals (Prompt 3). If the review is legally sensitive, involve HR or legal before the meeting, not after. What's the difference between an incident review and a retrospective? An incident review focuses on a specific failure or disruption: what happened, the impact, the causes, and the actions. A retrospective is broader, covering what worked and what didn't across a sprint or project. For retrospective prompts, the AI retrospective prompts guide covers that format specifically. How do I make sure action items from an incident review actually get done? Use Prompt 9 to check that action items are specific, owned, and deadline-bound before the review closes. Schedule the 30-day follow-through using Prompt 10. The most common failure mode is closing a review with action items that are vague, unowned, or not connected to an actual cause. What if the incident involves sensitive HR or legal issues? Stop and get the right people involved before using any AI tool. HR conversations, employee misconduct, legal disputes, and regulatory incidents require proper handling that AI prompts cannot provide. Use these prompts only for the operational and communication parts of an incident review once the sensitive elements have been appropriately contained. The whole point of an incident review is to learn something real and do something about it. AI helps with the drafting. The learning is still yours. If you want a sharper map for staying useful when everything is moving fast, Don't Replace Me is the field guide for exactly that, including why the human who can tell real lessons from polished ones is the one who keeps their seat. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Automation Prompts: Automate the Boring Stuff 10 copy-paste AI automation prompts for mapping workflows, writing SOPs, adding approval gates, and QA testing. Practical templates with safety rules included. ai automation promptsworkflow automationSOP templates AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts --- ## AI Killing Entry-Level Jobs: What to Do Now URL: https://dontreplace.me/articles/ai-is-killing-entry-level-jobs-what-new-grads-should-do entry-level jobsai job replacementnew gradscareer adviceai fluency AI Killing Entry-Level Jobs: What to Do Now Entry-level postings are down 35%. AI is reshaping the job market for new grads. Here's what's actually happening and what to do about it. March 28, 2026 Updated 12 May 2026 Entry-level job postings dropped 35% between 2022 and 2024. Not because the economy collapsed. Because AI got good enough to do the work those roles existed to do. AI killing entry-level jobs isn’t a prediction anymore. It’s already happened in enough sectors that new graduates are feeling it in real time. If you just graduated, or you’re about to, this is the thing nobody at your commencement ceremony mentioned. The entry-level jobs that were supposed to be your on-ramp? A lot of them are gone. And the ones that remain are being competed for by more people, paying less, and demanding more. This isn’t doom. It’s just the situation. And if you know what’s actually happening, you can do something about it. How AI killing entry-level jobs actually plays out The data is pretty clear. A 2024 analysis by Anthropic found that AI tools disproportionately affect early-career workers, the people who spend most of their time on the tasks AI does cheapest and fastest. Research writing, data formatting, basic code, first-draft copy, simple analysis. The work that used to take a junior hire a week, a senior person can now do in an afternoon with Claude. The World Economic Forum’s Future of Jobs Report puts it plainly: employers expect 39% of core skills to change by 2030. That’s not a rounding error. That’s almost half of what companies currently hire for becoming either less relevant or replaceable. Some CEOs are already saying it out loud. Duolingo cut contractors. IBM paused hiring for roles AI could handle. Klarna replaced 700 customer service workers with an AI agent and announced it publicly. These aren’t flukes. They’re signals about where the cost-cutting math leads when AI gets cheap enough. Why the traditional advice is actively bad for you right now “Get a degree, build your resume, apply for entry-level roles, work your way up.” This was reasonable advice in 2010. Right now it’s like telling someone to fax their CV. The ladder still exists for some people in some industries. But the bottom rungs are being sawed off. Companies are increasingly using AI to skip the junior layer entirely. A team of five senior people with AI tools can now do what used to require fifteen, including the three junior hires who would’ve learned by osmosis. The old model assumed companies needed bodies at every level. Cheaper bodies to do simpler things, more experienced bodies for the hard stuff. AI broke that assumption. A senior marketer who knows how to prompt Claude is now doing first drafts, research, and reporting herself. She doesn’t need an assistant. The assistant job is gone. Waiting for the market to return to normal is a losing bet. This is normal now. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What’s replacing entry-level: the gap job and the superjob Here’s the part that doesn’t make it into the doom articles. Every time AI gets better at something, a gap opens up between what it outputs and what humans actually need. Those gaps don’t disappear. They become jobs. Messy, ill-defined, weirdly compensated jobs at first. Then real ones. AI can write a product description. It can’t tell you if the tone is right for the brand, if the claim would make the legal team nervous, or if the regional sales team will actually use it. Someone who understands AI output AND has enough domain sense to bridge it to human needs? That person is currently drowning in work and underpaid because there isn’t yet a job title for what they do. That’s the gap job. It’s the space between AI output and human outcome. The superjob is the evolved version. It’s what happens when someone builds enough expertise in a domain, learns AI fluency on top of it, and can now do work that used to require a whole team. A single person who can research, analyze, write, edit, project manage, and ship. These hybrid roles are already emerging, and they pay well because there aren’t many people who can do them yet. Neither of these shows up on a job board. Not clearly, anyway. You have to build your way in. What industries are being hit hardest right now Not every sector is seeing the same damage. Some industries are cutting entry-level roles fast. Others are slower. Knowing the difference matters if you’re deciding where to aim. The hardest-hit categories right now are the ones built on structured, repeatable, computer-based work: Industry Entry-level roles most affected Why Finance Junior analysts, data entry, reporting AI handles the modeling and summarizing Legal Paralegal research, document review LLMs are fast and cheap for first-pass reading Marketing Junior copywriters, social media assistants AI drafts fast; one senior person can edit Tech QA testers, junior developers AI catches bugs and writes boilerplate code Customer service First-line support, ticket routing Chatbots handle volume; humans handle exceptions The common thread: anything that was mostly information processing on a computer. Those tasks didn’t require judgment, physical presence, or deep relationships. They required time and attention. AI has time and attention in unlimited quantities. The less-affected sectors involve physical work, high-stakes human relationships, or decisions with real legal and ethical weight. Construction, skilled trades, social work, medicine, senior legal counsel. Not immune, but slower-moving targets. If you’re choosing a field right now, that table is worth thinking about. The playbook: build, don’t apply Sending out 200 applications is not a strategy. It’s a slot machine. And right now, the slot machine is mostly broken. What works instead is building visible evidence that you can do the thing. This is Rule #14 from Don’t Replace Me, the field guide for exactly this situation: ship something ugly. Not polished. Not perfect. Visible. A project you did with AI tools. A case study of a problem you solved. A newsletter you started. A spec work piece you built for a company that didn’t hire you. The work doesn’t need to be impressive. It needs to prove you’re capable of doing instead of just studying. Hiring managers who are drowning in AI-polished resumes that all look identical respond to people who’ve actually built things. For people who are worried about whether AI will take their career entirely, the honest breakdown of what’s actually at risk is worth reading before you panic. What you should actually learn right now AI fluency. Not coding. Not a $997 prompt engineering course. Actual working familiarity with the tools. Pick one: Claude or ChatGPT. Use it every day for the next 30 days on real problems in your field. Not “what is photosynthesis” nonsense. Real problems. Draft this cover letter. Analyze this dataset. Summarize these 40 pages. Write a first draft of this report. The goal is to develop judgment about when AI helps and when it hallucinates, what to trust, and what to check. That judgment is an actual skill that non-technical people can build without a computer science degree. Most people don’t have it yet. That’s the gap you can be in. Here’s a rough stack for someone starting from zero: Skill What it means How to start AI fluency Using LLMs effectively for real work tasks Use Claude/ChatGPT daily for 30 days Domain knowledge Deep understanding of one field Double down on your actual area Output judgment Knowing when AI is wrong Compare AI output to expert opinion repeatedly Human bridge Translating AI output for non-technical people Practice explaining, editing, contextualizing Visible portfolio Proof you can build things Ship one project per month, post it publicly The combination of these is what Dee calls the “combination moat” in the book: the intersection of what you know, what AI can do, and the human judgment that holds it together. AI can’t replicate the full stack, because it can’t tell when its own output is wrong in context. The honest conversation about what’s actually hard here Some roles are genuinely going away. Not all of them. Not most of them. But entry-level paralegal work, basic data entry, first-pass content writing, simple QA testing, junior financial analysis. If your plan was to get a foot in the door with one of those and work your way up, the door is narrower than it used to be. That’s real, and it sucks, and anyone who tells you it doesn’t is either selling something or not paying attention. The actual data on AI job displacement isn’t as catastrophic as the headlines suggest overall, but for entry-level white-collar work, it’s genuinely significant. Here’s the thing though. The people who adapt now, while the field is still unsettled, have a real advantage. The superjob category is growing fast. There are organizations desperate for people who can work with AI and think critically about its output. Most of them can’t find those people yet because most people are still pretending the old model works. You don’t need to wait for the job market to figure itself out. You can start building toward the gap right now. Five things you can do this week Not eventually. This week. Open a Claude or ChatGPT account and use it for a real work task today. Not a test. A real thing you need done. Do a gap audit in your field. Find three things people in your industry complain about that AI should theoretically be able to help with. Write down why it doesn’t work perfectly. That gap is where you can position yourself. Build one ugly thing. A simple case study. A one-page analysis of something you care about. A spec project for a company you’d like to work for. Post it somewhere public. Stop sending generic applications. For every role you really want, send a personalized note and attach proof of relevant work. Ten targeted applications beat 200 generic ones. Read the room on anxiety. If the fear is blocking you from doing anything, that’s worth addressing directly. There’s a reason AI anxiety is its own problem right now. You can’t build while you’re frozen. The entry-level market is harder than it was three years ago. That’s true. It’s also not the end of your career before it starts. The people who treat this moment as a prompt to adapt, rather than a reason to panic, are going to look back on 2025 as the year they got ahead. The ladder is being rebuilt. You just have to climb what’s there instead of waiting for the old version to reappear. Frequently asked questions Is AI really killing entry-level jobs, or is this just hype? It’s real, not hype. Entry-level job postings in white-collar fields dropped significantly between 2022 and 2024, with some analyses showing declines of 35% or more. The WEF projects that 39% of core job skills will change by 2030, and early-career workers are disproportionately affected because their work overlaps most with what AI can currently do. It’s not every industry, but it’s enough sectors that it matters. What should new grads do instead of applying for traditional entry-level jobs? Build visible proof of work. One project that shows you can do something real is worth more than a polished resume right now. Combine that with genuine AI fluency, meaning actual daily use of Claude or ChatGPT for real tasks, and you’re ahead of most applicants. Target companies where your specific combination of domain knowledge and AI skills fills a gap. Do I need to learn to code to survive the AI job market? No. Coding helps, but it’s not required for most non-technical roles. What actually matters is AI fluency: knowing how to use the tools, how to judge their output, and how to bridge AI results to human needs. That’s a skill non-technical people can build in weeks, not years. See our breakdown of AI skills for non-technical people. Which entry-level jobs are most at risk from AI? Basic data entry, first-draft content writing, simple financial analysis, entry-level legal research, customer service routing, basic QA testing. Roles that consist mostly of structured, repeatable tasks on a computer are the most exposed. Roles that involve physical presence, complex human relationships, or highly contextual judgment are less at risk. What are "superjobs" and how do I get one? Superjobs are emerging hybrid roles where one person does the work of what used to require a team, using AI as a force multiplier. They combine domain expertise, AI fluency, and human judgment. You get there by building skills across multiple areas, shipping real work publicly, and positioning yourself as someone who bridges AI output and human needs. The guide to future-proofing your career has the full framework. Is it too late to adapt if I'm already a year or two into the job search? No. The window is closing slowly, not overnight. The people getting ahead right now are the ones who start building today rather than next year. A year of real AI fluency and a portfolio of actual work puts you ahead of most people who spent that same year sending applications. Start this week, not next month. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Job Replacement Statistics: What the Data Says Goldman Sachs, WEF, McKinsey: what AI job replacement statistics actually say vs. what the headlines tell you. Real numbers, real context. ai job replacement statisticsai job lossfuture of work AI Skills You Actually Need in 2026 (No Coding Required) If you're not technical, this is where to start. Build real AI fluency and the judgment employers now pay for. AI skillsnew gradsAI fluency New Jobs AI Is Creating (That Don't Have Names Yet) The old ladder is shifting, but new role categories are emerging. Here's where the gap jobs are appearing. new jobsAI careerssuperjobs --- ## AI Job Interview Prompts: 10 Copy-Paste Templates URL: https://dontreplace.me/articles/ai-job-interview-prompts-10-templates-that-work ai interview prepjob interview promptschatgpt interview practiceSTAR methodinterview preparation AI Job Interview Prompts: 10 Copy-Paste Templates 10 copy-paste AI job interview prompts for mock practice, STAR answers, salary talk, and more. No course required. Real templates that work. May 26, 2026 Updated May 26, 2026 Job interviews are a form of narrative work. You're not being evaluated on your resume. You're being evaluated on how well you tell the story of your resume, under pressure, to a stranger who's already tired. AI job interview prompts won't fix a bad story, but they'll help you find the story you already have and practice telling it clearly. This is a copy-paste guide. Ten prompts, each one built for a specific part of the prep process. No $200 course required. The reusable formula every AI job interview prompt needs Before the templates, a quick note on why most people get mediocre results from AI. They type something like "help me with interview questions" and get generic advice that could apply to literally any job in any country. The fix is simple. Every prompt you send needs five things: your role (who you are), context (the job and company), the actual task, any constraints, and the format you want. That's the structure Dee calls the "smart intern" framework in Don't Replace Me, and it's why the same tool gives wildly different output depending on how you use it. So before each session, set the stage. Paste in the job description. Name the company. Describe your background in two sentences. Then ask your specific question. Every prompt below is built on this logic. One more thing before you start: don't paste in anything confidential. No client names, no internal financial data, no private HR information, no details from your employer's systems. Anonymize what you need to and verify every number, title, and date you plan to claim before the actual interview. AI is your rehearsal partner, not your source of truth. AI job interview prompts: the full set These ten prompts cover the complete prep arc, from reading a job description on Sunday night to running a final honesty check the morning before. Use them in order for a full prep session, or dip into whichever one you need. Prompt 1: Analyzing a job description for what they actually want Copy this: I'm preparing to interview for this role. Here is the full job description: [PASTE JD]. I'm a [your background in 2 sentences]. Analyze the job description and tell me: (1) the three most important things this company is actually trying to solve, (2) the skills they'll probe hardest in an interview, and (3) any red flags or unusual requirements I should prepare for. Be specific and direct. This one sounds simple but it saves you from walking into an interview having prepped for the wrong priorities. You might think it's a marketing role. The JD tells you it's actually a sales role with a marketing title. Better to know that now. A useful follow-up is to ask it to compare the stated requirements against what similar job postings at comparable companies include. If this JD is an outlier in what it's asking for, you want to know before you show up. Prompt 2: Generating the questions they'll probably ask you Paste in the job description, your role, and the industry, then run this: Based on this job description and my background, give me the 12 most likely interview questions for this role. Include: 3 behavioral questions, 3 technical or functional questions, 3 situational "what would you do if" questions, and 3 questions that are likely to trip up candidates who aren't prepared. Don't give me generic questions. Tailor them to this specific role and company context. The "trip-up questions" section is the one most people skip. That's the section you actually need. Prompt 3: Turning resume bullets into interview stories This is where prep gets real. If you've already done the work of rewriting your resume bullets with AI, you're ahead. Now you need to turn those bullets into stories you can actually say out loud. Here are three bullet points from my resume: [PASTE BULLETS]. For each one, help me build a STAR story (Situation, Task, Action, Result). Ask me follow-up questions to fill in any gaps. My target role is [ROLE] at [COMPANY/TYPE OF COMPANY]. Keep the stories concise, under 90 seconds when spoken aloud. The key instruction is "ask me follow-up questions." That's what turns AI from a summary machine into an actual practice partner. Prompt 4: Practicing and tightening STAR answers Once you've drafted a story, run it through this: Here is my STAR answer for a question about [TOPIC]: [PASTE YOUR DRAFT ANSWER]. Review it for: (1) clarity, (2) whether the Result is specific enough, (3) any places where I'm being vague or taking too long to get to the point, and (4) whether it actually answers the question. Give me a rewritten version that is tighter, and tell me what you changed and why. The "tell me what you changed and why" part matters. You need to understand the edits, not just paste the cleaned-up version. You're prepping for a live conversation, not writing an essay. Prompt 5: Running a full mock interview This one requires you to stay in character and actually respond out loud (or type your answer, then have it reviewed): You are an interviewer for [COMPANY NAME], hiring for [ROLE TITLE]. I am a [YOUR BACKGROUND]. Run a 20-minute mock interview with me. Ask me questions one at a time. After I answer each one, give me brief feedback: what landed, what was unclear, what I should cut or expand. Start with a warm-up question, then move to behavioral questions, then a few functional/role-specific ones. Ask follow-up questions when my answer is vague. Don't let me off easy. "Don't let me off easy" is not just drama. It actually changes the output. Without it, most AI mock interviews are suspiciously gentle. You don't want gentle. Prompt 6: Challenging your vague answers This pairs with the mock interview but works on its own too. After you've drafted any answer, run this: Here is my answer to the interview question "[QUESTION]": [YOUR ANSWER]. Identify every vague or unsupported claim. For each one, give me the follow-up question a skeptical interviewer would ask. Then tell me what specific evidence I should add to make the answer credible. You'll be amazed how many times an answer contains phrases like "I improved team efficiency" without any actual evidence attached. A good interviewer will probe those gaps. Better to find them now. How to tailor AI job interview prompts for different audiences Not all interviewers ask the same questions for the same reasons. A recruiter screens for baseline fit and red flags. A hiring manager wants to know if you'll make their life easier. A panel wants to know if you'll cause problems. Same you, different framing. I'm preparing for a [recruiter screening / hiring manager interview / panel interview] for the role of [ROLE] at [COMPANY]. Here is my answer to this question: [QUESTION + YOUR ANSWER]. Rewrite or reframe this answer for the specific audience. Tell me what to emphasize differently and what to cut for each context. This is the prompt most interview prep guides don't include, and it's one of the more useful ones. The skills article on what you actually need in 2026 covers audience-reading as one of the human skills AI can't replicate. But it can still help you practice the adjustments. The audience prompt works especially well when you're interviewing at a company that does multiple rounds with different stakeholders. What impresses the recruiter on round one ("I'm a quick learner, I've been in this field for eight years") is often the wrong framing for the VP on round three ("Here's the revenue impact of the last initiative I owned"). Same experience, completely different story. If you're going through a panel interview specifically, add this line to the prompt: "The panel includes a technical lead, a peer from a different team, and the hiring manager. Adjust the emphasis for each listener." The output won't be perfect, but it'll remind you that there are three separate people evaluating you for three different things simultaneously. Prompt 7: Preparing smart questions for the interviewer "Do you have any questions for us?" is not a formality. It's an evaluation. Bad candidates ask nothing or ask about vacation policy. Prepared candidates ask something that shows they've thought about the role. I'm interviewing for [ROLE] at [COMPANY]. The job description mentions [SPECIFIC DETAIL FROM JD]. Based on this context, generate 8 smart questions I could ask the interviewer. The questions should: demonstrate that I've thought carefully about the role, not be answerable from the website's FAQ page, and open a real conversation about challenges or priorities rather than just showing off research. Run the output through your own judgment. Ask only the questions you'd actually want answered. You'll spot the ones that sound clever but feel hollow. A good tell: if the question could just as easily be asked by someone who never read the job description, cut it. Prompt 8: Salary and role-fit language This is the area where people either undersell themselves or make claims they can't back up in follow-up questions. The goal is confident, honest, and specific. I'm preparing to discuss compensation and why I'm a strong fit for [ROLE]. My background: [2-3 SENTENCES]. My target salary range is [RANGE]. Help me write language to: (1) discuss salary without sounding desperate or overclaiming, (2) articulate my fit for the role based on the JD, without using vague phrases like "I'm passionate about" or "I'm a quick learner." Keep it direct and grounded in evidence. The "without vague phrases" constraint is doing a lot of work here. Without it, you'll get back the same empty language you were already planning to use. Prompt 9: A final honesty check Run this before any interview. Seriously. Here are the three main claims I'm planning to make about my background and experience in this interview: [LIST YOUR CLAIMS]. For each one, ask me the hardest follow-up question a skeptical interviewer could ask. Then flag any claim that sounds inflated, unverifiable, or likely to fall apart under questioning. This is the prompt equivalent of a friend saying "are you sure you want to say that?" It won't catch everything, but it'll catch the obvious ones. The AI performance review prompts on this site use the same honesty-check logic, and it's worth running there too. Prompt 10: Full practice loop from JD to mock interview If you want to do a complete prep session in one go, chain these steps together: Step 1: Analyze this JD and tell me the three things they care most about. [PASTE JD] Step 2: Generate the 10 most likely interview questions for this role. Step 3: Run a mock interview with me. Ask me one question at a time and give feedback after each answer. Step 4: After we've finished, give me a summary of: my strongest answers, my weakest answers, and the three things I should work on before the actual interview. You can run this in one chat window. Takes about 30-40 minutes if you actually answer each question instead of skipping to the feedback. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What makes AI job interview prompts actually work The output quality lives or dies on the specificity of your input. Most people give AI about 20% of the context it needs and then complain the results are generic. Here's the exact information worth pasting into every session. The job description. The full thing, not a summary. Copy it straight from the posting. Your actual background. Two or three sentences. Not your LinkedIn headline. Something like: "I've been a project manager in construction for six years. Most of my work involves coordinating between subcontractors and municipal permitting offices. I've never managed a team larger than four people." The company name and any recent news. If the company just announced a major product launch or a round of layoffs, that context shapes what they're hiring for. Mention it. The specific question or scenario you're prepping for. "Help me with interviews" is a request for a lecture. "Help me answer behavioral questions about conflict with a stakeholder, because I have a complicated example from last year that I'm not sure how to frame" is a request for actual help. The no-BS guide to using AI at work covers this input-quality principle more broadly. Interview prep is just one application of it. The pattern is the same everywhere: the quality of what you put in determines the quality of what you get back. One more thing worth knowing: AI interview prep is most useful when your problem is not lack of experience, but messy communication. If you've ever walked out of an interview thinking "I knew the answer to that, I just didn't say it well," this is the tool for you. What AI can't do for you in an interview It can't tell you what actually happened. It doesn't know your numbers, your projects, your mistakes, or your wins. Every specific claim in every answer needs to come from you, checked by you, and defensible by you on the day. It also can't read the room. If the interviewer looks bored at minute three, you need to notice that and adjust. If they lean forward when you mention a specific experience, you need to know to expand on it. That's judgment, not language pattern matching. And it can't manufacture the thing good interviewers are actually looking for: believable evidence. A stumbled-over but genuine story about a real failure lands better than a perfectly constructed non-answer. AI can help you find the story. It cannot be the story. The full picture of human skills that remain genuinely resistant to AI is covered in the jobs-AI-can't-replace breakdown if you want to think about this more broadly. The short version: don't memorize a script. Use these prompts to find your stories and sharpen your language. Then go talk like a person. Frequently asked questions Can I use ChatGPT to prepare for a job interview? Yes, and it's genuinely useful if you use it correctly. The key is to give it real context: your background, the actual job description, and specific questions you want to practice. Generic prompts give generic output. Specific prompts give you tailored mock questions, story feedback, and sharper answers. Is it cheating to use AI for interview prep? No. Using AI to rehearse is the same as practicing with a friend or a career coach. What would be a problem is submitting fabricated claims or memorizing answers that don't reflect your real experience. The stories have to be yours. The AI just helps you find and polish them. What's the best AI job interview prompt to generate practice questions from a job description? Paste the full job description, add your background in two sentences, and ask for: behavioral questions, functional questions, situational questions, and the questions most likely to trip up unprepared candidates. The specificity in the request is what makes the output actually useful. See Prompt 2 above for the full template. How do I make my STAR answers less boring? The most common problem is a Result that isn't specific. "I improved team communication" means nothing. "Response time dropped from 48 hours to 6 hours over three months" means something. Prompt 4 above will flag vague results and push you to add evidence. Run every STAR draft through it before the interview. Should I memorize the AI-generated answers? No. You'll sound like you're reading from a teleprompter, and a good interviewer will notice. Use the prompts to find your stories and tighten your language. Then practice out loud until the answer sounds like you talking, not a polished document. What AI tool is best for interview prep? ChatGPT or Claude. Both work well for this. You don't need a specialized interview prep AI tool. As the practical AI guide for work covers, most of the "dedicated" tools are just wrappers around the same underlying models, often at higher prices. Use the general tool well. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance --- ## AI Job Replacement Statistics: What the Data Says URL: https://dontreplace.me/articles/ai-job-replacement-statistics-what-the-data-says AI job replacementAI statisticsfuture of workautomation dataAI anxiety AI Job Replacement Statistics: What the Data Says Goldman Sachs, WEF, McKinsey: what AI job replacement statistics actually say vs. what gets quoted in headlines. The real numbers, with context. May 09, 2026 Updated 15 May 2026 The Goldman Sachs number gets quoted constantly. "300 million jobs exposed to AI." It shows up in headlines, LinkedIn posts, conference slides, and the mouth of every consultant who charges $500 an hour to explain it to your leadership team. Here's the thing nobody mentions: "exposed to" does not mean "replaced by." In AI job replacement statistics, the gap between those two phrases is where the actual truth lives. This article is about that gap. It's going to give you the real numbers from the real reports, explain what the researchers actually said versus what got quoted in the headline, and leave you with something more useful than panic or blind optimism. Let's start with the data. What the AI job replacement statistics actually say Goldman Sachs published a research note in 2023 that put 300 million full-time jobs "exposed" to automation by generative AI. That number traveled fast. It's still traveling. What the same report said, in the same document, which almost nobody quoted: two-thirds of current jobs are exposed to "some degree" of AI automation. Not full replacement. Not even significant replacement. Some degree. The report also estimated that only about 7% of US workers face the scenario where AI could automate more than half of their tasks. Seven percent. Not 300 million. Seven percent of the workforce facing significant task-level exposure. That's still real. That's still worth taking seriously. But it's a different conversation than the one your LinkedIn feed is having. The other thing the Goldman report includes, buried behind the headline, is a genuine acknowledgment that historical automation waves have consistently created new roles faster than they've eliminated old ones. The researchers weren't writing a horror story. They were writing a productivity forecast. The horror story got written by everyone downstream. The WEF numbers: 85 million lost, 97 million created The World Economic Forum's Future of Jobs Report is the other set of numbers that gets half-quoted constantly. The headline: 85 million jobs displaced by AI and automation by 2025. Terrifying, right? The second half: 97 million new jobs created in the same period. Net positive. Twelve million jobs more than we started with. And this is from the organization most likely to be taken seriously at Davos, not some optimistic think tank trying to sell you on the bright future. Now, the honest caveat: jobs created and jobs displaced don't go to the same people in the same places. A data labeling job in Southeast Asia doesn't help an accounts payable clerk in Ohio. The transition is the hard part. The WEF acknowledges this. The headlines don't. If you want to understand what that transition looks like by industry, there's a full breakdown at what jobs AI will actually replace. McKinsey on tasks vs.jobs: this is the framing that matters McKinsey has been running task-level automation analysis longer than the AI hype cycle has existed. Their most recent estimates suggest that between 2030 and 2060, about half of today's work activities could be automated. That's a huge range and a long timeline, which tells you something about how confident even the best researchers are. More useful than the headline number is the framing underneath it. McKinsey doesn't analyze jobs. They analyze tasks within jobs. A lawyer's job has hundreds of tasks. Some of those tasks (document review, contract comparison, summarizing case law) are highly automatable. Others (advising a client through a divorce, arguing in front of a judge, reading the room in a negotiation) are not. This is why the book's approach matters. Dee Kargaev breaks down the task audit concept in Don't Replace Me, specifically the "find your appendix" framework, which is about identifying which parts of your job are basically appendixes: vestigial, repetitive, and ready to be cut. Those tasks go to AI. You keep the rest. You get faster. You look better. The point is that the McKinsey framing of tasks versus jobs is the correct unit of analysis. And once you use it, the scary number gets a lot less scary. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → The 65% who are quietly panicking EY's workforce survey found that 65% of employees report anxiety about AI replacing their job. That number is real and it tracks with what's happening on search engines, in HR departments, and at 2am when people can't sleep. The anxiety makes sense. The stimulus is everywhere. Every third article, every earnings call, every company announcement about "AI integration" adds a little more cortisol to the system. But here's what the same kind of survey data consistently shows: the people with the highest AI anxiety are also the least likely to have actually used AI tools. Once people start using the tools, the fear tends to drop. Not because AI turns out to be harmless, but because using it makes the threat concrete and manageable instead of abstract and infinite. Abstract threats are the most terrifying kind. If you've never opened ChatGPT for more than five minutes, your brain is filling in the blank with its worst hypothesis. For a practical way to close that gap, the no-BS starter guide to using AI at work is a reasonable place to start. How the statistics get mangled in transit Here's a pattern worth understanding, because it explains why your LinkedIn feed sounds so much worse than the actual research. A research firm publishes a nuanced report with multiple scenarios, careful caveats, and a 60-page methodology section. A journalist reads the executive summary and writes a headline. A content creator reads the headline and writes a thread. A thought leader quotes the thread in a post about why you need to buy their course. By that point, "7% of workers face significant automation exposure" has become "AI will replace 85 million jobs." Every step in that chain is incentivized to make the number bigger and the nuance smaller. Research firms get credibility from being cited. Journalists get clicks from scary headlines. Content creators get followers from alarm. Course sellers get sales from fear. The actual stat, with its caveats intact, is less useful to all of them. It's only useful to you. The panic is a business model. It doesn't have to be your nervous system's operating system. What "exposed to AI" actually means in practice Take a marketing manager. Their job has maybe 40 tasks in it across a week. Writing briefs. Reviewing copy. Running performance reports. Managing vendor relationships. Sitting in strategy meetings. Presenting to the leadership team. Responding to emails. Setting budgets. Some of those tasks are highly exposed to AI assistance. Running a performance report is one. Drafting a creative brief is another. Summarizing a competitor's ad strategy. Writing the first draft of a campaign email. AI can do a version of all of those faster than the marketing manager. But "exposed" means something can assist with it, not that the task disappears or that the person disappears. The marketing manager who uses AI to draft the brief in 10 minutes instead of 60 is now a better marketing manager, not an unemployed one. The people actually at risk are the ones whose entire job is a single automatable task with no variation and no relationship layer. Entry-level data entry. Certain categories of document processing. High-volume, low-judgment repetitive work. That's real displacement risk. But it's not the same as "300 million jobs gone." For a clearer picture of where actual risk concentrates, the breakdown at will AI replace my job goes sector by sector. AI job replacement statistics by sector: where the risk actually concentrates Not all industries are equally exposed, and this is where the aggregate numbers start to fall apart as a guide to anything useful. "Half of work activities could be automated" means something very different for an insurance underwriter than for a pediatric nurse. The Bureau of Labor Statistics tracks automation exposure by occupation. The pattern that emerges is consistent with what the McKinsey and Goldman research shows at the task level: clerical, administrative, and data-processing roles face the highest concentration of automatable tasks. Physical care, skilled trades, and high-judgment professional work face the least. Here's a rough breakdown based on current research: SectorAutomation exposureRealistic timeline Data entry / processingVery highAlready happening Basic customer serviceHigh2-5 years Paralegal / legal researchModerate-high3-7 years Financial analysisModerate5-10 years Marketing / copywritingModerateOngoing, not total Software engineeringModerateOngoing, not total Teaching / coachingLow-moderateLong timeline, partial Healthcare (clinical)LowVery long timeline Skilled tradesVery lowNot on the horizon The key column is the last one. "Already happening" and "2030-2060" are not the same threat. Your career planning for one of those looks nothing like career planning for the other. The other thing worth noting: even in the "very high" category, full elimination is rare. What actually happens first is that one person with AI does the work of three people without it. The headcount shrinks at the edges. It doesn't vanish overnight. What the data says about who actually gets replaced Research from MIT and Boston University that tracked automation over the past few decades found a consistent pattern: the workers who get displaced are disproportionately those doing routine, codifiable tasks, while workers doing non-routine tasks see their productivity and wages go up when automation improves. This pattern has held through calculators, spreadsheets, and digital workflows. There's no particular reason to think it breaks with AI. The distribution shifts. The people who adapt end up better off. The ones who don't get stuck. The honest version of the AI replacement story isn't "robots coming for everyone." It's more like: the gap between people who adapt and people who don't is going to widen. And the adaptations are genuinely not that hard for most people in most jobs. The data on new job categories actually being created by AI is, if anything, underreported relative to the displacement numbers. Check the breakdown of what new roles are emerging if you want the other side of the ledger. How to read AI statistics without losing your mind A few rules for whenever you see an AI jobs number in the wild: "Exposed to" is not "replaced by." Exposure means some task overlap is possible. It says nothing about whether that task disappears or just gets faster. Check the timeline. "By 2060" is a completely different claim than "by 2027." Models with 35-year timelines are basically climate change predictions. Useful for policy, useless for your current career anxiety. Check what's being measured. Jobs, tasks, and roles are different things. A job can have 30% of its tasks automated and the job still exists, just differently. Check who published it. A research firm's peer-reviewed methodology and a course seller's blog post are not the same thing. One has a 60-page appendix. The other has a $997 upsell. Check what was left out of the headline. If Goldman Sachs says 300 million jobs exposed, find the part where they also say only 7% face significant automation. The second number is always there. It just doesn't travel as well. The data is real. The risk is real. The hysteria is optional. Frequently asked questions How many jobs will AI actually replace? No one knows the exact number, but the Goldman Sachs estimate of 300 million "exposed" jobs is widely misquoted. The same report estimates only about 7% of US workers face automation of more than half their tasks. The WEF projects 85 million jobs displaced alongside 97 million new jobs created, a net positive of 12 million. Exposure to AI assistance is not the same as job elimination. What does "exposed to AI" mean in job replacement stats? "Exposed to AI" means some tasks within a job overlap with what AI can do. It doesn't mean the job disappears or that the person gets fired. A marketing manager "exposed" to AI might use it to write first drafts faster. That's a productivity shift, not a pink slip. Which jobs are most at risk from AI replacement? Jobs with high concentrations of routine, repetitive, codifiable tasks face the most genuine displacement risk. That includes certain categories of data entry, document processing, and high-volume, low-judgment administrative work. Jobs with significant relationship, judgment, and physical components are much harder to automate. The full breakdown by industry covers this in detail. Is the 85 million jobs figure from WEF accurate? The WEF's Future of Jobs Report does project 85 million jobs displaced, but the same report projects 97 million new jobs created. The net figure is positive. The displacement and creation don't affect the same workers in the same locations, which is the real challenge, but quoting only the first number without the second is incomplete. Why does AI job replacement data seem so scary in headlines? Because scary travels better than nuanced. A report saying "7% of workers face significant task-level automation exposure" doesn't drive clicks. "AI will replace 300 million jobs" does. Every link in the chain from researcher to headline to LinkedIn thread is incentivized to amplify the alarming number and drop the caveats. Should I be worried about AI replacing my job? Depends on what your job actually is. If most of your work is routine, codifiable, and doesn't require relationships or judgment, that's worth taking seriously. If you have a mix of tasks with some human complexity in it, the more productive question is which parts of your job could go to AI so you can focus on the parts that can't. The jobs AI genuinely can't replace is a good starting point for that assessment. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Anxiety at Work: How to Stop Spiraling 65% of employees feel AI anxiety at work. Here's how to convert that fear into something useful without buying a $997 course or losing more sleep. ai anxietyai job fearwill ai replace my job Jobs AI Can't Replace: Human Skills That Hold Up Jobs AI can't replace aren't myths. Empathy, taste, physical judgment, trust: here's what humans still own and why it matters for your career. jobs AI can't replaceAI-proof careershuman skills New Jobs AI Is Creating (With Evidence) AI is displacing 85M jobs but creating 97M new ones. Here's what the new jobs actually look like and why the pattern always repeats. new jobs created by AIAI job creationfuture of work --- ## AI Launch Checklist Prompts: Ship Without Gaps URL: https://dontreplace.me/articles/ai-launch-checklist-prompts-ship-without-gaps ai promptslaunch checklistproject managementai at workproduct launch AI Launch Checklist Prompts: Ship Without Gaps 10 copy-paste AI launch checklist prompts for ChatGPT and Claude. Turn messy plans into structured checklists with owners, risks, and rollback steps. July 17, 2026 Updated July 17, 2026 Most launches don't fail because the team was lazy. They fail because someone assumed someone else had checked the thing. The QA link was never shared. The rollback plan existed in one person's head. Support didn't know the feature was shipping until users started complaining. A good AI launch checklist prompt can't fix those problems by itself, but it can make the gaps impossible to ignore before you're dealing with them at 11pm on a Tuesday. This is a template page. You're here for prompts you can use today, not a seminar on launch theory. So here they are. What makes a good AI launch checklist prompt? Before the templates, one thing worth saying: a checklist that AI generates is only as good as what you put in. Vague input gets you vague output. "We're launching a new feature" produces a generic list that looks thorough and covers almost nothing specific to your situation. The prompts below use a consistent structure. You swap in your real details. Every good launch checklist prompt needs: what you're launching, who owns what, what "done" looks like, what the risks are, and what happens if it goes sideways. That's it. Anything AI makes from that is a starting structure, not a sign-off. One more thing before you paste anything anywhere. Don't put customer PII, security credentials, unreleased product details, legal disputes, financial forecasts, HR issues, or private client conversations into AI tools that aren't approved for that data. If you're unsure whether your company's AI tool is approved for the data you're working with, check with your IT or security team first. Nothing in these prompts requires you to include sensitive information. AI can build the scaffold. You still have to inspect the building. How to use these AI launch checklist prompts Copy the prompt. Fill in the brackets. Paste into ChatGPT, Claude, or whatever your team uses. Read the output critically. Every owner and date it suggests is a placeholder until a real person has said yes out loud. Every risk it flags should be checked against what you actually know, not accepted as settled. When AI says something sounds "low risk," that's a hypothesis worth verifying, not a clearance. If a prompt surfaces something involving legal, security, privacy, finance, HR, customer contracts, or exec-level decisions, escalate to the right person. AI is not your compliance team. These prompts pair well with the AI project kickoff prompts if you're still in the planning phase, or with the AI QA checklist prompts if you're focused on testing coverage rather than the full launch picture. 10 AI launch checklist prompts you can use today Prompt 1: Turn a messy launch plan into a structured checklist I'm preparing for a product/feature/campaign launch. Here's what I have so far: [Paste your messy notes, Slack summary, doc outline, or bullet list] Create a structured launch checklist organized by phase (pre-launch, launch day, post-launch). For each item, include: task description, suggested owner role (I'll assign real names), target completion date (I'll fill in real dates), and a "done" definition. Flag any items where you don't have enough information to be specific and list what I should clarify. Use this when you're working from a doc that's half planning notes and half someone's stream of consciousness. The output won't be perfect, but it'll be a structure you can actually edit instead of a blank page. Prompt 2: Find missing owners and dates Here is our current launch checklist: [Paste checklist] Review this checklist for: 1. Items with no named owner or vague owner (e.g., "the team") 2. Items with no target date or only relative dates (e.g., "before launch") 3. Dependencies where one item can't start until another is complete but the link isn't stated List each gap with a recommendation for what to specify. Don't invent names or dates. Just flag what's missing. This is one of the most useful things AI can do for a launch: spot the places where accountability has been quietly avoided. You still have to assign the real people. But at least now you know where the holes are. Prompt 3: Check launch scope against acceptance criteria Here is our launch scope: [Paste scope or feature list] Here are the acceptance criteria we're working from: [Paste acceptance criteria] Compare these two. Identify: 1. Scope items with no corresponding acceptance criteria 2. Acceptance criteria that don't map to anything in the current scope 3. Areas where the criteria are too vague to be testable List the gaps. Don't invent criteria. I need to know what's missing so I can write it properly. If you haven't written acceptance criteria yet, the AI acceptance criteria prompts are a good place to start before you run this one. Prompt 4: Review customer-facing copy and assets We're launching [brief description of product/feature/campaign]. Here is the customer-facing copy and asset list we plan to use: [Paste copy, or list the assets and their current status] Review for: 1. Anything that promises functionality we should verify is actually working at launch 2. Inconsistencies between what the copy says and what I've described as the product/feature 3. Missing assets (e.g., error states, mobile views, accessibility labels) based on what a typical [SaaS / ecommerce / B2B / consumer] launch would need 4. Tone or claims that might need legal or compliance review before going live Flag only. I'll make decisions on each item. Don't paste customer data or confidential contract terms into this. If the copy involves pricing, legal obligations, or regulated claims, get a human to review it before shipping. Prompt 5: Map dependencies and approvals Here is our launch plan: [Paste plan] Map the dependencies and approvals required. Specifically: 1. Which tasks must be complete before others can start? 2. Which approvals are needed and from which roles (I'll confirm real names)? 3. Where are the longest chains of dependencies? 4. What's the latest possible date each item can slip without affecting the launch date? Present this as a dependency table, not a paragraph. Flag any dependencies where you don't have enough information to be confident. Approval chains are where launches go quiet and then explode. This prompt doesn't fix that, but it makes the chain visible so you can pressure-test it in a real conversation with real people. Prompt 6: Pressure-test rollout risks We are launching [describe what you're launching, to whom, and at what scale]. Here is what we know about the risks: [Paste known risks, constraints, technical limitations, or anything your team has flagged] For each risk I've listed, assess: likelihood (low/medium/high), potential customer or business impact, and whether the risk is currently mitigated, partially mitigated, or open. Then identify any common launch risks for this type of product/audience that I haven't mentioned. Flag each as something I should verify with my team, not as confirmed risks. Do not invent statistics or claim any risk is acceptable. Flag anything that may need review by engineering, security, legal, or customer success. The AI risk assessment prompts go deeper on risk structure if you need more than a launch-specific view. Prompt 7: Build a rollback checklist We are launching [describe the launch]. Our tech stack / deployment approach is: [brief description]. Create a rollback checklist. Include: 1. The trigger conditions that would cause us to roll back (I'll validate these with engineering) 2. The steps to execute a rollback in order 3. Owner role for each step (I'll assign real names) 4. Communication tasks required if we roll back (who needs to be told, what they need to know) 5. Post-rollback review steps Flag any steps where you need more information about our environment to be specific. Do not invent technical details about our systems. Rollback plans that live in one engineer's head are not rollback plans. This prompt gets it out of someone's head and onto a document that other people can actually follow under pressure. Prompt 8: Prepare support and customer success notes We are launching [describe the launch]. Here is what's changing for customers: [Describe the change, any known edge cases, and what customers might be confused about] Create a support readiness brief that includes: 1. Common questions customers are likely to ask 2. Suggested responses (draft, not final, I'll review) 3. Known edge cases and how to handle them 4. Escalation paths for issues support can't resolve 5. What support should not say or commit to without checking with [relevant team] This is a starting draft. Flag anything where you'd need more product or policy information to give a good answer. Support finding out about a launch from a user complaint is one of the most preventable things that keeps happening. This prompt gives your support team something to work from before that happens. Prompt 9: Summarize launch readiness for stakeholders Here is our current launch status: [Paste checklist status, open items, risks, and any decisions made] Write a launch readiness summary for [executive / stakeholders / leadership]. Include: 1. What we're launching and when 2. What's complete 3. What's still open and who owns it 4. Known risks and current mitigation status 5. A clear go/no-go recommendation based on the information I've provided (flag if the information isn't sufficient to make a recommendation) 6. What we need a decision on before we can proceed Keep it to one page. Don't minimize open items to make the summary look better than it is. This pairs naturally with the AI stakeholder update prompts if you need variations for different audiences. The AI status report prompts are also useful if you're tracking across multiple workstreams. Prompt 10: Final pre-launch sanity check We are launching [describe the launch] on [date]. Here is the final state of our launch checklist: [Paste checklist] Run a final sanity check against this list. Specifically: 1. Any items marked "done" that lack a clear definition of done or a named verifier 2. Any open items that could block the launch or cause customer impact 3. Any approvals that appear to be missing or unclear 4. Any communication tasks (internal or external) that aren't assigned 5. Any rollback steps that aren't specified Present findings as a prioritized list: must resolve before launch / should resolve before launch / monitor post-launch. Be honest about uncertainty. Don't tell me we're ready if there are gaps. Don't run this prompt and take the output as a green light. Run it, review the gaps it surfaces, make real decisions with real people, and then decide whether you're launching. The judgment call is yours. Dmitry Kargaev covers exactly this dynamic in Don't Replace Me: the risk isn't that AI gives you bad advice, it's that a polished-looking output feels like a decision when it's actually just a draft. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What AI can't do on a launch checklist AI can spot structural gaps. It can ask useful questions. It can organize chaos into something readable. What it can't do is know whether a test actually passed, whether a customer-facing claim has been reviewed by your legal team, or whether the "probably fine" risk your senior engineer mentioned in Slack last Thursday is actually fine. Every item on a launch checklist needs a human who has checked the thing and can say they checked it. AI-generated checklists create the illusion of coverage without the substance. The substance comes from evidence: screenshots, test logs, written approvals, named people with dates. If you're working from AI-generated output and something feels uncertain, it probably is uncertain. Flag it. Escalate it. Don't ship through the fog because the checklist looks clean. For the bigger picture on what AI is genuinely useful for at work, and where it'll get you in trouble, the no-BS guide to using AI at work is a good place to ground your thinking. And if you want to understand why AI's confident tone doesn't mean confident output, what AI can and can't do explains it in plain language. Frequently asked questions Can I use AI to create a launch checklist from scratch? Yes, and it's one of the better uses of AI in project work. The key is giving it real context: what you're launching, known risks, owner roles, and what "done" looks like for each item. Generic input produces generic output. Treat the result as a draft that needs human review, not a finished document. What information should I not paste into AI tools when building a launch checklist? Don't paste customer PII, security credentials, unreleased financial forecasts, legal disputes, HR details, private client conversations, or confidential product strategy into AI tools that haven't been approved for that data. Most launch planning can be described in general terms without including the sensitive specifics. How do I know if my AI-generated launch checklist is actually ready? You don't, based on the checklist alone. Readiness comes from evidence: test results, written approvals, named owners who've confirmed their items, and a human decision-maker who has reviewed the open risks and said "we're go." A complete-looking checklist isn't the same as a complete launch. Should AI be making the go/no-go call on a launch? No. AI can summarize the state of your checklist and flag open items, but the go/no-go decision belongs to the humans accountable for the launch, the customers, and the business. That includes engineering, QA, product, legal, security, and whoever is on the hook if something goes wrong. What's the biggest mistake people make with AI launch checklists? Treating the output as an audit rather than a draft. AI will produce something that looks thorough, uses the right vocabulary, and hits every expected category. That doesn't mean it reflects your actual situation. You have to read it critically, fill in the specifics, and make real humans confirm the real items. When should I escalate a launch risk beyond the checklist? When the risk involves legal obligations, customer contracts, security vulnerabilities, regulated data, financial impact, or any scenario where a mistake affects people outside your immediate team. AI can flag that something might warrant escalation. The escalation itself is always a human conversation. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Migration Plan Prompts: 10 Templates That Work 10 copy-paste AI migration plan prompts for system mapping, dependency checks, rollback criteria, cutover plans, and stakeholder comms. With safety reminders. AI promptsmigration planningsystem migration AI Requirements Gathering Prompts: Stop Starting Vague 10 copy-paste AI requirements gathering prompts to turn vague asks into verified briefs. Includes safety rules and a reusable prompt formula. AI promptsrequirements gatheringproject management AI Assumption Log Prompts: Stop Planning on Guesswork Use 10 AI assumption log prompts to expose project guesswork, design evidence checks, assign review owners, and stop unverified beliefs becoming commitments. AI promptsproject managementassumption logs --- ## AI Localization Testing Prompts: Catch Locale Bugs URL: https://dontreplace.me/articles/ai-localization-testing-prompts-catch-language-and-locale-bugs AI promptslocalization testinginternationalizationquality assurancesoftware testing AI Localization Testing Prompts: Catch Locale Bugs 10 AI localization testing prompts for translations, text expansion, RTL layouts, formats, Unicode, accessibility, defects, and release decisions. August 19, 2026 Updated August 19, 2026 Localization testing is where “we translated the strings” meets the product and discovers the product has opinions about string length, calendars, names, currencies, and which direction text should travel. AI localization testing prompts can help organize locale requirements, draft test ideas, compare approved terminology, expose missing coverage, and summarize sanitized defects. They cannot see every rendered screen, understand every cultural context, certify a translation, or decide that a release is safe. AI can draft the checklist. Humans still read the language, use the product, inspect the evidence, fix the defects, and own the release. If your wider environment matrix is unclear, start with AI compatibility testing prompts. Use AI test case prompts when you need detailed checks. This guide focuses on language, locale, regional formats, and the interface failures hiding between them. What localization testing actually covers Localization testing is not proofreading with extra steps. It checks whether a product behaves correctly and communicates appropriately for a supported locale. Useful coverage includes: Translation completeness: missing, stale, duplicated, fallback, or exposed resource keys. Linguistic quality: meaning, grammar, tone, terminology, pluralization, gender, and context. Layout: expansion, wrapping, truncation, overlap, alignment, and controls that stop being usable. Writing direction: right-to-left layout, mixed-direction content, icons, navigation, and input behavior. Characters and fonts: Unicode storage, combining marks, emoji, fallback fonts, line breaking, and missing glyphs. Regional formats: dates, calendars, time zones, numbers, decimals, currencies, units, addresses, and phone numbers. Behavior: sorting, search, case conversion, input, validation, exports, notifications, and integrations. Culture and compliance: imagery, examples, color, idioms, legal text, consent, and regional expectations. Accessibility: localized labels, reading order, focus, zoom, text spacing, and assistive-technology output. Internationalization and localization are related but different. Internationalization makes the product capable of supporting locales: externalized strings, Unicode, flexible layouts, locale-aware formatting, and bidirectional rendering. Localization supplies and adapts language and regional content. Testing needs both. A flawless translation cannot rescue a button that only fits eight characters. Pseudolocalization is useful too. Accented characters, expanded strings, and simulated right-to-left text can reveal hard-coded copy and fragile layouts before translations arrive. It does not replace review by qualified people who understand the language, product, and culture. The reusable localization testing prompt formula Use this base prompt with one specialized template below: “Act as a localization test-planning assistant. I am evaluating [named product, build, locales, and critical journeys]. Use only the supplied locale requirements, approved glossary, style guide, resource inventory, designs, technical notes, and sanitized evidence. Produce [artifact] with coverage, priorities, setup, expected behavior, evidence fields, exclusions, assumptions, and open questions. Separate verified facts, proposed checks, observations, hypotheses, and human decisions. Do not invent translations, cultural rules, test execution, defect causes, or release approval.” That last sentence matters. Models confidently manufacture linguistic rules and test results when context is thin. A proposed check is not proof. A machine-generated translation is not approved copy. A screenshot in one viewport does not establish that a journey works. Never paste credentials, API keys, customer PII, production logs, private source code, proprietary translation memory, unreleased screenshots, regulated data, or confidential client material into an unapproved AI tool. Use synthetic users, redacted evidence, approved environments, placeholder secrets, and your organization’s privacy, security, vendor, and retention rules. What to collect before prompting Give the model a small, verified input packet rather than “check French.” InputWhy it mattersHuman check Supported locale listDefines promised coverageProduct and regional owners approve Approved glossaryGrounds terminologyLinguists resolve conflicts Style and voice guideDefines tone and conventionsContent owner confirms version String/resource inventoryReveals missing and reused copyEngineering verifies extraction Critical journeysFocuses testing on user outcomesProduct confirms priority Designs and content limitsGrounds layout expectationsDesign validates flexibility Format requirementsDefines dates, money, units, addressesRegional experts confirm Platform matrixExposes device and browser variationQA confirms real environments Accessibility requirementsPrevents localized barriersAccessibility owner reviews Evidence and severity rulesMakes defects actionableQA and engineering agree Record versions. Glossaries, translations, builds, and designs change independently. Without fingerprints, testers can spend a morning diagnosing a “translation defect” caused by an old bundle or cached content. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI localization testing prompts Replace bracketed text with verified, sanitized information. These prompts create plans and analysis artifacts. Humans must perform the checks in real products and approve language. 1. Turn locale promises into a test scope “Using these supported locales, markets, platforms, contractual commitments, critical journeys, release changes, translation status, regional requirements, and test resources: [paste], draft a risk-based localization test scope. Separate full, smoke, regression, linguistic, visual, accessibility, and out-of-scope coverage. Identify assumptions, dependencies, owners, evidence, and unresolved decisions.” A scope should name exact locale variants. Portuguese for Brazil is not interchangeable with Portuguese for Portugal. French for Canada can have different terminology and legal needs from French for France. Language alone is too vague. Prioritize by user impact, changed areas, transaction risk, regulatory obligations, and historical defects. Do not let AI invent market size or legal requirements. Product, legal, localization, support, and regional teams own those facts. 2. Audit translation completeness and resource use “Using this approved source-string inventory, locale resource files, build version, fallback rules, feature flags, screenshots, and known exceptions: [paste], create a translation-completeness audit. Flag missing keys, source-language leakage, stale translations, duplicate or conflicting keys, suspicious placeholders, concatenated fragments, fallback behavior, and strings requiring runtime verification. Do not translate missing content.” Completeness is more than comparing file counts. Strings may be assembled at runtime, supplied by a CMS, hidden behind permissions, or reused in contexts that need different grammar. Check errors, empty states, notifications, emails, exports, help text, metadata, and accessibility labels—not only the happy-path screens. Preserve variables and markup. A “helpful” model that rewrites {userName}, %1$s, or an ICU plural expression can create a production defect while appearing linguistically polished. 3. Test text expansion, wrapping, and truncation “Using these approved translations, component designs, viewport ranges, text-scaling requirements, character limits, responsive rules, and critical states: [paste], create visual localization scenarios. Include short and long strings, buttons, navigation, tables, dialogs, errors, empty states, dynamic values, zoom, larger text, orientation, virtual keyboards, and evidence checkpoints.” German is not magically “30% longer” in every sentence, and no universal expansion percentage replaces real content. Use approved translations plus deliberately extreme samples. Check whether controls grow, wrap, scroll, or expose the full meaning accessibly. Truncation can change meaning. A clipped destructive action, hidden unit, or missing negation is not cosmetic. Test localized UI behavior in realistic viewports and with increased text size. 4. Validate dates, times, numbers, and currencies “Using these supported locales, calendars, time zones, number rules, currencies, rounding policy, units, business deadlines, and critical workflows: [paste], create format and calculation checks. Cover display, input, parsing, storage, sorting, conversion, daylight-saving boundaries, midnight rollover, leap dates, negative values, zero, large values, and round-trip consistency.” A format bug can become a money bug. 1,234 and 1.234 do not mean the same thing everywhere. A date such as 03/04/2026 is ambiguous. A timezone conversion can move a booking or deadline into another day. Require unambiguous stored values and verify conversion at boundaries. Check currency symbol placement, decimal digits, negative formats, nonbreaking spaces, and whether changing locale accidentally changes the underlying amount. Humans must approve financial and legal behavior. 5. Check right-to-left and mixed-direction interfaces “Using these supported right-to-left locales, component library rules, navigation patterns, icons, charts, forms, tables, media controls, and mixed-direction samples: [paste], draft RTL test scenarios. Include mirroring rules, logical alignment, focus order, scrolling, punctuation, numbers, URLs, email addresses, phone numbers, code, cursor movement, selection, and assistive-technology reading order.” Right-to-left support is not transform: scaleX(-1) applied to civilization. Some directional icons mirror; others represent physical or universal concepts and should not. Numbers and Latin identifiers can remain left-to-right inside surrounding RTL text. Test with actual Arabic or Hebrew content reviewed by qualified people. Confirm reading order, keyboard movement, selection, copy and paste, validation, and mixed text. A mirrored screenshot can look tidy while interaction order is nonsense. 6. Stress Unicode, fonts, names, and input “Using these supported scripts, approved fonts, database and API constraints, input rules, search behavior, import/export formats, and representative synthetic names and addresses: [paste], create Unicode and input tests. Cover combining marks, composed and decomposed forms, emoji, surrogate pairs, non-Latin digits, bidirectional text, line breaking, fallback fonts, copy/paste, normalization, validation, storage, retrieval, and round trips.” Do not use real customer names as test data. Create synthetic examples that exercise spaces, apostrophes, hyphens, multiple family-name patterns, mononyms, long text, and multiple scripts. Avoid “validating” human names according to one culture’s assumptions. A string can render correctly and still fail after save, search, export, email, or API transfer. Verify complete round trips. Use AI test data prompts to structure synthetic coverage without copying production records. 7. Test search, sorting, and locale-aware behavior “Using these locale requirements, supported scripts, collation rules, search configuration, case and accent behavior, filters, indexing pipeline, and representative synthetic records: [paste], create locale-aware search and sorting checks. Include exact and partial matches, accents, case, normalization, transliteration policy, tokenization, punctuation, numeric strings, mixed scripts, no-result states, and ranking evidence.” Alphabetical order is locale-dependent. Case conversion has edge cases. Removing accents may improve discovery in one product while destroying meaningful distinctions in another. Search expectations must come from product requirements and regional expertise, not generic model intuition. Capture the query, locale, indexed data version, expected rule, actual ranking, and evidence. A model can compare structured results; it cannot decide what relevance should mean for your users. 8. Review imagery, tone, and cultural adaptation “Using this approved campaign brief, screenshots, icon inventory, examples, imagery descriptions, glossary, voice guide, market constraints, and reviewer notes: [paste], create a cultural-review checklist. Identify items requiring qualified regional review, explain the possible concern without asserting harm, and record owner, decision, replacement, and evidence. Do not make universal claims about a culture.” This is where overconfident AI becomes especially dangerous. Culture is not a lookup table. Meanings vary by region, community, context, age, and product. Treat model suggestions as questions for regional reviewers, not verdicts. Review names, humor, idioms, gestures, maps, flags, holidays, food, clothing, colors, examples, support language, and assumptions about addresses or households. Keep decisions documented so the next release does not rediscover the same issue through angry support tickets. 9. Triage a sanitized localization defect “Using this sanitized evidence—build, locale, platform, exact steps, source and approved target strings, resource key, screenshots, expected and observed behavior, reproduction rate, fallback behavior, and comparison locales: [paste]—draft a localization defect brief. Separate facts from hypotheses, classify linguistic, functional, visual, data, accessibility, or configuration risk, suggest safe isolation checks, and identify required reviewers. Do not invent a translation or root cause.” Compare one variable at a time: locale, platform, viewport, account, resource version, or previous build. A clipped string might be a translation-length issue, a component constraint, stale assets, a font failure, or unexpected dynamic data. Keep raw screenshots and logs in approved systems. Use AI bug report prompts to format verified evidence, but remove identities, tokens, private URLs, and customer content first. 10. Turn verified results into a release brief “Using only these verified localization results, locale scope, passed and failed journeys, linguistic approvals, defect severities, evidence links, known gaps, workarounds, remediation owners, and monitoring plans: [paste], draft a localization release brief. Include coverage by locale and platform, unresolved risk, retest status, fallback behavior, support communication, rollback triggers, and a decision table. Do not invent results or make the final decision.” Keep observations separate from decisions. “Checkout total displays the wrong decimal separator in de-DE” is evidence. Whether it blocks release depends on transaction integrity, affected users, workaround, commitments, remediation confidence, and rollback. Use AI go/no-go prompts to structure the governance meeting. Accountable humans still decide. A workflow that produces evidence instead of translation theater Define supported locales. Name language-region variants, platforms, and support tiers. Freeze inputs. Version source copy, glossary, translations, designs, build, and locale data. Sanitize context. Remove secrets, customer data, private screenshots, and proprietary assets. Pseudolocalize early. Find hard-coded strings, expansion failures, and basic RTL problems before handoff. Prioritize journeys. Focus on authentication, payments, creation, search, settings, support, and regulated flows. Run technical checks. Verify resources, layouts, formats, storage, search, inputs, exports, and integrations. Run human review. Qualified linguistic and regional reviewers judge meaning, tone, terminology, and culture. Check accessibility. Verify labels, reading order, focus, zoom, text spacing, and assistive technology. Use AI accessibility testing prompts for deeper planning. Capture evidence. Record exact locale, build, platform, viewport, resource version, steps, and results. Fix, retest, and decide. Humans verify remediation and own the release decision. Common ways localization testing goes wrong Treating translation as a final production step Localization defects get expensive when strings arrive after layout, validation, data models, and workflows are frozen. Internationalize early, pseudolocalize in development, and involve localization specialists before the release panic. Testing only the homepage The dangerous copy often lives in validation, checkout, account recovery, notifications, exports, legal consent, and error recovery. Test complete journeys and persisted outcomes. Using machine translation as linguistic approval Generated copy can help create layout stress data or a rough draft in an approved workflow. It does not certify meaning, terminology, tone, legal suitability, or cultural fit. Qualified humans approve user-facing language. Assuming one locale represents a language Regional variants can differ in vocabulary, formats, currency, regulation, and expectations. Define what you support instead of silently collapsing variants. Ignoring accessibility after translation Longer labels can break focus indicators and zoom. Translated accessible names can diverge from visible labels. RTL changes reading order. Localization and accessibility need coordinated evidence, not separate checkboxes. Pasting private translation assets into public AI tools Translation memories, screenshots, logs, source strings, and defect evidence may expose unreleased products, customer data, strategy, or licensed content. Redact aggressively and use approved systems. Understanding what AI can and cannot do keeps the boundary clear: generation and organization are fast; language judgment, cultural context, real execution, and accountability stay human. Frequently asked questions Can AI perform localization testing automatically? AI can help orchestrate approved automation when engineers deliberately connect it, but a chat response has tested nothing. Connected tools still need trusted translations, representative environments, reviewed assertions, stored evidence, and human linguistic interpretation. What is the best AI prompt for localization testing? Name the product build, exact locales, critical journeys, approved glossary, style guide, resource version, platforms, expected behavior, evidence requirements, and exclusions. Require assumptions and gaps to be explicit. Forbid invented translations, execution, and approval. Is pseudolocalization enough? No. It is excellent for finding hard-coded strings, expansion problems, encoding failures, and basic RTL issues early. It cannot judge translation accuracy, tone, terminology, cultural appropriateness, or every locale-specific behavior. Who should review localized software? The team usually needs QA, engineering, localization specialists, qualified native or near-native reviewers, product, design, accessibility expertise, and regional legal or compliance input where relevant. One reviewer rarely covers every dimension. How many locales should we test before release? There is no universal number. Test every promised locale to the depth required by risk, change, user impact, transaction severity, and obligations. Use tiers deliberately, document gaps, and never call an untested locale “supported” because its files compiled. Is localization testing the same as translation review? No. Translation review evaluates meaning, terminology, grammar, tone, and context. Localization testing also checks rendering, formats, inputs, storage, search, sorting, integrations, accessibility, and complete workflows. Both are necessary. Can AI decide whether a localization bug blocks release? No. It can organize verified evidence against severity rules. Humans must judge user impact, transaction risk, legal obligations, affected locales, workarounds, remediation, monitoring, support burden, and accountability. The useful boundary AI is good at turning scattered locale requirements into structured plans, generating test variations, finding obvious omissions, and formatting sanitized evidence. That can save real time. It is bad at knowing whether a phrase feels natural, whether a cultural concern matters in context, whether a screen actually works, or whether remaining risk is acceptable. Those are the valuable parts: observation, language, taste, context, and responsibility. Use the prompts to accelerate planning and analysis. Keep approved language, sensitive assets, real execution, remediation, and release authority with accountable people. That is the broader argument in Dmitry Kargaev’s Don’t Replace Me: use the machine for speed, then apply the human judgment it cannot supply. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance AI API Testing Prompts: Catch Broken Integrations AI API testing prompts for contracts, edge cases, authorization, retries, error responses, and human-verified integration evidence. AI promptsAPI testingquality assurance AI Compatibility Testing Prompts: Catch Breakage 10 AI compatibility testing prompts for browser, device, OS, responsive, locale, network, defect triage, and evidence-based release decisions. AI promptscompatibility testingcross-browser testing --- ## AI Manager Prompts: Better 1:1s, Feedback, Team Updates URL: https://dontreplace.me/articles/ai-manager-prompts-better-1-1s-feedback-team-updates ai manager promptsmanagementfeedback1:1steam communication AI Manager Prompts: Better 1:1s, Feedback, Team Updates 11 copy-paste AI manager prompts for 1:1 agendas, feedback, team updates, and hard conversations. Includes safety rules and a reusable prompt formula. May 31, 2026 Updated May 31, 2026 Management is mostly communication work. You're constantly preparing for conversations, summarizing what happened in conversations, and documenting what you said in conversations so nobody can pretend it didn't happen. AI manager prompts can cut the prep time for all of that without replacing the thing that actually matters: your judgment about what to say and whether to say it. This isn't about automating your leadership. It's about spending less time staring at a blank doc before a 1:1 and more time actually listening during one. Here are 11 prompts you can copy, adjust, and use today. Plus the formula that makes all of them work. The manager prompt formula that actually works Every prompt in this list follows the same structure. Role, context, task, constraints, format. That's it. You're not asking AI to be your manager. You're giving it enough information to draft something useful that you then read, edit, and take responsibility for. Think of it like a smart intern who types fast and has read every management book but has never met your team, doesn't know your company's history, and has no idea what happened in the all-hands last Tuesday. The more context you give, the better the output. The less you give, the more generic the result. Generic outputs waste time. Specific inputs save it. One hard rule before the prompts: don't paste real employee names, salary information, medical details, HR investigation notes, or anything your company considers confidential into a public AI tool unless your organization explicitly allows it. Anonymize. Use "the employee" or "my direct report" instead. More on this at the end. AI manager prompts for better 1:1s Good 1:1s don't happen by accident. They happen because someone thought about what matters before the meeting starts. AI is genuinely useful for that thinking-out-loud phase. Prompt 1: Prepare a 1:1 agenda You are a management coach. I'm preparing for a weekly 1:1 with a direct report. Here's what's been going on: [brief context, for example, they're leading a project that's two weeks behind, they mentioned feeling stretched last time we met, we've got a big deadline in three weeks]. Draft a 1:1 agenda with four to six items that balance project updates, their concerns, and any feedback I need to give. Format as a short bulleted list I can paste into our shared doc. Adjust the context block every week. This prompt works because you're making it specific. "Give me a 1:1 agenda" gives you a generic list of questions you could've Googled. The context makes it useful. Prompt 2: Turn your messy notes into a follow-up summary I just finished a 1:1. Here are my rough notes: [paste your notes]. Write a clean follow-up summary with three sections: what we discussed, what we agreed on, and open questions or next steps. Tone should be direct and professional. Under 200 words. You take better notes when you know you're going to hand them to something that cleans them up. It also means your direct report gets an actual summary instead of an email that says "good chat!" AI manager prompts for feedback that doesn't land like a wet towel Feedback is one of the hardest management skills to develop. Most managers either avoid it or deliver it in a way that makes the other person defensive before you get to the useful part. AI won't fix your courage, but it can fix your language. Prompt 3: Make vague feedback specific I want to give feedback to a direct report about [describe the behavior or situation, for example, they're often late to team meetings and when they do show up, they seem disengaged]. Help me rewrite this feedback so it's specific, behavior-focused, and not accusatory. Include one example sentence that names the behavior, one that explains the impact, and one that opens a conversation rather than closing it. The structure here (behavior, impact, open question) is a classic feedback framework. You're using AI to apply it, not to invent the conversation. Prompt 4: Turn messy notes into structured feedback Here are my rough observations about a direct report's recent performance: [paste notes]. Write this up as clear developmental feedback using the structure: situation, behavior, impact, and a forward-looking suggestion. Keep the tone direct and supportive, not harsh. Under 300 words. One warning: never send feedback you wouldn't personally stand behind. Read it. Edit it. Make sure it sounds like you, not like a management textbook. If it doesn't feel true, it isn't ready. The full performance review version of this is covered in the AI performance review prompts guide, if you're heading into review season. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → AI manager prompts for team communication Team updates are one of those management tasks that eats time disproportionate to their impact. You spend 45 minutes writing an update that people skim in 90 seconds. AI can help you write faster and cleaner, which at least gets you the 90-second skim instead of the delete. Prompt 5: Draft a team status update You are a communications assistant. I need to write a weekly team update email. Here's the context: [project name, what's on track, what's delayed and why, what decisions are needed from leadership, any shoutouts or wins]. Write a status update email under 250 words. Use three sections: progress this week, blockers or risks, and what's coming next. Tone should be clear and factual, not cheerleader-ish. "Not cheerleader-ish" is a legitimate instruction. AI defaults to relentlessly positive unless you tell it not to. Prompt 6: Write a delegation brief I need to delegate a project to a direct report. Here's what I know: [project goal, key deliverables, deadline, any constraints like budget or team members available, and what I need updates on and when]. Write a delegation brief that explains the goal, the scope, what good looks like at the end, and what decisions they can make independently versus what needs my sign-off. Under 400 words. A clear delegation brief is the thing that prevents you from having to answer the same clarifying questions three times. This prompt makes it easy enough that you'll actually write one. Prompt 7: Summarize team risks I need to present a summary of current team risks to my manager. Here are the risks I'm tracking: [list them in rough notes form]. Organize these into a short risk summary with three columns: risk, current status, and mitigation plan. Make the language direct. No jargon. This one pairs well with the AI meeting notes prompts if you're pulling risks out of standup summaries. AI manager prompts for hard conversations Hard conversations are the job. They're also the thing most managers avoid longest. AI can help you prepare, but it cannot have the conversation for you and it definitely cannot read the room when your direct report starts crying or gets defensive. Prompt 8: Prepare for a hard conversation I need to have a difficult conversation with a direct report about [describe the situation in general terms, for example, consistent lateness, missing a key commitment, a behavior that's affecting the team]. Help me prepare by: outlining what I want to say in the first two minutes, anticipating two or three ways they might respond, and suggesting how I can acknowledge their perspective without undermining the main message. Tone should be calm and direct. Important: Do not generate anything that sounds like a final written warning or a legal document. This is for conversation prep only. That last instruction matters. You don't want AI generating language that sounds like formal HR documentation when you're planning a regular management conversation. Dee covers the judgment layer in Don't Replace Me, specifically the idea that managers hold context AI can't touch: the employee's history, the team dynamics, the organizational politics, and the difference between someone having a bad month and someone who isn't working out. Prompts help you prepare. They don't replace knowing which situation you're in. Prompt 9: Turn a project postmortem into actionable lessons We just finished a project. Here are my notes from the postmortem: [paste notes]. Summarize the key lessons learned in three categories: what worked and should be repeated, what went wrong and how to prevent it next time, and what's still unclear and needs follow-up. Format as a short internal doc I can share with the team. Tone should be honest and forward-looking, not blame-y. "Not blame-y" is another legitimate instruction. Postmortems that read like blame docs don't help anyone and usually make things worse. Two more prompts that save real time Prompt 10: Coaching without micromanaging I want to coach a direct report through a challenge they're facing: [describe the situation]. They're capable but stuck. Help me write three coaching questions I can ask in our next 1:1 that would help them think through the problem themselves, rather than me just giving them the answer. Questions should be open-ended and genuinely curious, not leading. This one is underrated. If you've noticed you default to giving answers instead of asking questions, this prompt gives you better questions without having to come up with them mid-conversation. Prompt 11: Translate leadership strategy into team priorities My organization has communicated the following strategic priorities for this quarter: [paste the memo or key points]. Help me translate this into three to five concrete priorities my team can actually act on this week. For each priority, include a one-sentence "what this means for us" explanation and one example of a task or behavior that demonstrates it. Keep it brief and practical. Leadership strategy often arrives in language that sounds important and means nothing to the people who need to do the work. This prompt closes the gap. If you're looking for prompts across other work contexts, the general AI prompts for work guide covers a broader range, and ChatGPT at work has practical tips for getting better outputs from the tools themselves. What you must never skip: privacy, accuracy, and your own judgment These prompts are useful. They're not magic. And used carelessly, they can cause real problems. Privacy first. Don't paste real employee names, salary figures, medical information, HR investigation notes, disability accommodations, immigration status, or anything from an active legal matter into a public AI tool. Anonymize. Your company almost certainly has a policy on this. If you don't know what it is, find out before you start using these prompts. The disclosure question applies to you as a manager too. Verify everything. AI will confidently generate feedback language, policy-sounding sentences, and dates that are wrong. If a prompt output references a company policy, a promise you made, or a specific performance claim, verify it before you use it. You're responsible for what you send. Read it before you send it. AI-written feedback or performance documentation that you send without reading carefully is still your feedback and your documentation. If it doesn't sound like you, edit it. If it doesn't feel true, delete it. The words are yours the moment you hit send. The thing AI genuinely can't replace in management isn't the drafting. It's knowing when someone needs a direct conversation instead of another carefully worded email. That's the job. The prompts just handle the paperwork. Frequently asked questions Can I use AI to write performance reviews for my team? Yes, as a drafting and organizing tool, with important caveats. AI can help you structure your observations and make vague feedback more specific. But you must verify every claim, remove any confidential or HR-sensitive information before pasting, and read the final version carefully before submitting it. The AI performance review prompts guide has specific templates built for that use case. Is it cheating to use AI to prepare for 1:1s or write team updates? No. Using a tool to prepare better for a conversation is the same as writing an agenda beforehand or reviewing your notes. The 1:1 itself, the listening, the judgment calls, the follow-through, that's still your work. AI just saves you time on the paperwork that surrounds it. What information should I never paste into an AI tool as a manager? Never paste real employee names combined with performance details, salary or compensation data, medical or disability information, HR investigation notes, anything from an active legal matter, or confidential customer data. Use "my direct report" or "the employee" instead of real names. Check your company's AI use policy before you start. Will AI replace managers? Not in any near-term sense that should keep you up at night. Management is mostly context work and trust work: knowing your team, making judgment calls, navigating organizational politics, and being the person who makes a decision when nobody wants to. The jobs AI can't replace breakdown covers why accountability, coaching, and interpersonal judgment are genuinely hard to automate. How do I get better outputs from these prompts? The more specific your context block, the better the output. Don't just say "I have a struggling direct report." Say what they're struggling with, what you've already tried, and what outcome you're aiming for. The role-context-task-constraints-format structure works every time. Treat the AI like a fast assistant who knows nothing about your specific situation until you tell it. What if my company doesn't allow AI tools? Then don't use them for work without getting clarity first. Some organizations have approved specific tools, prohibited public ones like ChatGPT, or haven't made a policy yet. Ask your IT or legal team. If you're navigating this question, the should I tell my boss I use AI article covers the workplace politics of AI disclosure. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance --- ## 10 AI Marketing Prompts That Actually Work URL: https://dontreplace.me/articles/ai-marketing-prompts-campaigns-copy-content ai marketing promptsmarketing copychatgpt for marketingcontent marketingai at work 10 AI Marketing Prompts That Actually Work 10 copy-paste AI marketing prompts for copy, positioning, editing, and campaign retrospectives. Includes the context block method most marketers skip. June 02, 2026 Updated June 02, 2026 AI marketing prompts are everywhere right now. Blog posts, YouTube videos, LinkedIn threads full of people promising that the right prompt will make your campaigns write themselves. Here's what actually happens: you paste a generic prompt into ChatGPT, get back a list of "key value propositions" that could describe any company in your industry, and spend 20 minutes editing out phrases like "synergistic solutions" and "best-in-class service." The output is technically copy. It is not your copy. The problem isn't AI. The problem is that most people treat it like a vending machine: put in a vague request, expect a finished product. Good AI marketing prompts work the same way good briefs do. You have to put in the real stuff: the customer, the context, the constraint, the voice. This article gives you 10 prompts you can actually use. Not magic. Just a faster starting point, so you can spend your time on the part that actually matters: deciding if it's any good. The formula behind every AI marketing prompt that works Every prompt in this list follows the same basic structure. Role, context, task, constraints, format. That's it. Role: what kind of expert should the AI be acting as? A direct-response copywriter? A brand strategist? A skeptical editor? Context: what do you actually sell, who buys it, and what do they care about? Task: what specific thing do you need produced? Constraints: word limits, tone, what to avoid, what to include. Format: bullet list, paragraph, table, email, whatever you need to paste somewhere. You can read more about the role-context-task framework in the broader AI prompts guide for work, but the short version is: the more specific you are going in, the less editing you do coming out. Treat it like briefing a smart intern who's never heard of your company, your customers, or your category. The biggest mistake marketers make with this structure is skipping the context entirely. "Write me a landing page for a SaaS product" is not a brief. "Write me a landing page for a project management tool targeting operations managers at 50-200 person companies who are drowning in spreadsheets and need to justify the purchase to a CFO" is a brief. One of those gets you something you can work with. AI marketing prompts for understanding customers and positioning These are the prompts you should run before you write a single word of campaign copy. Most AI marketing fails because it skips this step entirely and jumps straight to drafting. That's backwards. Good positioning doesn't come from the AI. It comes from your customers. What it can do is help you process and organize what you already know faster, spot patterns in your research, and turn raw customer language into campaign angles you can test. Prompt 1: Extract pain points from customer notes Act as a customer research analyst. I'm going to paste anonymized notes from [X customer calls / support tickets / reviews]. Extract the top 5 recurring pain points, written in the customer's own language, not marketing language. For each one, note how often it appears and one direct quote that represents it. Do not invent problems that aren't in the notes. Format as a table. [Paste anonymized notes here] Prompt 2: Turn positioning into campaign angles Act as a senior brand strategist. Here is our positioning statement: [paste it]. Our primary audience is [describe them]. Generate 6 distinct campaign angles, each with a different emotional or rational hook. For each angle, write one potential headline and one sentence explaining the core message. Avoid superlatives, vague claims, and anything that sounds like every other company in our category. Prompt 3: Create a content brief Act as a content strategist. I need a brief for a [blog post / video / email series / whitepaper] targeting [audience]. The topic is [topic]. The business goal is [goal: awareness / leads / retention]. The reader's main question is: [question]. Include: recommended structure, key points to hit, what to avoid, tone guidance, and a suggested title. Format as a one-page brief. One important note before you paste anything: don't include real customer names, raw support tickets with identifying details, unreleased product specs, or anything your legal or compliance team would wince at. Anonymize before you prompt. Prompts for drafting copy Once you have the positioning and brief, these prompts do the drafting work. They won't get it right first time. That's normal. Your job is to edit, not to publish the first output. Think about it like this: a copywriter who's never worked on your account and had no briefing will produce something generic. An AI that's never worked on your account and had no briefing will produce something even more generic, because at least the copywriter can ask a follow-up question. The prompts below are designed to give AI the equivalent of a proper briefing before it starts typing. Prompt 4: Draft landing page copy Act as a direct-response copywriter. Write landing page copy for [product/service]. The offer is: [describe it]. The primary audience is: [describe them]. The main pain we solve is: [pain]. Our differentiator is: [differentiator]. Include: a headline, a subheadline, three benefit bullets (outcome-focused, not feature-focused), a short social proof placeholder, and a CTA. Under 250 words total. Write in a [confident / friendly / no-nonsense] tone. Do not make any claims we can't prove. Prompt 5: Write email campaign variants Act as an email marketing specialist. Write 3 variants of a [welcome / nurture / promotional / reactivation] email for [audience]. Each variant should lead with a different angle: [angle 1], [angle 2], [angle 3]. Keep each email under 200 words. Subject line for each should be under 50 characters. Use a [tone] voice. Avoid spam trigger words. End each email with one clear CTA. More email-specific prompt templates are in the AI email prompts guide if you want to go deeper on that format specifically. Prompt 6: Create social post variations Act as a social media copywriter. Repurpose this key message for 4 different social formats: a short LinkedIn post (150 words), a Twitter/X thread opener (2 tweets), an Instagram caption with hashtag suggestions, and a short-form video script hook (first 10 seconds only). Key message: [paste it]. Audience: [describe]. Brand voice: [describe or paste a sample]. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Prompts for editing, repurposing, and quality control Writing is actually the easy part. The prompts in this section are the ones that separate people who get value from AI from people who publish embarrassing copy with incorrect statistics and hollow claims. The editing and quality control step is where most marketers give up the real advantage. They use AI to draft, then publish without review. That's how you end up with a case study that references a client you lost, a statistic that doesn't exist, or a headline that makes a claim your product can't back up. Not a hypothetical. This happens constantly. Prompt 7: Repurpose a long article into multiple formats Act as a content repurposing specialist. I have an article on [topic]. Repurpose it into: a 5-bullet LinkedIn post, a 3-email newsletter sequence outline, one short FAQ (5 questions with brief answers), and two short-form video script hooks. Do not add claims or facts that aren't in the original article. Here is the article: [paste it] Prompt 8: Review copy for generic claims This is possibly the most useful prompt in this list, and the one people skip most. Act as a brutally honest copy editor. Review the following marketing copy. Flag every claim that is: (1) unverifiable or unsupported, (2) so generic it could describe any competitor, (3) jargon that a customer wouldn't actually use, (4) a superlative with no evidence. For each flag, suggest a more specific alternative or note that we need a real proof point. Here is the copy: [paste it] Run your AI-generated drafts through this before you touch publish. Understanding what AI is actually good at versus what it's bad at will help you know why this step matters: AI generates plausible-sounding text, not verified claims. Prompt 9: Plan a lightweight experiment Act as a growth marketer. I want to test [hypothesis: e.g., "shorter CTAs convert better than longer ones" / "pain-point subject lines outperform benefit-led ones"]. Help me design a simple A/B test. Include: what exactly to test, how to set up a clean control vs. variant, what metric to track as primary success signal, what sample size or time window is needed before drawing a conclusion, and what result would change my behavior. Keep it practical for a team without a data scientist. Prompt 10: Post-campaign retrospective Act as a marketing analyst. Help me run a retrospective on a recent campaign. I'll paste the brief and the results. I want you to: summarize what the campaign set out to do, identify the gaps between expectation and outcome, suggest 3 questions I should be asking about why it performed the way it did, and propose 2 test ideas for next time. Do not fill in data I haven't given you. Here is the brief: [paste it]. Here are the results: [paste anonymized performance data] How to get better AI marketing prompt output without better prompts This sounds counterintuitive, but stick with it. The single biggest improvement most marketers can make isn't writing more sophisticated prompts. It's building a reusable context block. A context block is a short document you paste at the top of almost any prompt. It covers your product, your audience, your tone of voice, your key differentiators, and what you're not. Once you have it, you stop briefing AI from scratch every time. You just drop it in. Here's what a minimal version looks like: Company: [name] What we sell: [one sentence] Who buys it: [specific audience, not "businesses" or "marketers"] Why they buy it: [the real reason, not the feature list] Main competitor we're often compared to: [name] Why customers choose us over them: [specific, honest answer] Tone: [2-3 adjectives, with one example sentence] Things we never say: [list 3-5 phrases or approaches to avoid] Takes about 20 minutes to write once. Saves you the same clarification dance every single time you open a new chat. Paste it before any prompt in this list and the output quality goes up noticeably, because you've eliminated the ambiguity that produces generic output. Some teams build a more detailed version that includes real customer quotes, example campaign lines they've approved in the past, and notes about what's flopped. That's the version worth building over time. What you still own, no matter how good the prompts get Here's the thing about AI marketing prompts: none of them replace the judgment calls. AI can draft a landing page. It cannot tell you whether the offer is actually compelling, whether the pricing is right, or whether you're targeting the right audience segment. It can suggest campaign angles. It can't tell you which one aligns with where your company is headed, what the sales team is hearing on the ground, or what your competitors just announced. Dee covers this in Don't Replace Me, calling it the taste moat: the ability to look at a page of plausible output and know whether it's differentiated, true, and worth publishing. That judgment is still yours. The AI doesn't know your customers. It knows what texts about customers generally sound like. You also remain responsible for everything that goes out the door. That means: verifying every product claim before it goes live. Double-checking pricing, legal language, and compliance requirements. Confirming any case study or testimonial actually happened. Making sure any statistic came from a real source, not a model that confidently invented it. Never publish AI-written marketing copy that makes a claim you cannot defend. That's not a disclaimer. That's just basic brand accountability. The people who are building a real advantage with AI marketing tools aren't the ones using the cleverest prompts. They're the ones who understand their customers deeply, have a strong point of view on positioning, and use AI to move faster through the drafting work so they can spend more time on the strategy. If you're curious which parts of marketing humans still genuinely own and why, that breakdown is here. A few things you should never paste into a prompt Before you run any of these prompts: a short list of what not to include. Don't paste real customer names or identifying information. Don't include raw data with private details from your CRM or support system. Don't paste unreleased product plans, pricing strategy, or M&A information. Don't include contract terms, legal agreements, or anything that would give your legal team a headache. Don't paste revenue figures or performance data that's commercially sensitive unless your company has a clear policy allowing it. Anonymize, paraphrase, or summarize the sensitive parts. The AI doesn't need the real names to help you with the copy. This applies to free tools especially. Review OpenAI's enterprise privacy notes and Anthropic's privacy policy if you're using these tools in a professional context. Most enterprise plans have different terms than free tiers. Check what your company is actually using before you paste anything sensitive. Frequently asked questions What's the best AI marketing prompt for writing copy? There's no single best prompt, but the most effective ones follow the same pattern: specify the role (e.g., "act as a direct-response copywriter"), give real context about your audience and offer, set clear constraints on length and tone, and ask for a specific format. Generic prompts get generic output. The prompt for landing page copy in this article (Prompt 4) is a solid starting point for most marketers. Can I trust AI to write my marketing copy? Use it to draft, not to publish. AI can produce plausible-sounding copy quickly, but it doesn't know your customers, can't verify its own claims, and has no idea whether your product actually delivers what the copy promises. Every piece of AI-generated copy needs a human to check it for accuracy, brand fit, and whether it's actually true. Is it safe to paste customer data into ChatGPT or Claude? No. Don't paste real customer names, emails, support tickets with identifying information, or any personal data without proper authorization and an understanding of what the tool does with that data. Anonymize before you prompt. When in doubt, check with your legal or compliance team before pasting anything sensitive. Will AI replace marketing jobs? Not the whole job. AI is getting fast at producing drafts, variations, and summaries. It's not good at knowing which of those variations is worth sending, what the customer actually cares about, or how to position a product honestly in a crowded market. The marketers who will feel the squeeze are the ones whose entire job was producing volume of mediocre copy. Strategy, taste, and customer understanding still require a person. How do I make AI marketing prompts produce less generic output? Run prompt 8 from this article: the copy review prompt that flags unverifiable claims, generic phrases, and jargon. Beyond that, make your prompts more specific, and build a reusable context block with your product, audience, differentiators, and tone that you paste at the top of every chat. Generic output usually means generic input. What should I use AI for in a marketing workflow? Good uses: drafting, brainstorming angles, creating variations, repurposing content, pressure-testing copy against a skeptical editor, building briefs, and running retrospective analysis. Bad uses: sourcing statistics, making product claims, replacing customer research, or hitting publish without review. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Automation Prompts: Automate the Boring Stuff 10 copy-paste AI automation prompts for mapping workflows, writing SOPs, adding approval gates, and QA testing. Practical templates with safety rules included. ai automation promptsworkflow automationSOP templates AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts --- ## AI Meeting Agenda Prompts: 10 Templates That Work URL: https://dontreplace.me/articles/ai-meeting-agenda-prompts-10-templates-that-work ai meeting agendameeting promptsai at workchatgpt promptsmeeting productivity AI Meeting Agenda Prompts: 10 Templates That Work 10 copy-paste AI meeting agenda prompts for decision meetings, kickoffs, 1:1s, retros, and more. Includes the reusable formula behind every good prompt. June 25, 2026 Updated June 25, 2026 Most meetings are not meetings. They're someone's unprocessed anxiety formatted as a calendar invite. You've been in them. The 60-minute "sync" that could have been a Slack message. The "quick alignment call" where nobody agreed on what needed aligning. The kickoff that launched nothing. If you're using AI meeting agenda prompts to fix this, you're on the right track, but only if you're honest about what AI can and can't do here. AI can turn your vague intent into a structured agenda fast. What it can't do is tell you whether the meeting should exist, who actually needs to be there, or what the real decision is hiding behind the polite agenda item labeled "general updates." That part's still on you. What AI meeting agenda prompts can actually do (and what they can't) AI is genuinely useful for meeting prep. Give it a clear goal, some context, and a list of attendees, and it'll produce a tidy agenda with timeboxes, facilitation questions, and a parking lot section faster than you'd type it yourself. What it can't do: read your org chart, understand the political tension between your VP and the product team, know which decision has been quietly made already, or realize that half the people on the invite shouldn't be there. It also can't tell the difference between a real deadline and one somebody invented to create urgency. The risk with AI-generated agendas isn't that they're wrong. It's that they look right. A polished agenda with clean timeboxes can paper over the fact that nobody agreed on the outcome. That's worse than a messy agenda, because now everyone shows up confident, then spends 45 minutes discovering they had different expectations the whole time. Use AI to build the structure. Use your brain to supply the judgment. The formula behind every good AI meeting agenda prompt Before you get to the 10 templates, here's the reusable formula. Every good prompt includes: Context: What is this meeting about and why now? Desired outcome: What does "good" look like when the meeting ends? Decision owner: Who has the authority to make the call? Required attendees: Who needs to be in the room and why? Source of truth: What documents, data, or prior decisions does this meeting build on? Constraints: Time available, topics that are off the table, decisions already made. Timebox: How long is each section? Parking lot: Where do tangents go? The more of this you give the prompt, the less AI has to invent. And AI inventing things is where meetings go wrong before they even start. 10 AI meeting agenda prompts you can copy and use today Prompt 1: Should this meeting even happen? Before you build an agenda, check if you need one. I'm considering scheduling a meeting about [topic]. The goal is [outcome]. The people who would attend are [names/roles]. Can you help me decide: - Could this be resolved async (email, doc, Slack)? - What's the minimum viable meeting if it must happen? - What information do I need before scheduling? Don't suggest the meeting is necessary. Be honest if it's not. Use this before you send a single invite. If the answer is "this could be an email," believe it. Prompt 2: Turn a vague topic into a clear agenda Most meeting requests start as vibes. "We should get everyone together about the launch." Fine. Now what? I need to run a meeting about [vague topic]. Here's the context: [2-3 sentences about what's happening]. Desired outcome: [what we need to decide or align on]. Attendees: [names/roles and why each is there]. Time available: [45 minutes / 1 hour / etc.]. Create a structured agenda with: - A single clear objective at the top - Agenda items with timeboxes - One decision or output per section - A parking lot placeholder - Pre-meeting prep for attendees (1-2 questions they should answer beforehand) Do not add agenda items I haven't mentioned. Flag if the outcome isn't achievable in the time available. The "do not add items I haven't mentioned" instruction matters. AI will helpfully pad agendas if you don't constrain it. Prompt 3: Decision meeting agenda This is the meeting where something actually gets decided. The agenda prompt needs to be explicit about who owns the decision, or you'll end up with a great discussion and zero outcome. I'm running a decision meeting. Decision to be made: [specific decision]. Decision owner (who has final call): [name/role]. Context and background: [relevant docs, prior discussions, constraints]. Options being considered: [list them]. Attendees: [names and their role in the decision: informed, consulted, decider]. Time: [length]. Create an agenda that: - States the decision clearly at the top - Allocates time to surface objections, not just present options - Includes a section for explicit disagreement - Ends with a decision log template (decision, owner, rationale, date, follow-up actions) Do not list anyone as a decision owner unless I've specified it. For more help with decision structure, the AI decision-making prompts article has templates specifically for thinking through options before you even get to the meeting. Prompt 4: Brainstorming meeting agenda Brainstorming meetings are the most likely to go shapeless. Everyone talks, nobody captures anything, and the follow-up email says "great ideas shared!" with zero specifics. I'm running a brainstorming meeting on [topic]. Goal: Generate [number] actionable ideas on [specific question]. Attendees: [names/roles]. Constraints: [budget, timeline, things we've already tried, things that are off the table]. Time: [length]. Create an agenda that: - Opens with a crisp problem statement (not open to debate in this meeting) - Includes structured diverge-then-converge time - Has a method for capturing and ranking ideas during the meeting - Ends with 2-3 ideas selected for follow-up, with a named owner for each Do not pre-populate the ideas themselves. Prompt 5: Project kickoff agenda Kickoffs fail when they're treated as orientation rather than alignment. The goal isn't to inform people. It's to get everyone pointed in the same direction before the first task starts. I'm running a kickoff for [project name]. Project summary: [2-3 sentences]. Key stakeholders in the room: [names and roles]. Known risks or dependencies: [list]. Decisions that have already been made: [list]. Decisions still open: [list]. Timeline: [key milestones]. Meeting length: [time]. Create a kickoff agenda that: - Covers scope, roles, timeline, and success criteria - Surfaces open decisions and assigns owners - Includes a "questions we can't answer yet" log - Ends with next steps and dates Do not imply any budget, headcount, or technical commitments I haven't specified. If this kickoff is part of a larger project, the AI project management prompts article has templates for the ongoing work that follows. Prompt 6: Stakeholder update meeting agenda These meetings exist to inform people who need to know things but don't need to make decisions. The trap is turning them into decision meetings by accident. I'm running a stakeholder update on [project/initiative]. Audience: [names/roles and their relationship to the project]. What's new since the last update: [status, blockers, changes]. What I need from them (if anything): [approvals, awareness, input]. What is NOT on the table for discussion: [list anything already decided]. Time: [length]. Create an agenda that: - Leads with the headline (status in one sentence) - Covers what changed, what's at risk, what's decided - Has a clear section for Q&A - Ends with explicit "no action required" or specific asks by name Do not suggest decisions that are already made are still open. After the meeting, turn the update into a report using the AI status report prompts templates. Prompt 7: 1:1 meeting agenda One-on-ones should not be status meetings. They're for the things that don't fit anywhere else: blockers, development, trust, honest feedback. I'm preparing a 1:1 with [role/relationship, e.g., "a direct report who's been struggling with workload"]. Topics I want to cover: [list]. Things I want to understand better: [questions or concerns]. Context: [recent feedback, projects, dynamics I'm navigating]. Time: [30 or 60 minutes]. Create an agenda that: - Opens with space for them to raise what's on their mind - Covers my topics without dominating the conversation - Includes one forward-looking question ("what would make next week easier?") - Ends with 1-2 clear actions and owners Do not suggest I raise performance issues, salary discussions, or HR matters without specialist guidance. For more on 1:1 structure, the AI manager prompts article covers the full range of management conversations, not just kickoffs. Prompt 8: Retrospective agenda Retros fail when they turn into complaint sessions with no output, or when the facilitator is too close to the work to run an honest one. I'm running a retrospective on [project/sprint/period]. Team: [roles, not names, keep it general]. What went well that we should repeat: [optional starter list]. What was painful: [optional starter list]. What we tried that didn't work: [optional]. Time: [45 minutes / 1 hour]. Create a retro agenda that: - Separates "what happened" from "what we should change" - Has a structured method for surfacing input without groupthink - Ends with 1-3 specific process changes, each with a named owner and a review date - Includes a "close the loop" section on actions from the previous retro Do not editorialize about team performance or assign blame to individuals. The AI retrospective prompts article has the full set of retro templates if you need more depth. Prompt 9: Pre-read questions for attendees Most meetings suffer from attendees who show up cold. A short pre-read prompt fixes this. I'm sending a pre-read for an upcoming meeting on [topic]. The meeting agenda is: [paste agenda]. The key question we're trying to answer: [one sentence]. Background documents they should review: [list links or descriptions]. Write 3-5 questions attendees should be able to answer before they arrive. Keep them short. Focus on the decision or output, not background knowledge. Tell me if the agenda is too complex to prepare for in [X minutes] of pre-read time. Prompt 10: Turn the agenda into notes and action items after the meeting The meeting happened. Now what? Here is the meeting agenda we used: [paste agenda]. Here are my rough notes from the meeting: [paste notes]. Create: - A summary of what was discussed (3-5 bullets) - A decision log: each decision made, who made it, and the rationale - An action item list: task, owner, deadline - A parking lot: items raised but not resolved, with a suggested next step for each Flag any action items where the owner or deadline is unclear. Do not invent owners, deadlines, or commitments that aren't in my notes. The AI meeting notes prompts article has more templates specifically for the notes and follow-up side of this. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What AI should never do in your meeting prep A few things worth being direct about. AI should not be inventing attendees, commitments, deadlines, or budget approvals. If your prompt is vague about who owns something, it'll guess. That guess will look credible. Check every output for assumed context before you send anything to real people. Don't paste sensitive information into AI tools that haven't been approved for it. That means no customer PII, employee performance records, HR issues, confidential financials, legal disputes, board materials, security incidents, contracts, or anything your legal or IT team would be uncomfortable seeing in a consumer AI tool. If the meeting is about sensitive stuff, build the agenda structure in AI first using placeholder descriptions, then add the real details yourself in your actual doc. AI builds the scaffolding. You supply the content that matters. Rule #13 in Don't Replace Me is "Garbage In, Garbage Out," and meeting agendas are a perfect example. A vague goal produces a polished agenda for a meeting that still has no real purpose. The AI made it look cleaner, which made it harder to spot the problem. And Rule #7 applies here too: taste is the moat. Knowing which meeting shouldn't happen, which conversation should be async, which decision is actually already made and just needs announcement rather than discussion, those are judgment calls. AI doesn't have them. You do. That's not nothing. A note on what to review before you hit send Every AI-generated agenda should pass a quick human check: Is the objective specific enough that we'd know if we achieved it? Is the decision owner explicitly named? Are there people on the invite who don't need to be there? Is there anything in the agenda that assumes a decision hasn't been made that actually has been? Would anyone reading this be surprised or blindsided by what's on it? If any of those land wrong, fix it before it goes out. The AI did the formatting work. The accountability is still yours. Frequently asked questions What's the best AI prompt for writing a meeting agenda? The most effective prompts include: the meeting's single desired outcome, the decision owner (if applicable), who's attending and why, any constraints or already-made decisions, and the time available. Without those inputs, AI defaults to a generic structure that fits any meeting and therefore fits none of them. Can AI tell me if a meeting is necessary? It can help you think it through. If you describe your goal and ask whether it could be resolved async, a good AI tool will push back honestly. But the final call is yours. AI doesn't know your org, your team dynamics, or which conversations are too loaded for a doc comment. Is it safe to paste meeting notes into ChatGPT? It depends on what's in them. Meeting notes that include customer PII, employee performance issues, confidential strategy, legal matters, security incidents, or financial details should not go into consumer AI tools unless your organization has approved that tool for sensitive data. When in doubt, use placeholder descriptions and fill in the real details yourself afterward. How do AI meeting agendas handle conflicts between team members? They don't, and they shouldn't try. AI can create structure for surfacing disagreement (an objections section, an explicit dissent log), but it can't navigate interpersonal conflict, political dynamics, or sensitive personnel situations. Those require human judgment, and sometimes HR. What's the difference between an AI meeting agenda and AI meeting notes? An agenda is built before the meeting to set the structure. Meeting notes capture what actually happened: decisions made, actions assigned, things left open. They require different prompts and different levels of review. The AI meeting notes prompts article covers the post-meeting side specifically. Do I need to learn any special skills to use these prompts? No. Copy the prompt, fill in the bracketed sections with your actual context, and review the output before you use it. The skill isn't prompt engineering. It's knowing what a good meeting needs, which you likely already know. You're just using AI to format it faster. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts AI Project Management Prompts That Actually Work 10 copy-paste AI project management prompts for briefs, risk logs, stakeholder updates, and recovery plans. Real templates, no hype. ai project managementproject management promptsai at work --- ## AI Meeting Follow-Up Prompts (10 Templates) URL: https://dontreplace.me/articles/ai-meeting-follow-up-prompts-turn-talk-into-work ai promptsmeeting follow-upaction itemsproductivityai at work AI Meeting Follow-Up Prompts (10 Templates) 10 copy-paste AI meeting follow-up prompts for ChatGPT and Claude. Turn messy notes into recaps, action items, and exec summaries without inventing decisions. July 13, 2026 Updated July 13, 2026 Meeting ended 40 minutes ago. You have four pages of notes, two voice memos, and a Slack message from someone asking "what did we actually decide?" You know what you need: AI meeting follow-up prompts that turn that mess into something people will actually read and act on. The good news is AI is genuinely useful here. The bad news is it's useful in a specific, narrow way that most people get wrong. AI is fast. It's not smart about your meeting. It doesn't know who's actually going to do the thing, whether the "decision" was real or just someone talking loudly, or that the timeline your VP mentioned was aspirational at best. It will write a beautiful follow-up that sounds completely authoritative and contains at least one thing nobody agreed to. That's the trap. A polished recap isn't the same as real accountability. The prompts below help you avoid it. What makes AI meeting follow-ups go wrong? The problem isn't the tool. It's what you feed it, and what you believe it produced. Vague notes create confident fake action items. This is Rule #13 from Don't Replace Me: garbage in, garbage out. If your input is "we talked about the timeline and John said something about Q3," the AI will confidently output "John will deliver the project by Q3." That's not an action item. That's a misquote waiting to cause a fight. The other failure mode is treating structure as agreement. An AI can organize your notes into a clean three-section email with decisions, action items, and next steps in under 30 seconds. That formatting does not mean anyone agreed to those decisions, owns those items, or remembers the meeting the same way you do. Before you paste anything: don't put customer PII, employee records, confidential strategy, unreleased product details, legal disputes, financial forecasts, HR issues, or private client conversations into unapproved AI tools. Check your company's policy first. If the meeting was sensitive, the follow-up might not belong in a public AI tool at all. The reusable formula for AI meeting follow-up prompts Every prompt in this list follows the same structure. Once you understand it, you can write your own. [Role] + [Context] + [Source material] + [Desired output format] + [Constraints] In practice that looks like: "You are a meeting facilitator. Here are my rough notes from a 45-minute product review with our engineering lead and two stakeholders. Turn these into [specific output]. Do not invent owners, deadlines, or decisions that aren't explicitly in the notes. Flag anything ambiguous as an open question." That last sentence does more work than the rest combined. Telling the AI what not to do is as important as telling it what to produce. The other thing worth saying: context matters more than cleverness. A long, elaborate prompt with vague source notes will produce vague output. A simple prompt with specific, detailed notes will produce something usable. Clean up your notes first. Then run the prompt. 10 copy-paste AI meeting follow-up prompts Prompt 1: Turn rough notes into a follow-up email You are a meeting facilitator. Here are my rough notes from [meeting name, date, attendees]: [PASTE NOTES]. Write a professional follow-up email summarizing what was covered. Include: key discussion points, any decisions made, action items with owners and due dates (only if explicitly mentioned in the notes), and next steps. If something is ambiguous or unclear, write it as an open question rather than a stated fact. Do not invent owners, deadlines, or decisions. Keep it under 250 words. Prompt 2: Extract action items with owners and dates Here are notes from a meeting: [PASTE NOTES]. Extract all action items. For each one, list: the task, the named owner (only if explicitly mentioned), the due date (only if explicitly mentioned), and any dependencies. If an owner or date is missing, write "OWNER NEEDED" or "DATE NEEDED" rather than guessing. Flag any action item that seems assumed rather than explicitly agreed. Prompt 3: Separate decisions from open questions Here are my meeting notes: [PASTE NOTES]. Create two separate lists: (1) Decisions that were clearly made and agreed upon during this meeting, and (2) Open questions that were raised but not resolved. Be conservative. If something sounds like a tentative discussion rather than a firm decision, put it in open questions. Do not present uncertainty as resolution. Prompt 4: Write a client recap You are a client success manager. Here are notes from a client meeting: [PASTE NOTES]. Write a professional recap email to send to the client. Summarize what was discussed, confirm any decisions the client made, list agreed next steps with owners and dates, and note any items we're waiting on from their side. Tone: clear and professional. Do not include internal commentary, pricing details, or anything marked [INTERNAL]. If the notes are unclear on a point, omit it rather than guess. Prompt 5: Write an internal team recap Here are notes from an internal team meeting: [PASTE NOTES]. Write a short internal recap for the team. Include: what we covered, what we decided, who owns what by when, and what's still unresolved. Keep it to bullet points. Skip anything that wasn't directly relevant to decisions or work. Flag anything where ownership is unclear so we can resolve it before the next meeting. Prompt 6: Convert a transcript into next steps Here is a meeting transcript: [PASTE TRANSCRIPT]. Read through it and produce only the next steps. Format as: [Owner] will [action] by [date]. Only include items where there's clear evidence of commitment in the transcript. If you're uncertain whether something was agreed, add a note saying "(unconfirmed, needs verification)". Do not pad the list. Prompt 7: Find risks and dependencies Here are notes from a project meeting: [PASTE NOTES]. Identify any risks, blockers, or dependencies mentioned. For each one, note: what the risk or dependency is, who it affects, whether it was flagged as a concern or just mentioned in passing, and whether it needs escalation. If any risk involves legal, security, financial, HR, or compliance issues, flag it clearly so I can escalate to the right person. For this one especially: if the AI surfaces something that looks like a real legal, security, or compliance risk, don't handle it with a follow-up email. Escalate to whoever owns that function before the email goes anywhere. Prompt 8: Draft a no-owner escalation note Here are action items from a recent meeting where owners weren't assigned: [PASTE LIST]. Write a short, professional internal message to the team noting that these items came out of [meeting name] and need owners assigned before [date]. Keep the tone neutral and action-focused. Do not assign blame. Include a simple table: Task | Suggested Owner (blank) | Due Date | Notes. Prompt 9: Create a one-paragraph executive summary Here are notes from [meeting type]: [PASTE NOTES]. Write a single paragraph executive summary. Cover: what was decided, what the key outstanding questions are, and what happens next. Audience: [name or role]. Write for someone who wasn't in the meeting and has 30 seconds. Avoid jargon and hedged language. Be direct. Prompt 10: Check whether the follow-up is safe to send Here is a meeting follow-up I've drafted: [PASTE DRAFT]. Review it for the following: (1) Any statements presented as decided that may actually be open questions, (2) Any owners or deadlines that weren't explicitly agreed, (3) Any language that could be misread as a commitment the company hasn't made, (4) Any sensitive information that probably shouldn't be in a written follow-up. List your concerns. Don't rewrite the email, just flag the issues for my review. Run your own drafts through this one before anything goes to a client, an executive, or anyone outside your team. It takes 20 seconds and saves the kind of conversations nobody enjoys having. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → AI meeting follow-up prompts that need extra care Some meetings shouldn't be processed through an AI tool at all, or need significant stripping before they are. Know the difference. Meetings that are usually fine: general project updates, team standups, planning sessions, retrospectives, internal product reviews with no sensitive details. Meetings that need care: anything involving individual employee performance, compensation, legal strategy, unannounced product decisions, client contract terms, security incidents, financial projections, medical or safety issues, or anything your legal or HR team would be unhappy seeing in a third-party tool. The test is simple: if this follow-up ended up somewhere it wasn't supposed to, would it cause a problem? If yes, either don't use an AI tool, use one your company has approved with proper data handling, or scrub the sensitive parts before pasting. This isn't about being paranoid. It's about not creating liability with a convenience tool. If you're building a broader habit around AI at work, the no-BS starter guide is the right place to start, and what AI can and can't do will give you a clear picture of where it helps and where it confidently lies. How to get better output from these prompts The prompts above are starting points. The quality of what comes back depends almost entirely on what you put in. Here's what actually moves the needle. Write notes as you go, not after. Real-time notes, even messy ones, are infinitely more useful than reconstructed notes written 20 minutes after the meeting ended. The AI can clean up messy. It can't fill in forgotten. If you're capturing discussion live, you'll have specific names, specific commitments, and specific disagreements. Those details are what make the difference between a follow-up that creates accountability and one that just sounds like it does. Label the structure before you paste. A quick [DECISION], [ACTION], [DISCUSSION], [OPEN QUESTION] tag in your raw notes takes 30 seconds and dramatically improves what the AI produces. You're not organizing the notes. You're just flagging intent. The AI will follow the signals. Use the output as a draft, not a finished product. Read everything before it goes anywhere. Not a skim. An actual read. You're the one who was in the room. You know whether "we'll revisit the budget" was a polite dismissal or a genuine commitment to reconvene. The AI doesn't. That judgment call belongs to you. Iterate once if the first output is off. If the draft has the wrong tone, is too long, or misses something, don't start over. Paste the output back in and say: "This is too formal / too long / missing the risk we discussed about the vendor contract. Revise accordingly." One iteration usually fixes it. These aren't special tricks. They're just discipline. The people who get consistently useful output from AI aren't using better tools. They're putting in better inputs and doing a proper review before anything leaves their inbox. What to do after AI writes the draft This is where most people skip a step. The draft looks good. It's organized. It has bullet points and headers and an action item table. So they send it. Don't do that. Read it against your actual notes. Check every named owner. Verify every deadline. Look at every "decision" and ask yourself: was that actually decided, or did someone say it and nobody pushed back? Those are different things. If you took notes from a transcript, check the relevant section for anything the AI flagged as unconfirmed. A confident tone in a follow-up email doesn't create agreement retroactively. It just makes the disagreement happen later, with more confusion about what people thought they signed up for. One other thing worth doing: send the draft to one person who was in the meeting before it goes to everyone. Not for approval, just a quick "does this match what you remember?" That 60-second check has saved more than a few awkward threads. It's also good politics. People trust follow-ups more when they've been verified by someone other than the person who wrote them. For anything that ends up in a permanent record, like a status report or a stakeholder update, the bar is higher. Those documents get referenced for weeks. Make sure what's in them is true. The AI meeting notes templates and decision log prompts pair naturally with this workflow if you want to build the full loop from capture to follow-up to record. Frequently asked questions What's the best AI prompt for writing a meeting follow-up email? Start with your role, context (meeting type and attendees), your raw notes, and a clear instruction not to invent owners, deadlines, or decisions. The most effective prompts explicitly tell the AI to flag ambiguity as open questions rather than resolve it into false certainty. Prompt 1 above is a solid starting point for most professional settings. Can AI reliably extract action items from meeting notes? It can extract items that are clearly stated in the source material. Where it fails is inventing owners, assuming deadlines, and presenting uncertain discussions as confirmed commitments. Always cross-check AI-extracted action items against your original notes before sending. Use Prompt 2 and watch for anything labeled "OWNER NEEDED" as a flag that requires human follow-up. Is it safe to paste meeting notes into ChatGPT or Claude? Depends on the meeting. General project discussions are usually fine. Anything involving individual performance, client contracts, legal strategy, security issues, financial forecasts, or HR matters should not go into an unapproved AI tool. Check your company's AI use policy first. When in doubt, strip sensitive details before pasting. How do I make sure my AI meeting follow-up doesn't create false commitments? Use Prompt 10 (the safety check prompt) before sending anything to clients or executives. It reviews your draft for statements that could be read as commitments you haven't actually made. Also get in the habit of distinguishing "decisions" from "discussions" in your notes before you run them through AI. What's the difference between a client recap and an internal team recap? Audience and content. A client recap confirms what they decided, what you're doing next, and what you need from them. It excludes internal commentary, pricing deliberations, and anything marked for internal use only. An internal recap focuses on ownership, unresolved questions, and honest risk flags. Running them through separate prompts (Prompts 4 and 5 above) keeps that separation clean. How do I handle a meeting where nobody was assigned to own things? Use Prompt 8 to draft a no-owner escalation note that surfaces the gap without assigning blame. Send it quickly, before tasks go cold and people assume someone else is handling it. If this happens frequently, the project kickoff prompts can help establish ownership structure before the meeting ends rather than chasing it afterward. The tidy follow-up email is not the hard part. The hard part is knowing what was actually decided, who actually owns it, and whether the nice-looking document you just sent reflects reality or just sounds like it does. Dmitry Kargaev's field guide Don't Replace Me covers exactly this: how to stay the person who exercises judgment while everyone else mistakes a polished output for actual accountability. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Customer Service Prompts: 10 Templates That Work 10 copy-paste AI customer service prompts for ChatGPT and Claude. Draft replies, de-escalate complaints, and summarize tickets without sounding like a robot. AI promptscustomer serviceChatGPT at work AI Executive Summary Prompts: 10 Templates 10 copy-paste AI executive summary prompts for project updates, board briefs, risk summaries, and more. Includes the 5-part prompt formula. ai promptsexecutive summaryworkplace ai --- ## AI Meeting Notes Prompts: 10 Templates That Work URL: https://dontreplace.me/articles/ai-meeting-notes-prompts-10-templates-that-work AI promptsmeeting notesChatGPT at workAI toolsproductivity AI Meeting Notes Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into summaries, action items, and client emails. With a reusable formula and safety reminders. May 23, 2026 Updated May 23, 2026 Your last meeting produced 47 minutes of audio, three people talking over each other, and one action item that nobody wrote down. You know this because you were there. Now it's two days later and someone is asking for the notes. AI meeting notes prompts can fix this. Not perfectly, not magically, but fast enough to matter. This article gives you a reusable formula and 10 copy-paste prompts you can use today without buying any special software. Why meeting notes are the perfect AI task AI is good at one specific category of work: taking messy language and turning it into clean, structured language. That's it. It can't tell you if the decision was right, it can't know if the client was lying, and it definitely can't read the room when your VP went quiet for 10 seconds after hearing the budget number. But taking a wall of rambling text and pulling out the decisions and action items? That's exactly the kind of pattern-matching it was built for. The Microsoft Work Trend Index found that AI users increasingly report being able to produce work they couldn't a year ago. Meeting notes are the most obvious place to start, because the raw material (the transcript or your rough notes) already exists. You're not asking AI to invent anything. You're asking it to compress and reshape something you already have. That's the right use of the tool. The one formula behind every good meeting prompt Before the 10 prompts, here's what makes any of them work. Rule #13 in Don't Replace Me puts it plainly: useful AI output starts with role, context, task, constraints, and desired format. Think of it as writing a brief for a smart intern who's fast, tireless, and occasionally wrong. Every prompt in this article follows the same structure: Role (who the AI is playing) + Context (what the meeting was) + Task (what you want done) + Constraints (what to avoid or include) + Format (how you want the output) When a prompt gives you garbage output, it's almost always because one of those five pieces is missing. Usually context or constraints. You don't need a $200/month AI meeting tool. You need ChatGPT or Claude and this formula. The full starter guide for using AI at work walks through the same logic for other work tasks if you want to go deeper. Before you paste anything: a safety check This is not a legal disclaimer you should skim. It matters. A lot of meeting transcripts contain things you should not paste into a commercial AI tool: client names, revenue figures, unreleased product details, HR conversations, legal strategy, anything covered by an NDA. Before you paste a transcript anywhere: Check your company's AI policy. Many companies now have one. If yours doesn't, assume the conservative position. Remove customer names and replace with "Client" or "Customer." Strip out specific revenue numbers, headcount figures, or pricing details if they're sensitive. Never paste a recording from an HR, legal, or executive strategy session without explicit clearance. If your company uses Microsoft 365 Copilot or Google's Workspace AI, those tools process data within your company's environment, which is a different security posture than pasting into a public chatbot. Know which situation you're in. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → The 10 prompts Use these as-is or adjust them. The square brackets are your fill-in spots. Prompt 1: Clean up a raw transcript What it's for: You have an auto-generated transcript full of speaker errors, filler words, and run-on sentences. You want it readable before you do anything else. You are an editor cleaning up a meeting transcript for internal use. The meeting was [a weekly project standup / a client kickoff / a team planning session]. Remove filler words, fix obvious transcription errors, and break the text into clear speaker turns. Do not add information that isn't in the transcript. Do not summarize yet. Output clean, readable paragraphs by speaker. Here is the transcript: [paste transcript] Prompt 2: Executive summary What it's for: Your boss wasn't in the meeting and needs the three-sentence version. You are a business analyst summarizing a meeting for a senior leader who was not present. The meeting was [describe topic and attendees]. Write a 3-sentence executive summary covering: what was decided, what is in progress, and what requires a decision or follow-up. Be specific and factual. Do not include anything that wasn't clearly stated in the notes. Here are the raw notes: [paste notes] Prompt 3: Action items only What it's for: You need a clean list of who does what by when. You are an executive assistant extracting action items from meeting notes. For each action item, identify: the task, the person responsible (if named), and the deadline (if mentioned). If a deadline was not mentioned, write "deadline not set." Do not invent owners or dates. Format as a numbered list. Here are the notes: [paste notes] Prompt 4: Decisions log What it's for: Tracking what was actually decided, separate from what was just discussed. You are a project manager documenting decisions from a meeting. Extract only the decisions that were clearly made during this meeting. A decision is something that was agreed upon, not something that was discussed or proposed. For each decision, write: what was decided, who made or approved it (if stated), and any noted conditions or caveats. If something was discussed but not decided, do not include it. Here are the notes: [paste notes] Prompt 5: Risks and blockers What it's for: Flagging what could go wrong before it does. You are a risk analyst reviewing meeting notes. Identify any risks, blockers, dependencies, or concerns that were raised during this meeting. For each one, note: what the issue is, who raised it (if named), and whether any mitigation or next step was discussed. Format as a bulleted list. Do not include items that were mentioned only in passing without any concern being expressed. Here are the notes: [paste notes] Prompt 6: Client recap email What it's for: Sending a professional follow-up to a client after a call without spending 45 minutes drafting it. You are a client-facing account manager writing a follow-up email after a call with [client type, e. g., "a marketing agency client"]. Use a professional but warm tone. The email should include: a brief thank-you, a summary of what was discussed, confirmed next steps with owners and deadlines, and any open questions or items we said we'd follow up on. Do not include internal commentary or anything marked as confidential. Subject line should be specific, not generic. Here are the call notes: [paste notes] Prompt 7: Standup summary What it's for: Turning daily standup notes into a two-minute async read for the team. You are a team lead writing a brief standup summary for a development/marketing/ops team [pick one]. The standup covered [number] people. For each person or workstream, summarize: what they completed yesterday, what they're working on today, and any blockers they mentioned. Keep each entry to 1-2 sentences. Do not pad. Here are the standup notes: [paste notes] Prompt 8: Project status update What it's for: Taking messy internal notes and turning them into a structured update for stakeholders. You are a project coordinator writing a weekly status update for a [type of project] project. The audience is [internal team / executive leadership / a client]. Write the update in the following sections: Overall Status (one sentence with a RAG rating: Red / Amber / Green), Progress This Week, Planned Next Week, Key Decisions Made, and Issues or Risks. Use bullet points within each section. Base this only on what's in the notes. Here are the notes: [paste notes] Prompt 9: Follow-up questions to send What it's for: Identifying what was left unclear that you need to go back and ask. You are a project manager reviewing meeting notes for gaps. Identify questions that were left unanswered, topics that were discussed but not resolved, and anything that seems to be missing context or a decision. Write these as clear, specific questions ready to send in a follow-up email. Do not include questions that were already answered in the notes. Here are the notes: [paste notes] Prompt 10: Turn notes into a task list What it's for: Getting everything into your task manager without retyping it all by hand. You are a productivity assistant converting meeting notes into a task list. Extract every action, task, or commitment mentioned. For each task write: the task description, the assignee (if named, or "unassigned" if not), the due date (if mentioned, or "TBD" if not), and the priority if it was mentioned. Format as a table with four columns: Task, Owner, Due Date, Priority. Here are the notes: [paste notes] These same patterns work across many kinds of work documents, not just meetings. The 12 AI prompt templates for general work tasks covers the same structure applied to emails, reports, and decisions. Quality checks you cannot skip This section is short because it should be obvious, but people skip it anyway. Read the output before you send it. Every time. AI will occasionally invent an action item, misattribute something to the wrong person, or confidently state a decision that was actually still being debated. It does this with complete calm and zero indication that it's guessing. Verify every owner, date, and decision. Cross-check against your own memory or the recording. If you can't verify it, flag it as "to confirm" rather than asserting it happened. Do not pretend a decision was made if it wasn't. If the prompt produces a decision that wasn't in your notes, delete it. "We should probably think about switching vendors" is not a decision. Check the tone on client-facing outputs. The client recap email prompt is useful, but AI defaults to a particular flavor of professional warmth that might not match your relationship with that specific client. Adjust it. The ChatGPT at work prompt guide has more on how to iterate when you get a bad first output. What this won't fix AI meeting notes prompts won't fix a meeting that shouldn't have happened. They won't help if nobody was taking notes at all, if the recording cut out halfway through, or if the "decision" was really just whoever spoke last. They also won't save you from the political dimension of meetings. If the real outcome of that call was that your VP is quietly furious about the budget and nobody said anything directly, no prompt on earth is going to surface that. Understanding what AI can and can't do matters here: the tool compresses language, it doesn't read subtext. What these prompts will do is compress the busywork. If you're spending 30 minutes after every call writing up notes, that's a reasonable task to get to 10 minutes. Use the saved time to actually think about what happened in the meeting, which is the part that requires a human. Frequently asked questions What's the best AI tool for meeting notes? You don't need a dedicated tool. ChatGPT (GPT-4) or Claude work fine with the prompts in this article. Dedicated tools like Otter.ai, Fireflies, or Microsoft Copilot in Teams add transcription and real-time features, which is genuinely useful, but the prompting logic is the same. Start with what you already have before adding another subscription. Is it safe to paste meeting transcripts into ChatGPT? It depends on what's in the transcript and what your employer's AI policy says. Remove customer names, financial figures, and anything covered by confidentiality obligations before you paste. If you're on a paid ChatGPT or Claude plan, your data handling is different from the free tier. When in doubt, check with your IT or legal team. How do I get better output when the prompt doesn't work? Add more context. The most common reason a prompt fails is that the AI doesn't know what kind of meeting it was, who the audience is, or what you actually need. Reread the formula: role, context, task, constraints, format. Something is usually missing. See the no-BS guide to using AI at work for more on this. Should I tell my manager I'm using AI to write meeting notes? Probably, yes, and sooner is better than later. The AI disclosure question at work has a full breakdown, but the short version is: if your company has no policy against it and you're not compromising confidentiality, transparency protects you more than secrecy does. Can AI make up things that weren't in the meeting? Yes. It's called hallucination, and it happens. The model will sometimes invent an owner for an action item or state a decision with more certainty than the conversation warranted. This is why the verification step is not optional. Read the output against your notes or the recording before you send anything. Do I need a meeting recording or can I use rough notes? Both work. A clean transcript gives you more to work with and more accurate attribution. But rough notes, even messy ones, are usually enough for the summary, action items, and email prompts. The cleaner your input, the better your output. That's what Rule #13 in the book means by "garbage in, garbage out." You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Customer Service Prompts: 10 Templates That Work 10 copy-paste AI customer service prompts for ChatGPT and Claude. Draft replies, de-escalate complaints, and summarize tickets without sounding like a robot. AI promptscustomer serviceChatGPT at work AI Executive Summary Prompts: 10 Templates 10 copy-paste AI executive summary prompts for project updates, board briefs, risk summaries, and more. Includes the 5-part prompt formula. ai promptsexecutive summaryworkplace ai AI Meeting Follow-Up Prompts (10 Templates) 10 copy-paste AI meeting follow-up prompts for ChatGPT and Claude. Turn messy notes into recaps, action items, and exec summaries without inventing decisions. ai promptsmeeting follow-upaction items --- ## AI Migration Plan Prompts: 10 Templates That Work URL: https://dontreplace.me/articles/ai-migration-plan-prompts-10-templates-that-work AI promptsmigration planningsystem migrationAI at workproject management AI Migration Plan Prompts: 10 Templates That Work 10 copy-paste AI migration plan prompts for system mapping, dependency checks, rollback criteria, cutover plans, and stakeholder comms. With safety reminders. July 19, 2026 Updated July 19, 2026 A migration that looks clean on paper has ended more than a few careers. The plan was thorough, the checklist was complete, and then someone forgot that the billing integration still pointed at the old server. These AI migration plan prompts exist to help you catch that before production does. AI is genuinely useful here. It can turn your messy bullet points into a structured plan, spot gaps in your dependency list, and draft the stakeholder email you've been avoiding. What it cannot do is know your actual system topology, verify your backups exist, or feel the specific dread of a Friday night cutover gone wrong. That part is still yours. Use these prompts as a planning assistant. Feed them real, sanitized context. Check every output against your actual environment. And do not paste credentials, PII, access tokens, production secrets, or anything your security team would be upset about into any AI tool that hasn't been approved for that data. With that said, let's get you a migration plan that doesn't blow up. The AI migration plan prompt formula Every prompt in this article follows the same structure. Once you understand it, you can adapt any of them for your specific situation. Role + context + constraint + output format Role: tell the AI what kind of expert it's acting as (migration architect, technical writer, QA analyst). Context: give it the facts it needs (systems involved, scope, timeline, known issues). Constraint: tell it what you don't want (no invented integrations, flag gaps instead of filling them). Output format: specify what you need back (structured list, table, plain-language draft). The constraint is the one people skip. Without it, AI will fill gaps with plausible-sounding fiction. A dependency it made up looks identical to one you told it about. The constraint forces it to flag uncertainty instead of hiding it. Prompt 1: Turn messy migration notes into a structured plan You've got scattered notes from three different meetings, a Confluence page nobody updated since last year, and a Slack thread that goes back six weeks. This prompt turns that pile into something usable. Act as a migration planning specialist. I'm going to give you raw notes about a planned system migration. Your job is to organize them into a structured migration plan with the following sections: overview and scope, source and target systems, key milestones, known risks, open questions, and items that require human verification before proceeding. Do not invent details I haven't provided. If something is unclear or missing, flag it as a gap rather than filling it in. Here are my notes: [paste sanitized notes] Review the output for anything it invented. If a milestone appears that you didn't mention, that's AI padding the plan to look complete. Delete it and fill the gap yourself. AI migration plan prompts for system mapping System mapping is where migrations fall apart most often. Not because people don't know their main systems, but because the integrations between them are tribal knowledge living in one engineer's head. These prompts help you surface what you know and identify what you need to find out. Prompt 2: Map source and target systems Act as a systems architect reviewing a migration plan. Based on the system information I'm providing, create a comparison table of source and target systems covering: system name, owner, current version or spec, target version or spec, data it holds, integrations it has, and migration approach (lift-and-shift, rebuild, retire, or TBD). Flag any row where I haven't confirmed the integration owner or the migration approach. Do not assume. Here are the systems I know about: [paste sanitized list] Prompt 3: Find hidden dependencies This one has saved migrations. The prompt doesn't find your hidden dependencies, but it asks the questions that make you find them. Act as a migration risk analyst. I'm migrating [describe system, no PII or secrets]. Based on common dependency patterns for this type of system, give me a checklist of dependency categories I should verify before migrating. For each category, include a question I should answer and who typically owns that answer (e.g., database admin, integration owner, security team). Do not assume I've already checked any of these. Present as a table. Cross-reference the output with your actual architecture. Anything it lists that you can't immediately confirm? That's your to-do list. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Prompts for data handling, access, and backups Data migrations are where compliance teams get involved and careers get affected. Do not skip this section. And again, do not paste real customer data, PII, regulated records, or database schemas containing sensitive fields into any AI tool. Prompt 4: Create a data and access checklist Act as a data migration specialist. I need a pre-migration checklist covering: data classification (what types of data are in scope), backup verification steps, access control changes needed during and after migration, data retention requirements to confirm, and a list of roles that need updated permissions in the target system. Present as a checklist with a "confirmed by" column for human sign-off. I'll fill in the specifics. Give me the structure. Use the structure it gives you. Fill in the actual values yourself from your data classification documentation and access control records. Prompt 5: Draft a cutover plan Act as a migration engineer. Create a cutover plan template for a [type of system] migration. The plan should include: cutover window and duration, pre-cutover verification steps, step-by-step cutover sequence with time estimates and owners, parallel run period if applicable, go/no-go decision point, and immediate post-cutover checks. Use a table format with columns for step, owner, time estimate, and status. Flag any step where I'll need engineering sign-off before we proceed. The owner column matters. Every step needs a real human name or role attached. "TBD" is not an owner. Prompts for rollback, testing, and QA The honest answer to "what's your rollback plan?" should not be "we'll figure it out if we need to." These prompts help you build rollback criteria before the migration starts, not during a 2am incident call. Prompt 6: Create rollback criteria Act as a resilience engineer reviewing a migration plan. Based on a [type of system] migration, help me define rollback criteria. I need: a list of failure conditions that should trigger rollback, rollback decision owner (who has authority to call it), rollback time window (how long after cutover we can still roll back), rollback steps at a high level, and communication steps if rollback happens. Present as a structured document. Flag any criteria that should be validated by an engineering lead before we finalize. Pair this with the AI risk assessment prompts article if you want a fuller risk picture. Prompt 7: Check testing and QA coverage Before you migrate, you need test evidence. Not test plans. Actual results. Act as a QA analyst reviewing pre-migration test coverage. Based on a [type of system] migration, give me a testing checklist covering: functional testing areas, integration testing requirements, data integrity checks, performance baseline comparisons, user acceptance testing scope, and security validation steps. For each area, include a column for test evidence required (e.g., pass/fail report, screenshot, sign-off). Present as a table. Do not mark anything as complete. I'll fill in status. The AI QA checklist prompts article has more depth on the testing side if you need it. Prompts for communications and support readiness Technical migrations fail technically sometimes. They also fail because users weren't warned, support teams didn't know what changed, and nobody drafted the "yes the system is down, here's what to do" message until 11pm on cutover night. Prompt 8: Write stakeholder communications Act as a technical communications writer. I need to draft stakeholder communications for a planned system migration. Audience: [describe: e.g., end users, department heads, external clients]. Write three versions: a pre-migration announcement (timing, what's changing, what users need to do), a day-of reminder (what to expect, downtime window, who to contact), and a post-migration confirmation (what changed, what to do if something doesn't work, who to contact). Use plain language. Flag any section where I'll need to insert specific technical details or get approval from a communication owner. The "flag" instruction matters here. AI will write confident-sounding placeholders. You want it to mark them as placeholders, not hide them in professional-looking prose. Prompt 9: Prepare support and training notes Act as a support team lead preparing for a system migration. I need a support readiness brief covering: what's changing for end users, common questions users will ask and suggested answers, known limitations in the first few days post-migration, escalation path for issues the support team can't resolve, and training or reference materials users will need. Present as a brief I can share with the support team before go-live. Flag anything I need to verify with the engineering team before this goes out. Share this with actual support leads before the migration. Not after. The AI handoff document prompts article has useful templates if you're also managing team transitions around the migration. The go/no-go prompt: your final migration plan check This is the prompt you run the day before cutover. Not to get AI approval, which means nothing, but to force yourself to check whether every critical item actually has a human signature on it. Prompt 10: Run a final go/no-go review Act as a migration readiness reviewer. I'm going to give you my migration plan and current status. Your job is to identify any item that does not have: a confirmed owner, a completed verification step, a documented approval, or a defined next action if it fails. Present the results as a table with three columns: item, current status, and what's missing. Do not tell me the migration is ready. Tell me what still isn't confirmed. That last instruction is the important one. Ask AI whether you're ready to launch and it will usually say yes with a few caveats. Tell it to find what's missing and it will find things. The framing changes the output. For the full launch readiness picture, the AI launch checklist prompts article covers go-live beyond just migration scenarios. What AI can't do in a migration Rule #5 in Don't Replace Me puts it plainly: AI is fast, not smart. Speed is not operational safety. A migration plan that looks clean on a Tuesday afternoon can still hide a broken rollback path, an unconfirmed backup, or an integration owner who hasn't been looped in. AI will not tell you your backup actually completed. It will not know that the "billing-v2" integration in your notes is the old name for three different services. It will not feel the difference between downtime your users can absorb and downtime that triggers an SLA breach. The what AI can and can't do article covers the general shape of this well. In migrations specifically, the things AI can't do are: verify your actual system state, confirm integration owners exist and are available, validate that test evidence covers real failure modes, or sign off on anything. Every one of those needs a human. Your job is judgment. Use AI for structure, translation, gap-prompting, and communication drafts. Keep the accountability for yourself. A note on what not to paste into AI tools Before you run any of these prompts, check what you're about to paste. Do not include: Customer PII or any personally identifiable information Credentials, API keys, access tokens, or secrets Production database schemas with real data Security vulnerabilities or open findings Unreleased product strategy Legal disputes, contracts, or financial records HR issues or employee information Private client conversations Regulated data (health, financial, legal) Anything your security team hasn't approved for the tools you're using If the context you need to give AI requires any of the above, anonymize or abstract it first. "Our CRM migrates to a new platform" is enough context for most of these prompts. You don't need to name the specific vendor, the customer count, or the contract terms. Frequently asked questions Can AI write a complete migration plan for me? AI can draft a structured migration plan from your notes, map known systems, and flag common gap categories. It cannot verify your actual system state, confirm integration owners, or validate that your backups exist. Every plan it produces needs human review against your real environment before anyone treats it as authoritative. What should I never paste into an AI tool when planning a migration? Never paste credentials, API keys, access tokens, PII, production secrets, security findings, regulated data, legal or HR records, or anything your organization's security policy doesn't approve for the AI tools you're using. Sanitize or abstract sensitive context before using these prompts. How do I make sure the AI migration plan reflects my actual systems? Start with a verified source-of-truth inventory from your own documentation, not AI-generated assumptions. Use the prompts to structure and gap-check what you already know. Flag any output where AI has added details you didn't provide, and resolve those gaps through your actual system owners, not by trusting the AI's inference. What's the difference between an AI-assisted migration plan and a safe migration? An AI-assisted plan is a planning artifact. A safe migration requires engineering validation, QA confirmation, security and privacy review, confirmed backups, tested rollback procedures, communication owner sign-off, and management approval. AI can help produce the documents. It can't perform the verification or hold the accountability. How do I use the rollback criteria prompt effectively? Run it before you finalize your cutover plan, not after. Treat the output as a starting template, then validate every rollback trigger and time window against your actual engineering constraints. The rollback decision owner should be a named human being with the authority and availability to make that call at 2am if needed. Where do these prompts fit in a broader project process? Migration planning sits between requirements gathering and launch. For the upstream work, see the AI project kickoff prompts and requirements gathering prompts. For what comes after the plan is drafted, the stakeholder update prompts help you keep people informed through the actual execution. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Launch Checklist Prompts: Ship Without Gaps 10 copy-paste AI launch checklist prompts for ChatGPT and Claude. Turn messy plans into structured checklists with owners, risks, and rollback steps. ai promptslaunch checklistproject management AI Requirements Gathering Prompts: Stop Starting Vague 10 copy-paste AI requirements gathering prompts to turn vague asks into verified briefs. Includes safety rules and a reusable prompt formula. AI promptsrequirements gatheringproject management AI Assumption Log Prompts: Stop Planning on Guesswork Use 10 AI assumption log prompts to expose project guesswork, design evidence checks, assign review owners, and stop unverified beliefs becoming commitments. AI promptsproject managementassumption logs --- ## AI Mobile App Testing Prompts: Catch Device Bugs URL: https://dontreplace.me/articles/ai-mobile-app-testing-prompts-catch-bugs-across-devices AI promptsmobile app testingquality assurancesoftware testingmobile apps AI Mobile App Testing Prompts: Catch Device Bugs AI mobile app testing prompts for device coverage, permissions, offline use, interruptions, accessibility, deep links, and human-verified evidence. August 14, 2026 Updated August 14, 2026 A mobile app can pass every desktop-shaped assumption and still collapse the moment a real phone gets involved. The keyboard covers the submit button. A permission denial creates a dead end. A call interrupts checkout. The connection changes from Wi-Fi to cellular and the app decides the user never existed. AI mobile app testing prompts can help turn verified requirements, supported-device policies, and known risks into a useful test plan. They can organize coverage, propose scenarios, identify unanswered questions, and summarize sanitized observations. They cannot hold a phone, feel a broken gesture, reproduce every hardware quirk, verify an operating system's behavior, or approve a release. AI can draft the map. Humans still have to test the build on real devices and own the decision. Start with a broader AI test plan if the release strategy is unclear. Use AI test case prompts when requirements need traceable checks. Mobile testing adds a messy physical world: radios, sensors, permissions, storage pressure, interruptions, accessibility settings, and many device configurations. Why mobile app testing is its own problem A browser flow usually runs inside a fairly stable window. A mobile app lives on a device the user rotates, locks, backgrounds, starves of storage, disconnects, upgrades, and interrupts. The operating system may kill the process. Notifications may arrive late. Biometrics may fail. Text may be enlarged far beyond the designer's tasteful mockup. Useful mobile coverage can include: installation, first launch, upgrade, downgrade policy, and uninstall behavior; supported phones, tablets, operating systems, screen sizes, and hardware capabilities; app permissions, privacy choices, and changes made in system settings; Wi-Fi, cellular, weak connections, captive portals, airplane mode, and recovery; backgrounding, process termination, screen locking, calls, alarms, and notifications; portrait, landscape, split-screen, gestures, keyboards, and safe areas; large text, screen readers, switch access, reduced motion, contrast, and zoom; storage, memory, battery, thermal, startup, and responsiveness risks; deep links, universal links, push notifications, and external-app handoffs; local state, synchronization, retries, duplicate actions, and conflict handling. The correct matrix depends on actual users and risk. A warehouse scanner app and a meditation timer do not deserve the same test plan. Neither does a banking approval flow and a marketing quiz. The model should structure your known context, not manufacture universal coverage from vibes. For interface-specific checks, pair this guide with AI UI testing prompts. Mobile testing includes the interface, but it also includes the device and operating-system lifecycle around it. The reusable mobile testing prompt formula Use this base prompt, then add one specialized template below: “Act as a mobile testing planning assistant. I am testing [named flow] in app build [version] for [user role] on [approved environment]. Use only the supplied requirements, supported OS/device matrix, permission states, network conditions, accessibility expectations, known risks, and sanitized evidence. Produce [artifact] with setup, actions, expected results traced to a source, observable evidence, cleanup, stop conditions, and unanswered questions. Separate confirmed behavior from hypotheses. Do not invent device behavior, execution results, compliance, severity, or release decisions.” Traceability is the anti-bullshit mechanism. If an expected result cannot be tied to an approved requirement, design, platform rule, or confirmed product decision, label it as a question. Familiar app behavior is not necessarily your app's intended behavior. Never paste credentials, customer PII, production screenshots, private logs, payment data, health information, exact location histories, advertising identifiers, device IDs, push tokens, unreleased builds, vulnerability details, regulated data, or confidential client material into an unapproved AI tool. Use synthetic accounts, approved test environments, redacted evidence, placeholder secrets, and human privacy and security review. What to collect before prompting “Test our app” gets you an enormous generic checklist that nobody will execute. Give the model a narrow, safe packet of verified context. InputWhy it mattersHuman check Named flow and user goalDefines success and scopeProduct owner confirms boundaries Build and environmentPrevents evidence from floating between versionsTester records exact build Supported OS/device matrixBounds compatibility workConfirm current analytics and policy Requirement and approved designSupplies expected behaviorResolve stale or conflicting sources Permission and account statesExposes first-use and denial pathsUse approved synthetic accounts Network and interruption risksCovers real mobile conditionsDefine safe simulations Accessibility settingsChanges layout and interactionTest with actual assistive technology Local and server stateReveals sync and retry risksDefine cleanup and reset steps Evidence and severity rulesMakes findings actionableHumans own judgment and storage Keep unknowns visible. If nobody knows whether an interrupted upload resumes or restarts, the output should contain a decision request—not a beautifully formatted invention. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI mobile app testing prompts Replace bracketed text with verified, non-sensitive information. These prompts create planning artifacts. They do not execute tests. 1. Build a risk-based device and OS matrix “Using our supported OS policy, anonymized device distribution, feature dependencies, user roles, and release risks: [paste], draft a prioritized mobile test matrix. Group coverage into must-test physical devices, targeted secondary devices, emulator or simulator checks, and unsupported combinations. Explain each priority using supplied evidence. Flag missing data instead of inventing market share.” A matrix should reduce risk, not become a museum of every phone ever sold. Prioritize combinations used by customers, devices affected by the change, oldest supported operating systems, relevant screen sizes, and hardware dependencies such as cameras, NFC, biometrics, Bluetooth, or GPS. Emulators are useful for fast breadth. They are weak substitutes for physical checks involving performance, sensors, cameras, notifications, thermal behavior, radios, keyboards, and vendor-specific quirks. A human decides where real-device evidence is mandatory. 2. Test installation, first launch, and upgrades “From these supported upgrade paths, storage requirements, onboarding rules, migration notes, and account states: [paste], create checks for fresh install, first launch, accepted and denied permissions, upgrade from supported prior versions, preserved settings, migrated local data, interrupted upgrade, low storage, logout, uninstall, and reinstall. Cite the expected source for each result.” Upgrade bugs punish existing users—the people who trusted you enough to keep the app. Confirm local data migrations, authentication state, cached content, notification preferences, consent choices, drafts, and feature flags. Do not casually test downgrade paths unless the product supports them. State what should happen when a build cannot read newer data. Back up approved test fixtures and define recovery before experimenting with migration failures. 3. Challenge permissions and privacy flows “Using these verified permission needs and platform requirements: [paste], draft tests for permission not requested, first request, allow, deny, limited access, ask-later behavior, permanent denial, changing the choice in system settings, revoked access while backgrounded, and returning to the flow. Include user explanation, fallback, stored state, evidence, and privacy questions.” Test camera, microphone, photos, contacts, Bluetooth, notifications, location, motion, and other permissions only when relevant. Check that the request appears at a meaningful moment and that denial does not strand the user. The model cannot decide whether collection is legally permitted or proportionate. Privacy, security, legal, and product owners must review the actual data flow. Never reward a generated checklist for confidently normalizing unnecessary access. 4. Test weak, changing, and absent networks “For this flow, API dependency map, retry policy, timeout behavior, and approved network simulation methods: [paste], create tests for offline start, connection loss during read and write actions, weak latency, Wi-Fi-to-cellular transition, captive portal, timeout, server rejection, retry, duplicate prevention, queued work, and recovery after reconnect. Include expected local and server state.” “Works offline” is not one condition. Does the app open? Can the user read cached information? Can they create work? Is it visibly queued? What happens when two devices change the same record? Observe both interface and backend outcome. A success toast does not prove a write landed. A spinner stopping does not prove it failed. Confirm identifiers, timestamps, retries, duplicates, and eventual state using approved tools and synthetic records. 5. Cover interruptions and app lifecycle “Using this named flow, persistence requirements, security rules, and platform lifecycle notes: [paste], draft tests for backgrounding, foregrounding, screen lock, process termination, memory eviction, incoming call, alarm, notification tap, audio route change, session expiry, and device restart. For each, define safe interruption timing, expected preserved state, security behavior, and recovery.” Interrupt at meaningful moments: before submission, during upload, after server acceptance but before confirmation, while editing, and during authentication. Check whether users lose work, repeat an irreversible action, return to stale content, or see sensitive information after a lock. Some lifecycle behavior cannot be forced reliably on every device. Record the method used, distinguish observed from assumed outcomes, and avoid turning one successful run into a universal claim. 6. Test orientation, gestures, keyboards, and safe areas “From these supported orientations, interaction designs, device classes, keyboard types, and layout rules: [paste], create checks for rotation, gesture navigation, system back, edge swipes, notches, camera cutouts, home indicators, software keyboard appearance, hardware keyboard use, autofill, paste, long text, and task completion. Flag unspecified behavior.” Mobile layouts share space with the operating system. The keyboard can cover fields. A bottom action can collide with the home indicator. An edge gesture can fight the app's custom interaction. Rotating midway through a form can reset everything because apparently gravity also clears state. Run these checks on representative devices. Screenshots cannot tell you whether a gesture feels unreliable, focus jumps strangely, or a control is reachable with one hand. 7. Plan mobile accessibility checks “Using our adopted accessibility requirements, supported assistive technologies, semantic design, and named flow: [paste], draft manual checks for large text, display zoom, screen-reader names and order, focus, custom gestures, switch access, voice control, contrast, reduced motion, orientation, error announcements, and task completion. Do not claim conformance from requirements or generated output.” AI can organize an accessibility review. It cannot certify the app. Test with actual platform tools such as screen readers and enlarged text settings, using people with appropriate expertise. Check that content reflows, controls remain available, focus follows meaningful changes, labels distinguish repeated actions, and alternatives exist for gesture-only interactions. False confidence here is not a paperwork issue; it blocks people from using the product. 8. Investigate performance, battery, memory, and storage risks “Given these verified performance budgets, telemetry definitions, device tiers, flow steps, and safe profiling tools: [paste], create an investigation plan for startup, responsiveness, scrolling, memory growth, background work, battery use, network volume, thermal behavior, low storage, cache growth, and repeated sessions. Define measurements and thresholds only when supplied.” Do not ask a language model to estimate battery consumption from source descriptions. Measure the running build with approved profiling tools. Compare like with like: same build, device, OS, account state, network, and scenario. Treat unexplained heat, crashes under memory pressure, runaway cache growth, and lag on supported lower-tier devices as evidence requiring investigation. Humans interpret measurements and decide severity. 9. Verify deep links, notifications, and external handoffs “Using these approved deep-link routes, notification payload types, authentication rules, app states, and external integrations: [paste], draft tests for installed and uninstalled states, logged-in and logged-out users, foreground/background/terminated app, expired content, malformed or unauthorized links, duplicate taps, browser fallback, return from external apps, and analytics expectations.” A link should not bypass authorization merely because it knows the destination. Test what happens when the linked record is deleted, belongs to another account, requires fresh authentication, or points to an old app version. Use synthetic payloads and approved tooling. Push tokens, real customer messages, magic login links, and production analytics identifiers do not belong in a general chatbot prompt. 10. Turn observations into reproducible bug reports “Convert these sanitized, human-observed mobile test notes into draft bug reports: [paste]. For each report include build, device, OS, account and permission state, network, preconditions, exact actions, observed result, sourced expected result, reproducibility, redacted evidence, cleanup, and open questions. Do not invent missing steps, logs, severity, root cause, or customer impact.” Good mobile defects capture the environment. “Screen broken on Android” is not a report; it is a distress signal. Record the exact device model, OS, build, orientation, accessibility settings, permission state, network, and whether the issue survives restart. Use AI bug report prompts for deeper formatting help. The tester must verify the draft against the actual evidence before filing it. How to use these prompts without creating checklist theater First, choose one risky flow—not the whole app. Gather verified requirements and a current device policy. Remove sensitive information. Ask for a draft matrix, then have product, engineering, design, QA, accessibility, privacy, or security owners review the parts they own. Second, execute on approved emulators and physical devices. Record actual outcomes. If the generated plan includes a scenario you cannot safely simulate, mark it blocked or redesign the test. Do not change “not tested” to “passed” because the release calendar has developed feelings. Third, separate three labels in every artifact: Confirmed expectation: tied to an approved source. Observed result: produced by a named human test on a recorded build and device. Hypothesis or question: requires investigation or a decision. That distinction prevents fluent text from laundering assumptions into evidence. It also makes exploratory testing more useful: the model can propose charters, while humans investigate surprises. Finally, keep release ownership human. Use a smoke test for fast build confidence and regression testing prompts for change-focused coverage. Neither replaces accountable sign-off. Frequently asked questions Can AI test a mobile app by itself? Not from a text prompt. An agent connected to approved automation may execute defined actions, but it still needs trustworthy oracles, controlled environments, evidence review, and human ownership. Physical-device behavior, usability, accessibility, privacy, and release judgment remain human work. Should I use emulators or real devices? Use both according to risk. Emulators and simulators provide fast, repeatable breadth. Physical devices are essential for hardware, radios, notifications, gestures, vendor behavior, performance, thermal conditions, and real assistive-technology checks. Document which produced each result. How many mobile devices should we test? There is no magic number. Use supported-platform policy, anonymized usage, feature dependencies, device tiers, OS versions, change risk, and customer impact. A prioritized matrix with explicit gaps beats fifty arbitrary combinations nobody can maintain. Can AI generate mobile test cases from requirements? Yes, as a draft. Require every expected result to cite supplied requirements or be labeled for confirmation. Review for missing states, impossible setup, privacy risks, and invented platform behavior before execution. Is it safe to paste crash logs into an AI tool? Only after approval and careful sanitization. Logs can contain tokens, identifiers, paths, URLs, user content, location, or internal architecture. Follow company policy, remove sensitive values, use approved tools, and preserve original evidence in the authorized system. Can AI verify mobile accessibility? No. It can draft checks and help organize observations. Conformance and usable task completion require actual assistive technologies, qualified human review, the adopted standard, and evidence from the running app. Who decides whether a mobile bug blocks release? The accountable human owners using agreed severity and release criteria. AI may summarize evidence or expose missing facts. It should not assign impact, accept risk, or approve release without human judgment. The useful boundary AI is fast at turning structured facts into candidate coverage. That is valuable. It is not the phone, the user, the accessibility specialist, the privacy reviewer, the engineer, or the release owner. Give it verified context. Demand traceability. Test the real build. Keep evidence separate from guesses. Let humans own consequences. That boundary is the point of Don’t Replace Me by Dmitry Kargaev: use the speed without outsourcing the judgment that makes the work trustworthy. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance AI API Testing Prompts: Catch Broken Integrations AI API testing prompts for contracts, edge cases, authorization, retries, error responses, and human-verified integration evidence. AI promptsAPI testingquality assurance AI Compatibility Testing Prompts: Catch Breakage 10 AI compatibility testing prompts for browser, device, OS, responsive, locale, network, defect triage, and evidence-based release decisions. AI promptscompatibility testingcross-browser testing --- ## AI Onboarding Prompts: 10 Templates for New Hires URL: https://dontreplace.me/articles/ai-onboarding-prompts-10-templates-for-new-hires ai onboarding promptsonboarding templatesnew hire trainingai at workhr ai tools AI Onboarding Prompts: 10 Templates for New Hires 10 copy-paste AI onboarding prompts to build first-week plans, role guides, SOPs, and check-in agendas. With safety rules your HR team will thank you for. June 14, 2026 Updated June 14, 2026 Onboarding is one of the most expensive things a company does badly. A new hire's first week is a blur of Slack channels, acronyms nobody explains, and a 47-page employee handbook that reads like it was written by a lawyer who hates people. Then the manager disappears for two weeks and wonders why the person is still confused in month three. AI onboarding prompts can fix a surprising amount of this. Not the judgment part. Not the relationship part. But the "nobody ever wrote this down" part? AI is genuinely useful there. This article gives you 10 copy-paste prompts for real onboarding tasks, plus the formula for writing your own. It also covers where AI will confidently make things up and ruin your new hire's first impression of the company. The formula behind every good AI onboarding prompt Before the templates, here's the pattern. Every prompt that works follows the same structure: Role + task + context + format + constraints Role: who the AI is acting as ("Act as an experienced L&D coordinator") Task: what you actually need ("Write a first-week schedule") Context: what's specific to this situation ("This is a mid-level marketing hire joining a 40-person SaaS company") Format: how you want the output ("Bullet list, day by day") Constraints: what to avoid ("Do not include specific tool names we haven't approved yet") Vague prompts produce corporate soup. "Write an onboarding plan" gets you something that looks like it came from a LinkedIn carousel. The more context you give, the more useful the output. One more thing before the list: do not paste private employee information into AI tools your company hasn't approved for that use. That means no candidate notes, compensation details, performance history, health information, or anything from an HR incident. Treat the AI like a contractor who does good work but doesn't need to know everything. 10 AI onboarding prompts you can use today Prompt 1: Build a first-week plan Act as an experienced onboarding coordinator. Write a day-by-day first-week schedule for a new [job title] joining a [company size and type] company. Include time for IT setup, team introductions, role overview, tool walkthroughs, and one check-in with their manager. Format as a simple table: Day / Time Block / Activity / Owner. Use this for: Creating a draft schedule before the hire starts. Manager reviews and adjusts. This is a starting point, not a final document. Prompt 2: Explain the role in plain English Rewrite this job description in plain, direct language a smart person would understand on their first day. Remove jargon. Replace vague phrases like "drive cross-functional alignment" with what that actually means in practice. Keep it under 200 words. Here is the original: [paste job description] Use this for: Creating a role summary card new hires can actually read. You wrote the job description for recruiting. This version is for the person who took the job. Prompt 3: Create a team glossary I'm going to give you a list of acronyms and internal terms our team uses. For each one, write a one-sentence plain-English explanation. Flag any you're unsure about with [VERIFY]. Terms: [paste list] Use this for: Turning the "I didn't want to ask" problem into a searchable doc. The [VERIFY] flag matters. AI will confidently invent what an acronym stands for. Every definition needs a human to check it before it goes anywhere official. Prompt 4: Turn an SOP into a new-hire checklist Take the following standard operating procedure and convert it into a numbered checklist a new employee could follow on their own. Use plain, active language. Add a checkbox for each step. If any step is ambiguous or requires a decision, note it as [NEEDS CLARIFICATION]. Here is the SOP: [paste document] Use this for: Making existing documentation usable. Most SOPs are written for the person who already knows what they're doing. This flips that. Prompt 5: Draft a manager check-in plan Write a 30-day manager check-in plan for a new hire. Include suggested talking points for week 1, week 2, week 4, and a 30-day review conversation. Focus on understanding the new hire's experience, answering questions, and giving early feedback. Format as a simple agenda for each conversation. Use this for: Giving managers a scaffold so check-ins actually happen. If you want the fuller framework for running these conversations well, the AI manager prompts guide covers 1:1s, feedback, and team updates in more depth. Prompt 6: Create a practice scenario Write a realistic practice scenario for a new [job title]. The scenario should present a situation they'll commonly face in the first 60 days, give them a problem to solve or decision to make, and ask them three questions about how they'd handle it. Keep the scenario grounded in [industry/department type]. Do not reference specific clients, internal systems, or proprietary process details. Use this for: Building low-stakes practice before someone faces the real thing. Works well for customer-facing roles, where getting it wrong in front of a client is expensive. Prompt 7: Summarize a tool workflow Summarize how to use [tool name] for the most common tasks a new [job title] would need to do in their first month. Use a step-by-step format. If a step requires admin access or a specific permission level, note it. Assume the reader has never used the tool before. Use this for: Turning "you'll figure it out" into an actual walkthrough. Works best when you paste in official help documentation rather than asking AI to recall it from memory. AI gets tool-specific details wrong. Documentation prompts can help you build these from source material. Prompt 8: Build a buddy guide Write a short guide for an onboarding buddy (an experienced teammate paired with a new hire). Include what the buddy role actually involves, suggested topics for their first three conversations, how to answer questions they don't know the answer to, and when to escalate to the manager or HR. Keep the tone friendly and practical. Use this for: Making the buddy system work instead of just naming someone and hoping for the best. The buddy relationship is one of the few things in onboarding that genuinely can't be automated. This prompt helps the buddy show up prepared. Prompt 9: Write a 30/60/90-day expectations draft Draft a 30/60/90-day expectations document for a new [job title]. For each phase, describe what "getting up to speed" looks like, one or two key deliverables or milestones, and what success would look like by the end of that period. Keep expectations realistic for someone still learning the role. Format as three sections. Use this for: Creating a starting draft that the hiring manager edits to reflect the actual role. Do not send this to the new hire unreviewed. AI does not know your team's real priorities or what good performance looks like. That's a judgment call, and it belongs to the manager. This is also where Rule #7 from Don't Replace Me comes in. Taste is a moat. Knowing what "good" looks like in your specific context, and helping someone build that judgment, is exactly the thing AI can't do for you. Prompt 10: Create a safe FAQ from approved docs I'm going to paste content from our approved employee handbook and onboarding materials. Based only on this content, create a FAQ document with the 15 most common questions a new hire might ask, along with accurate answers drawn from the source text. If the answer isn't clearly covered in the source, write "Ask your manager or HR" instead of guessing. Here is the source content: [paste approved content only] Use this for: Building an FAQ that's actually accurate. The "based only on this content" and "ask your manager or HR instead of guessing" instructions are doing real work here. Without them, AI invents answers. On benefits questions, compliance questions, and anything legal, invented answers can cause real problems. See the broader AI training prompts guide for more on building learning materials safely. What AI should never decide in onboarding AI is fast. It's also confident when it shouldn't be. Here's where you do not let it make the call: Policy. AI should draft. HR and legal should approve. Full stop. Benefits and compensation. Never let AI explain what an employee is entitled to. These answers need to come from approved documentation or HR directly. Compliance and security steps. If a step is wrong, the consequences can be serious. Verify every item against your actual security and compliance requirements. Performance expectations. What "good" looks like in your specific team is a manager judgment call, not a prompt output. Legal or employment promises. Anything that could be read as a commitment about employment conditions needs a human and probably a lawyer. Promotion criteria. Do not let a new hire read AI-generated text about career growth and assume it's official policy. The short version: AI drafts. Humans decide. If you're not sure whether something needs review, it does. For a broader look at where AI genuinely helps versus where it confidently makes things up, what AI can and can't do is worth five minutes. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → The privacy rules you can't skip A few things that are easy to forget when you're moving fast: Don't paste into unapproved AI tools: Candidate evaluation notes or hiring committee discussions Compensation details (salaries, bonus structures, equity) Performance improvement plans or disciplinary records Customer data, client names, or deal information Internal credentials, system access details, or security procedures Health or personal information of any kind Unreleased product strategy or anything under NDA If your company doesn't have a policy on which AI tools are approved for which types of data, that's worth sorting out before you build any onboarding workflows. The AI policy prompts guide walks through how to get that on paper. What still requires a human Onboarding done well isn't about the documents. It's about the relationship. The buddy who remembers what it felt like to not know anything. The manager who spots that a new hire is confused and asks before it becomes a problem. The teammate who explains the unwritten rules, the ones that never make it into any handbook. AI can write the checklist. It cannot read the room. It cannot notice that someone seems overwhelmed on Wednesday and adjust. It cannot build the trust that makes a person want to stay. The prompts above are tools for getting the boring structural stuff out of the way faster, so the humans in the process have more time to do the part that actually matters. That's a real use. Don't confuse it for leadership. Frequently asked questions Can I use AI to write a full onboarding plan from scratch? You can use AI to draft the structure, but the content needs to come from your approved materials and be reviewed by a manager or HR before it goes to a new hire. AI doesn't know your actual policies, team norms, or role-specific expectations. Use it to save time on the scaffolding, not to replace the thinking. What should I never paste into an AI tool when building onboarding materials? Avoid pasting anything that identifies specific employees, includes compensation or benefits details, covers performance or HR incidents, contains customer data, or includes security credentials. Stick to content from approved internal documentation and generic role descriptions. How do I stop AI from inventing policy answers in an onboarding FAQ? The most effective way is to include explicit instructions in the prompt: "Answer only from the source content I provide. If the answer isn't clearly covered, write 'Ask your manager or HR.'" This doesn't eliminate hallucination entirely, but it significantly reduces it. Always have HR review any FAQ before it goes to new hires. Should new hire onboarding materials go through HR review even if AI only drafted them? Yes. The origin of a document doesn't change whether it needs review. If a document contains anything touching on policy, benefits, performance expectations, compliance, or employment conditions, it needs to be reviewed by the appropriate people before it reaches a new hire. Is there a good AI tool specifically for onboarding? Most teams use ChatGPT, Claude, or their internal AI platform for this kind of work. There are dedicated HR tools that include onboarding features, but for building materials and practice scenarios, a general-purpose model with clear prompts works well. What matters more than the tool is the quality of the prompt and the human review that follows. How to use AI at work covers how to get started without overcomplicating it. How do I make sure AI-generated onboarding materials are inclusive? Add inclusion requirements directly to the prompt: "Use gender-neutral language. Avoid idioms or cultural references that may not translate. Use plain language accessible to non-native English speakers." Then have someone review the output with that lens. Inclusive onboarding is a policy and culture question, not something you can fully solve with a prompt. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Automation Prompts: Automate the Boring Stuff 10 copy-paste AI automation prompts for mapping workflows, writing SOPs, adding approval gates, and QA testing. Practical templates with safety rules included. ai automation promptsworkflow automationSOP templates AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts --- ## AI Operational Readiness Prompts: 10 Templates URL: https://dontreplace.me/articles/ai-operational-readiness-prompts-find-the-gaps operational readinessAI promptslaunch readinessgo-live checklistrisk management AI Operational Readiness Prompts: 10 Templates 10 copy-paste AI operational readiness prompts for go-live checklists, dependency checks, owner gaps, and stakeholder updates. Includes the core formula. July 28, 2026 Updated July 28, 2026 A launch is never actually ready. Someone always has a note in a doc that says "TBD – check with Dave" and Dave has been on PTO since Tuesday. Your AI operational readiness prompts won't fix Dave, but they will make every gap visible before you go live instead of after. That's the whole point of operational readiness work: find the holes while you can still do something about them. The problem is that most readiness reviews are either too shallow (a 12-item checklist that everyone rubber-stamps) or too chaotic (three Slack threads, a spreadsheet from Q3, and a meeting where nobody took notes). AI can help you build the structure. What it can't do is verify anything. That part is still yours. Here's a reusable formula and 10 copy-paste prompts to get through a readiness review without missing something that bites you at 11pm on launch night. The AI operational readiness prompt formula that actually works Before the individual prompts, one rule: garbage in, garbage out. If you paste in stale SOP notes written by someone who left six months ago, AI will organize them beautifully and they'll still be wrong. Every prompt below assumes you're supplying current information, real names, and honest status updates. The formula is: "I'm preparing for [what is going live / what is changing]. Here is [context you're providing]. Help me [specific output you want]. Flag anything that looks incomplete, unowned, or unclear. Do not invent facts, owners, system states, or approvals." That last sentence matters. AI tools will fill gaps with plausible-sounding fiction if you let them. The explicit instruction to flag rather than invent keeps outputs honest. One more thing before you start. Do not paste customer PII, credentials, access tokens, production logs, private tickets, unreleased strategy, legal disputes, security vulnerabilities, financial records, HR issues, regulated data, medical information, or private client conversations into any AI tool that hasn't been formally approved for that data type. Know your organization's AI use policy. If you're not sure what's approved, ask before you paste. Prompt 1: Build a go-live readiness checklist Use this when you're starting a readiness review and need to make sure you haven't forgotten a whole category. "I'm preparing a go-live readiness checklist for [brief description of what's launching: feature, system, process, product]. The launch affects [who: customers, internal teams, partners]. Key areas I know about: [list what you already have]. Generate a structured readiness checklist covering technical, operational, support, communications, monitoring, rollback, and stakeholder approval. Flag any category where my notes are missing or vague. Do not invent owners or system states." What you get back: a structured checklist with visible gaps. What you still need to do: assign a real human to verify each item. For more on shipping without gaps, the launch checklist prompt templates cover this from a different angle. Prompt 2: Find missing owners Unowned items don't get resolved. They get politely ignored until they become incidents. "Here is my current readiness checklist: [paste checklist]. For each item, I've noted the owner where I know it. Identify every item that has no named owner, a vague owner (like 'team' or 'TBD'), or a single-point-of-failure owner with no backup. Return a list of unowned or at-risk items and suggest what type of role should own each one. Do not assign owners yourself." This prompt is specifically designed not to let AI invent accountabilities. You still have to go find the actual humans. Single-point-of-failure owners are worth flagging separately. It's not enough to have a name. If that person is sick on launch day, who picks it up? AI won't know. But it will surface the fact that you haven't written a backup down anywhere. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Prompt 3: Check dependencies Dependencies are the most common source of launch-day surprises. "We assumed X was ready" is how postmortems start. "Here is a list of the systems, teams, vendors, and approvals our launch depends on: [list]. For each dependency, I'll note what I know about its status: [add statuses where you have them]. Identify any dependencies with unclear status, no confirmation, no fallback, or a timing conflict with our go-live window. Flag dependencies where I've noted 'assume' or 'should be' rather than confirmed. Do not invent status information." If the dependency list is long, do it in sections. Trying to paste an entire architecture diagram and hoping AI summarizes it reliably is optimistic. Pay particular attention to vendor dependencies. Internal teams will tell you when something's broken. Vendors will not, unless you ask them directly and in writing. If you've got a third-party API, a payment processor, or an external data feed in your dependency list, "confirmed" means a reply from an actual human at that vendor, not a green status page from last week. Prompt 4: Pressure-test support coverage "We are launching [brief description] on [date/time window]. Our expected support load based on [historical data or estimates] is [volume]. Here is our support coverage plan: [paste plan]. Identify gaps in: staffing levels, escalation paths, knowledge base coverage, handoff points between shifts or teams, and whether the on-call rotation is confirmed and named. Flag anything that looks like an assumption rather than a confirmed arrangement. Do not fabricate staffing numbers or SLA commitments." Support readiness is the one that gets skipped most often in a crunch. Then the feature launches and three people are handling a queue that needs eight. The knowledge base gap is the one that causes the most damage in that first hour. Support staff can be on-call and present but completely unable to help customers because the article doesn't exist yet, or it exists but describes the old version of the feature. AI can help you check whether your KB topics map to your customer-impact risks. It can't tell you whether the articles are accurate. Prompt 5: Review customer-impact risks "Here is a description of what we're launching and how it changes the customer experience: [describe the change]. Here are the customer segments affected: [list]. Help me identify potential customer-impact risks: confusion, broken workflows, data visibility issues, communication gaps, or trust problems. For each risk, flag whether we have a mitigation in place, who owns the customer response if the risk materializes, and whether the customer comms plan covers it. Do not invent customer data, complaint volumes, or segment sizes." This pairs well with the risk assessment prompt templates for a fuller picture of blast radius. Prompt 6: Turn SOP notes into readiness questions Sometimes the readiness gap isn't a missing checklist item. It's a SOP that nobody has actually read in a year. "Here are notes from our standard operating procedures relevant to this launch: [paste SOP excerpts, no confidential data]. Convert these SOP notes into a set of readiness verification questions that a team lead or ops manager should be able to answer before go-live. Questions should test whether the SOP is understood, current, and executable, not just whether it exists. Flag any section of the SOP that is vague, outdated-sounding, or references roles or systems that might have changed." This is surprisingly useful. SOPs that reference "the legacy billing system" or "contact the operations lead Sarah" (Sarah left in 2022) are everywhere. You just don't notice until someone tries to follow the SOP at 11pm with an incident in progress. For building proper SOPs from scratch, the SOP prompt templates walk through that process separately. Prompt 7: Prepare a stakeholder readiness update "I need to write a readiness update for [audience: executive team, steering committee, client, etc.] ahead of our [launch/change/go-live]. Here is the current readiness status: [paste summary]. Here are the open items: [list with owners and target resolution dates]. Write a structured update that covers: what's confirmed ready, what's open with a clear resolution path, what's blocked and needs a decision, and what the go/no-go recommendation is based on current status. Do not imply anything is resolved that isn't. Do not add false reassurance." The stakeholder update prompt templates have more variations on this format for different audiences. What AI gets wrong about operational readiness (and why it matters) This is the part most AI productivity articles skip, so it's worth saying plainly. AI will produce a beautiful, complete-looking readiness document from incomplete inputs. It doesn't know that the monitoring dashboard you referenced hasn't been updated since the last platform migration. It doesn't know that the person listed as the security approver left last month. It doesn't know that "confirmed with vendor" means one Slack message that nobody replied to. It will organize all of this into a professional-looking summary and you will feel prepared. Rule #5 in Don't Replace Me puts it bluntly: it's not smart, it's fast. Speed is valuable. But tidy notes are not verified readiness. The human value here is knowing which gaps are real and which risks are theatrical. A polished checklist can hide a critical dependency just as effectively as a chaotic spreadsheet can. There's a specific failure mode worth naming. When AI generates a readiness checklist from your notes, it creates a document that looks authoritative. That document then gets shared. People see it, assume someone verified it, and stop asking questions. The checklist becomes the evidence of readiness rather than a prompt for the work that proves readiness. That's the opposite of what you want. Treat AI like a sharp intern who's good at structure and terrible at judgment. They'll give you the framework. You supply the facts. Prompt 8: Define monitoring and escalation triggers "We are going live with [brief description]. Here are the metrics we plan to monitor post-launch: [list metrics and sources]. Help me define: specific threshold values that should trigger an alert (I'll fill in the actual numbers based on our baselines), escalation paths for each alert type, who should be notified at each escalation level, and how we define a severity level that requires a rollback decision. Do not invent thresholds, escalation contacts, or SLA numbers. Leave those as blanks for me to fill in." The prompts above pair directly with the post-launch monitoring templates for the post-go-live phase. Prompt 9: Decide ready vs. blocked vs. needs review "Here is my complete readiness checklist with current status for each item: [paste checklist]. Categorize every item as: READY (confirmed, owned, evidenced), BLOCKED (cannot proceed without resolution), or NEEDS REVIEW (has an owner and a plan but isn't confirmed yet). Based on this categorization, summarize: how many items are in each category, which BLOCKED items are on the critical path, which NEEDS REVIEW items have a resolution deadline before go-live, and what the overall readiness posture looks like. Do not change any status I've provided. Do not imply NEEDS REVIEW items are READY." For the actual go/no-go decision conversation, the go/no-go decision prompts are built specifically for that moment. The three-category system matters because "not ready" has two very different flavors. A BLOCKED item needs a decision or an action before anything moves. A NEEDS REVIEW item has a path, just not confirmation yet. Conflating those two creates the wrong urgency in the wrong places. AI won't know the difference. You annotate the status, it organizes the picture. Prompt 10: Convert readiness gaps into action items "Here are the open gaps from my readiness review: [paste gaps with current status]. For each gap, help me draft an action item that includes: what needs to happen, who should own it (I'll assign the actual person, use [OWNER] as a placeholder), the latest acceptable resolution date, and what 'done' looks like so the item can be closed. Format as a table. Do not invent owners, dates, or definitions of done." The output gives you a clean action register. Before you share it, add the real owners, verify the dates against the actual go-live timeline, and confirm what done actually looks like with the people doing the work. The rollback plan prompts are worth running in parallel, because gaps that don't close in time become rollback decisions. How to use AI operational readiness prompts across different launch types The 10 prompts above apply broadly, but the emphasis shifts depending on what you're launching. Software feature or product release: Prompts 3 (dependencies) and 8 (monitoring triggers) are the most critical. Technical launches have more invisible dependencies than any other type, and the most common failure is discovering a monitoring gap after the fact. Run those two first. Process or policy change: Prompts 6 (SOP review) and 7 (stakeholder update) carry more weight. The risk isn't usually technical. It's people not knowing the new process exists, or knowing it exists but not understanding what's changed for them specifically. Vendor or system migration: Prompt 3 (dependencies) and Prompt 9 (ready/blocked/needs review) are your anchors. Migrations have the longest dependency chains and the most external parties who have their own definitions of "done." Confirm every external status directly. Organizational or operational change: Prompt 2 (missing owners) and Prompt 5 (customer-impact risks) are the ones that bite hardest. Org changes create ownership vacuums that don't show up until something goes wrong and nobody knows who to call. In every case, run Prompt 10 last. The action register is how you turn a readiness snapshot into forward movement. Final human sign-off is not optional Every readiness review needs a person who says "yes, we're going." Not a checklist. Not a summary. A human who has reviewed the evidence, understands the open items, accepts the residual risk, and is accountable for the call. AI can prepare you for that conversation. It can't replace it. Anyone who tells you otherwise is selling something. The difference between a launch that goes wrong and a launch that goes wrong but gets fixed fast usually isn't the quality of the checklist. It's whether someone actually owns the call and knows what they signed up for. Build that accountability into the process before you're in the middle of an incident. If you want the broader framework for staying useful while everyone around you mistakes a neat checklist for actual readiness, that's exactly what Don't Replace Me is built for. Frequently asked questions What are AI operational readiness prompts? AI operational readiness prompts are structured instructions you give to tools like ChatGPT or Claude to help you build checklists, find missing owners, check dependencies, and prepare stakeholder updates before a launch or major change. They speed up the structure work. They don't verify facts, confirm approvals, or replace human judgment on go/no-go decisions. Can I use AI to run a readiness review without human oversight? No. AI will organize whatever you give it, but it can't confirm that a system is healthy, that an owner has actually accepted accountability, or that an approval is real. Every AI-generated readiness output needs a human to verify the facts before it's used to make a launch decision. What should I never paste into AI tools during a readiness review? Don't paste customer PII, credentials, access tokens, production logs, unreleased financials, HR records, legal disputes, security vulnerabilities, regulated data, or private client information into any AI tool that hasn't been formally approved for that data type. Check your organization's AI use policy first. How do I make AI operational readiness prompts more accurate? Supply current, specific information rather than stale docs or vague placeholders. Use the explicit instruction "do not invent facts, owners, or system states" in every prompt. Treat AI output as a draft that needs verification, not a finished deliverable. What's the difference between a readiness checklist and a go/no-go decision? A readiness checklist tells you where you stand. A go/no-go decision is the human judgment call about whether the risk profile is acceptable given what the checklist shows. See the go/no-go decision prompts for the decision conversation specifically. How do I handle a readiness gap that can't be resolved before go-live? Escalate it. Document it. Make sure the decision-maker who signs off on go-live explicitly accepts the residual risk. A gap that's invisible is a problem. A gap that's acknowledged, owned, and accepted with a mitigation plan is a managed risk. Those are very different things. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Contingency Plan Prompts: 10 Templates 10 copy-paste AI contingency plan prompts for ChatGPT and Claude. Map failures, assign owners, draft comms, and verify assumptions before a crisis hits. contingency planningAI promptsbackup plans AI Dependency Register Prompts: Keep Projects From Surprising You Use AI dependency register prompts to map owners, blockers, assumptions, and downstream impact before one small change becomes a three-week incident. AI dependency mappingproject managementAI prompts AI Go/No-Go Decision Prompts (10 Templates) 10 copy-paste AI go/no-go decision prompts for launch readiness. Structure evidence, find gaps, draft memos. AI helps you think. Humans make the call. go/no-go decisionlaunch readinessAI prompts --- ## AI Performance Review Prompts That Work URL: https://dontreplace.me/articles/ai-performance-review-prompts-that-actually-work performance reviewAI promptsself-reviewChatGPT at workcareer AI Performance Review Prompts That Work 10 copy-paste AI performance review prompts that turn your messy notes into polished self-review language. Real templates, no fluff. May 25, 2026 Updated May 25, 2026 Performance review season. The time of year when you're expected to summarize twelve months of work in three paragraphs, sound confident without sounding arrogant, and somehow remember what you did last February. And you have to do all of it in "professional" language that doesn't make you sound like a robot or a LinkedIn influencer. AI performance review prompts can cut that process in half. Not by inventing your accomplishments. You still have to do that part. But by taking your messy notes, half-remembered wins, and "I know I did something good here" brain fog and turning them into clean, credible review language. Here's how to actually use it. What AI is good at in this context (and what it's not) AI is fast at language-shaped work. Drafting, rewriting, restructuring, adjusting tone. Give it raw material and a clear instruction, and it'll produce a usable first draft in seconds. That's it. That's the whole trick. What it can't do: remember your year, verify your numbers, know your manager's priorities, or understand the political context of why you missed that Q3 goal. All of that stays with you. So the process works like this: you bring the evidence, AI shapes the language. If you walk in with nothing and ask AI to write your self-review from scratch, you'll get generic corporate filler that sounds like it could belong to anyone in any industry. Your manager will notice. Don't do that. One more thing before the prompts. Don't paste confidential information into any AI tool unless your company policy explicitly allows it. Remove client names, internal project code names, specific financial figures tied to deals, HR information, and any private employee details. Use "[client name]" or "[Project X]" instead. Most AI tools don't need the real names to help you write better sentences. The reusable formula behind every good AI performance review prompt Every prompt in this list follows the same structure. Once you see it, you can write your own. Role: Tell the AI what kind of reviewer you are (industry, level, function). Task: Tell it exactly what you want (rewrite this, structure this, make this more concise). Input: Give it your raw material. Constraints: Tell it what to avoid (jargon, puffery, passive voice, anything that sounds like ChatGPT wrote it). Format: Specify the output (2-3 sentences, bullet points, 150 words, etc.). This is Rule #13 from Don't Replace Me: garbage in, garbage out. The more context you give, the better the output. Treat it like a smart intern who's fast, available at midnight, and completely clueless about your specific job until you brief them properly. 10 copy-paste AI performance review prompts These are ready to use. Replace the bracketed parts with your actual details. Then read the output, fix what's wrong, and rewrite the whole thing in your own voice before submitting. Prompt 1: Turn project notes into accomplishments You're helping a [job title] at a [industry] company write their annual self-review. I have rough notes about a project I worked on. Turn them into 2-3 polished accomplishment statements that describe what I did, the challenge I faced, and the result. Avoid buzzwords and passive voice. Here are my notes: [paste your rough notes] Prompt 2: Rewrite a vague win with more specificity I have this sentence from my self-review: "[paste your sentence]" It's too vague. I want to rewrite it to be more specific and impact-focused. Here are some additional details I can add: [add any metrics, timelines, context you have] Write 2-3 versions I can choose from. Keep each under 40 words. Prompt 3: Use the CAR structure (Challenge, Action, Result) Help me structure this accomplishment using the CAR format: Challenge, Action, Result. Here's what I did: [describe what happened in plain language] Write a 3-sentence version that follows that structure. Don't use corporate jargon. Sound like a competent human wrote it. If you've used this format for resume bullets before, it transfers directly. The AI resume bullet prompts article has more variations on this structure if you want to go deeper. Prompt 4: Summarize feedback you received I received the following feedback from colleagues or my manager over the past year. Summarize this into 2-3 themes I could reference in my self-review without copying the exact wording. Feedback: [paste anonymized feedback notes or paraphrased comments] Keep the summary factual and specific. Avoid generic phrases like "team player." Prompt 5: Write a promotion-focused self-review I'm writing a self-review with the goal of making a case for a promotion from [current role] to [target role]. Here are my main accomplishments this year: [list them] Here are examples where I operated above my current level: [list them] Write a 250-word self-review section that frames this as readiness for the next level. Don't make it sound like a sales pitch. Make it sound like evidence. Prompt 6: Document collaboration and cross-functional work Help me write 2-3 sentences about my collaboration and teamwork this year. Here's the raw information: [describe who you worked with, what you helped them with, what the outcomes were] Make it specific and avoid phrases like "collaborated effectively" or "worked cross-functionally" without backing them up. Prompt 7: Explain a missed goal without sounding defensive I need to address a goal I didn't fully hit in my self-review. Here's what the goal was: [describe it] Here's why it didn't happen: [be honest, include what was in and out of your control] Here's what I learned or what changed as a result: [describe it] Write a 3-4 sentence version I can include in my review that's honest, takes ownership where appropriate, and doesn't sound like I'm making excuses. Don't be dramatic. Keep the tone professional and matter-of-fact. This is one of the harder prompts to get right, because the AI doesn't know your manager or your company culture. Use the output as a draft, then adjust the tone based on what you know about how these things land. Prompt 8: Convert a brag doc into review language I've been keeping notes about my work throughout the year. Here's a list of things I did: [paste your brag doc / running list] Turn this into polished self-review language. Group related items where it makes sense. Write in first person. Avoid passive voice and corporate jargon. Target length: [200 / 300 / 400 words, your call]. If you don't have a brag doc, start one now. A running note in your phone works fine. Just log things as they happen instead of trying to reconstruct the year in November. Prompt 9: Make review language more concise This paragraph from my self-review is too long. Rewrite it to be tighter and more direct. Cut anything that's filler. Keep every sentence earning its place. Here's the paragraph: [paste it] Target: under [word count] words. Prompt 10: Final honesty check before submission I'm about to submit this self-review. Read it and flag anything that: - Sounds exaggerated or hard to believe - Is vague or unsupported by evidence - Uses corporate jargon or filler phrases - Sounds like it was written by AI Here's my self-review: [paste it] Give me a list of specific lines to revisit, with a brief note on why. This last one is genuinely useful. AI is good at pattern-matching for "this sounds like fluff." Use it as a sanity check, then make the final calls yourself. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What makes AI performance review prompts actually work Most people use AI wrong for this. They open a chat window, type "write my performance review," stare at the generic output, decide AI is useless, and go back to staring at a blank Google Doc. The problem isn't the tool. It's the input. There are three things that separate a useful AI output from a filler-filled mess. Specificity of the role. Saying "I'm a project manager at a tech company" gives the AI something to work with. "I'm a senior technical program manager at a Series B SaaS company overseeing enterprise integrations" gives it a lot more. The more specific you are about your context, the more calibrated the language will be. Quality of your raw material. If you give the AI a two-sentence description of a project, you'll get a two-sentence result with padding. If you give it a paragraph of messy notes that includes timelines, what went wrong, what you did about it, and what happened next, you'll get something much closer to usable. The AI can't hallucinate your year into existence. But it can organize and elevate what you actually give it. Clear output constraints. "Write my review" gets you something generic. "Write three bullet points, each under 30 words, in past tense, with at least one specific outcome in each" gets you something you can actually use. The constraints feel like extra work. They're not. They save you editing time on the back end. This is also why the prompts above have specific lengths and formats baked in. Not because there's magic in "40 words," but because any constraint forces the AI to prioritize instead of rambling. When your review has to go through a system, not just a manager A lot of companies now run performance reviews through HR platforms, not just a conversation. Workday, Lattice, 15Five, Culture Amp. These systems often have specific competency categories, rating scales, and character limits that don't care about your carefully structured CAR format. The prompts above still work, but you need one extra step. Before running any of them, check what the actual form asks for. Copy the section headings and any guidance text from the platform into your prompt, and tell the AI to structure the output to match. Something like this: I'm filling out a performance review in [platform name]. The section is called "[exact section heading]" and the guidance says: "[paste the prompt text from the form]" There's a [X] character limit. Here's my raw material: [your notes] Write a response that fits the format and stays under the character limit. This is especially useful for competency sections, where the platform wants you to rate yourself on things like "drives results" or "builds relationships" and then provide evidence. AI is good at matching your real examples to those abstract categories. You're still supplying the examples. It's just helping you translate. How to use these prompts without getting burned A few things worth saying directly. Verify every fact before you submit. If you give AI a rough note that says "we hit our targets" and it writes "exceeded targets by 20%," that number came from nowhere. Check every metric, date, project name, and outcome claim. If you can't back it up with a Slack message, an email, or a spreadsheet, take it out. Rewrite the output in your voice. The most common mistake is copy-pasting AI output directly into the review form. Your manager has read enough AI-generated text to recognize the patterns. More importantly, your voice matters. Reviews that sound generic get treated as generic. Take the AI draft as scaffolding, then rebuild it sentence by sentence in language that actually sounds like you. Don't let it invent context. AI will sometimes add plausible-sounding details you didn't give it. It's pattern-matching, not remembering. If something appears in the output that you didn't put in the input, delete it. The political stuff is still yours. Whether to emphasize a specific project, how to frame a difficult manager relationship, what not to mention because it'll open a conversation you don't want. That's judgment. AI has no idea. For the broader question of what your employer knows about how you use AI at work, the should-you-tell-your-boss-you-use-ai article covers that territory in more detail. For more general prompt templates you can use across different work tasks, the 12 AI prompts for work article has a broader set that covers emails, decisions, and summaries. What tools should you actually use for AI performance review prompts? ChatGPT or Claude. That's it. You don't need a "performance review AI app" or a specialized HR writing tool. Those are wrappers around the same underlying technology, usually with worse interfaces and a subscription fee on top. The free tiers of both tools handle this task just fine. You're working with text, not images or code or anything that requires the premium versions. If you're already paying for ChatGPT Plus or Claude Pro for other work tasks, great, use those. If you're not, don't buy a subscription just for review season. One thing that does matter: session memory. If you're doing multiple rounds of drafting, you want to stay in the same conversation window so the AI has context from your earlier inputs. Starting a new chat every time you try a different prompt means re-explaining your role and situation from scratch. Work in one thread, build on earlier outputs, and only start fresh when you're moving to a completely different section. If you want to get more out of either of these for general work tasks, the how-to-use-chatgpt-at-work prompts guide covers the setup and the framing in more detail. The short version: paste your context, give clear instructions, review the output critically, rewrite before submitting. That's the whole system. No course required. Frequently asked questions Can I use AI to write my entire performance review? You can use it to draft and structure the language, but the content has to come from you. AI can't know what you accomplished, how your team works, or what your manager values. Use it as an editor and a drafting tool, not a ghostwriter working from nothing. Will my manager know I used AI to write my self-review? Probably not if you rewrite the output in your own voice. If you copy-paste AI text without editing, experienced reviewers often notice the patterns. The goal isn't to hide AI use, it's to use it well enough that the final product sounds like you. What information should I NOT paste into an AI tool for my review? Don't paste client names, private employee information, confidential financial figures, unreleased product details, or anything covered by your company's data policy. Use placeholders like "[client]" or "[Project X]" instead. Most AI tools don't need the real names to help you write better sentences. Is the CAR format (Challenge, Action, Result) actually useful for self-reviews? Yes. It forces you to move from vague claims ("I worked on the rebrand") to structured evidence ("The brand refresh had stalled for six months. I took ownership of the stakeholder alignment process and got sign-off in three weeks, which unblocked the launch."). It works for both reviews and resume bullets. What if I don't have good notes about my year? Start with whatever you have: emails, Slack messages, project files, meeting notes, old to-do lists. Give the AI fragments and ask it to help you identify themes. Then go back and verify which things are actually strong enough to include. A brag doc you start today will make next year's review much easier. How long should my prompts be? Long enough to give the AI real context. A one-sentence prompt gets a generic output. Include your role, the specific task, your raw material, and any constraints on tone or length. Two to four sentences of setup plus your actual content is usually enough. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Customer Service Prompts: 10 Templates That Work 10 copy-paste AI customer service prompts for ChatGPT and Claude. Draft replies, de-escalate complaints, and summarize tickets without sounding like a robot. AI promptscustomer serviceChatGPT at work AI Meeting Notes Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into summaries, action items, and client emails. With a reusable formula and safety reminders. AI promptsmeeting notesChatGPT at work AI Resume Bullet Prompts That Actually Work 8 copy-paste AI prompts for resume bullets that get results. Includes the CAR formula, before/after examples, and safety rules you can't skip. resumeAI promptscareer --- ## AI Performance Testing Prompts: Find Bottlenecks URL: https://dontreplace.me/articles/ai-performance-testing-prompts-find-bottlenecks-before-users-do AI promptsperformance testingload testingquality assurancesoftware testing AI Performance Testing Prompts: Find Bottlenecks 10 AI performance testing prompts for workload models, load, stress, spike, soak, telemetry, bottleneck analysis, and evidence-based release decisions. August 17, 2026 Updated August 17, 2026 A performance test is not “send lots of requests and watch the dashboard become festive.” It is a controlled experiment with a representative workload, measurable thresholds, useful telemetry, and humans who know when to stop. AI performance testing prompts can make that experiment easier to plan. A model can organize requirements, draft workload profiles, suggest scenarios, expose missing assumptions, and summarize sanitized results. It cannot generate trustworthy evidence by describing a test it never ran. It also cannot predict production capacity from a toy environment, diagnose every bottleneck from one graph, or approve a release. AI can draft the experiment. Humans still build the environment, run the tools, inspect evidence, diagnose the system, control risk, and own the decision. Start with an AI test plan if the overall strategy is fuzzy. Use AI operational readiness prompts when performance is one part of a larger launch decision. This guide focuses on the part where traffic, latency, throughput, errors, resources, and dependencies meet reality. What performance testing actually measures Performance is not one number. A system can have a respectable average response time while a painful minority of users wait five seconds. It can handle normal traffic and collapse during a sharp spike. It can look healthy for twenty minutes and leak memory during an eight-hour soak. It can make the application fast by quietly overwhelming a database or third-party service. A useful performance plan connects business behavior to technical evidence: Latency: response time at meaningful percentiles such as p50, p95, and p99. Throughput: completed requests, jobs, transactions, or user journeys per unit of time. Concurrency or arrival rate: how work enters the system and overlaps. Errors: failures, timeouts, retries, rejected work, and incorrect responses. Saturation: CPU, memory, pools, queues, locks, storage, network, and dependency limits. Stability: whether behavior degrades over time or recovers after pressure. User outcome: whether critical journeys finish correctly within the agreed objective. Cost: what the tested capacity requires in infrastructure and paid dependencies. Different tests answer different questions. A baseline records current behavior under a controlled workload. A load test checks expected demand. A stress test pushes beyond it to find degradation and recovery behavior. A spike test applies sudden change. A soak test looks for slow failure over time. Scalability and capacity tests examine how resources and throughput change together. Those names are not magic. Define the question, workload, environment, metric, threshold, duration, and evidence for each one. Otherwise the team gets a colorful report that says “2,000 users” without explaining what any user did. The reusable performance testing prompt formula Use this base prompt, then add one specialized template below: “Act as a performance test-planning assistant. I am evaluating [named system, build, and user journeys] in [approved environment]. Use only the supplied requirements, architecture, traffic evidence, environment details, dependency limits, telemetry, and sanitized results. Produce [artifact] with assumptions, workload phases, measurable thresholds, setup, data needs, observability, stop conditions, evidence fields, and open questions. Separate planned work, observed facts, hypotheses, and decisions. Do not invent execution, benchmark results, production capacity, root causes, costs, or release approval.” The separation is crucial. A workload model is a forecast. A test is an experiment. A metric is an observation. A bottleneck is a diagnosis supported by multiple signals. A release recommendation is a human decision made in business context. Autocomplete enjoys blending these into one confident paragraph. Do not let it. Never paste credentials, API keys, customer PII, production traces, proprietary traffic details, private source code, network diagrams, logs, regulated information, or confidential client material into an unapproved AI tool. Use synthetic data, redacted telemetry, approved environments, placeholder secrets, and your organization’s security, privacy, vendor, and retention rules. What to collect before prompting “Make it fast” is not a requirement. Give the model a small packet of verified, sanitized context. InputWhy it mattersHuman check Critical user journeysDefines what success meansProduct owner confirms priorities Service objectives and thresholdsCreates pass/fail criteriaEngineering and product approve Current traffic evidenceGrounds the workload modelAnalytics owner validates source Growth and event assumptionsDefines expected peaksBusiness owner confirms forecast Environment topologyReveals production differencesPlatform team verifies configuration Test data and account planPrevents unrealistic shortcutsData owner confirms representativeness Dependency quotas and limitsAvoids collateral damageService owners approve ceilings Telemetry and trace coverageEnables diagnosisSRE verifies dashboards and alerts Stop and recovery conditionsControls test riskIncident owner confirms procedure Cost constraintsKeeps the plan practicalFinance or platform owner validates Keep unknowns visible. If nobody knows the expected checkout rate during a campaign, the output should ask for that number or propose clearly labeled scenarios. It should not quietly invent a forecast and call the resulting graph scientific. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI performance testing prompts Replace bracketed text with verified, sanitized information. These prompts produce planning and analysis artifacts. Humans must review them and operate real testing tools. 1. Turn vague expectations into measurable requirements “Using these critical journeys, business events, current service objectives, user expectations, dependency limits, and failure policies: [paste], draft measurable performance requirements. For each journey, propose latency percentiles, throughput or arrival rate, error limits, availability expectations during load, data volume, duration, and recovery criteria. Trace every existing requirement to its source and label proposed thresholds for human approval.” This prompt is useful when the brief says “the site should handle launch traffic.” Ask what launch traffic means: arrivals per second, active sessions, transactions, geographic mix, payload size, account state, and duration. Include background jobs and downstream effects. Do not accept averages alone. Averages can hide a miserable tail. Require percentiles, error classifications, and correctness checks. A fast response containing stale or incomplete data is not a performance win. 2. Build a representative workload model “Using these sanitized analytics, transaction counts, journey mix, concurrency evidence, peak windows, geographic distribution, device or client mix, payload sizes, data volumes, seasonality, and growth assumptions: [paste], create a workload model. Show normal, peak, and plausible high-growth scenarios; arrival rates; journey weights; pacing; session behavior; data needs; assumptions; confidence; and missing evidence. Do not convert page views directly into concurrent users without an explicit model.” A workload model should explain behavior, not only volume. Ten thousand users reading cached pages differ from ten thousand users searching, uploading files, generating reports, and checking out simultaneously. Validate the model with product, analytics, engineering, and operations. Watch for coordinated events: email campaigns, scheduled jobs, billing runs, imports, and mobile clients reconnecting after an outage. These produce traffic shapes that daily averages erase. 3. Select critical journeys and checks “Given these user journeys, API calls, background processes, business priorities, historical incidents, and workload model: [paste], create a performance coverage matrix. Include journey steps, transaction weights, preconditions, data variation, correctness assertions, expected latency and error thresholds, dependencies touched, telemetry needed, and excluded flows. Rank by user impact and uncertainty.” Test complete outcomes. A login endpoint may be quick while session creation, profile loading, feature flags, and post-login data calls make the experience slow. A purchase is not successful merely because the API returns 200; the order, payment state, inventory, confirmation, and asynchronous jobs must remain correct. Use AI test case prompts to sharpen assertions, but avoid turning a performance test into every functional test ever written. Cover critical correctness without burying the experiment in noise. 4. Plan a repeatable baseline test “Using this environment, build, dataset, journey mix, test tool, warm-up behavior, telemetry, and service objectives: [paste], draft a repeatable baseline test. Specify setup, cache state, warm-up, low controlled workload, duration, repetitions, metric collection, environment fingerprint, evidence naming, variability checks, cleanup, and comparison method.” Baselines make later comparisons meaningful. Record application version, infrastructure size, configuration, feature flags, database state, dependency versions, test tool version, region, and time. If those change, say so. Run enough repetitions to see normal variance. One unusually good run is not a baseline; it is a screenshot with ambition. Check whether caches, autoscaling, connection pools, and runtime compilation behave differently between cold and warm states. 5. Design an expected-load test “Using this approved workload model, service objectives, environment capacity, journey mix, test accounts, data plan, dependency quotas, and telemetry: [paste], draft an expected-load test with warm-up, ramp, steady-state, and recovery phases. Include arrival rates, pacing, duration, thresholds, correctness checks, monitoring, stop conditions, evidence, and production-equivalence caveats.” The steady phase must last long enough for queues, autoscaling, pools, and background work to settle. A two-minute victory lap may miss the actual behavior. Coordinate with dependency owners. Email, payment, identity, analytics, and vendor sandboxes may have strict quotas or different performance from production. Stubbed dependencies can isolate the application, but they cannot prove end-to-end capacity. Label the evidence honestly. 6. Find degradation with a controlled stress test “Given this baseline, expected peak, infrastructure model, known limits, incident procedure, and approved maximum pressure: [paste], draft a stepped stress test. Increase load in controlled stages, hold each stage, capture latency percentiles, throughput, errors, saturation, queues, dependency behavior, and correctness. Define breakpoint indicators, stop conditions, rollback, recovery observation, and questions for diagnosis.” Stress testing is not permission to assault production. Use an approved environment and ceilings. Define what triggers an immediate stop: data corruption, unexpected production calls, uncontrolled cost, shared-service impact, severe error rates, monitoring alarms, or loss of observability. The useful result is not merely “it broke at X users.” Record the actual arrival rate and journey mix, first violated objective, saturation signals, queued work, failure mode, and recovery. Capacity depends on behavior and environment, not a theatrical virtual-user count. 7. Plan spike and recovery testing “Using these event scenarios, normal traffic, peak arrival estimates, autoscaling behavior, queue design, cache behavior, dependency limits, and recovery objectives: [paste], create spike tests for sudden increases and decreases. Specify starting state, rise time, peak, hold time, drop pattern, correctness checks, telemetry, alert expectations, backpressure behavior, stop conditions, and recovery criteria.” Spikes reveal whether scaling reacts quickly enough and whether work is rejected, queued, retried, or duplicated. Watch retry storms: a struggling system may receive more traffic because clients retry aggressively. Test the way traffic arrives in reality. A product announcement, push notification, ticket sale, scheduled integration, and regional reconnect have different shapes. Also observe the drop. Systems can remain overloaded after incoming traffic falls because queues and retries keep work alive. 8. Design a soak test for slow failures “Using this realistic steady workload, expected long-running behavior, data growth, scheduled jobs, resource limits, deployment topology, and observability: [paste], draft a soak test. Define duration, workload variation, data rotation, account reuse, checkpoints, memory and connection trends, queue depth, storage growth, error drift, dependency health, log volume, stop conditions, and post-test recovery checks.” Soak tests look for leaks, fragmentation, pool exhaustion, accumulating queues, runaway logs, cache churn, expiring credentials, scheduled-job collisions, and degradation that short tests miss. Do not repeat one identical request for eight hours and call it representative. Rotate data and journeys safely. Monitor trends rather than only final values. A resource that rises slowly but never returns after work completes deserves investigation even if the test finishes under the threshold. 9. Analyze sanitized telemetry without inventing a root cause “Using these sanitized test phases, timestamps, latency percentiles, throughput, error categories, resource metrics, queue depths, traces, profiles, dependency timings, changes, and environment notes: [paste], create an evidence matrix. Correlate signals by time, identify supported observations, list multiple root-cause hypotheses, name evidence for and against each, and propose the smallest safe follow-up experiments. Do not claim causation from correlation.” A CPU spike beside slow requests does not automatically mean CPU caused the slowdown. Both may be symptoms of retries, serialization, garbage collection, lock contention, or another dependency. Ask for competing hypotheses. Humans inspect raw evidence in approved systems. Keep trace identifiers and sensitive payloads out of the prompt. Use AI bug report prompts to convert a verified performance defect into a reproducible ticket. 10. Turn results into an action and release brief “Using only these verified test results, requirements, environment caveats, open defects, mitigations, capacity estimates, cost data, and owner decisions: [paste], draft a performance test summary. Include scope, build, environment, workload, passed and failed thresholds, evidence links, limitations, unresolved risks, remediation options, owners, retest plan, monitoring changes, rollback triggers, and a decision table. Do not invent results or make the final release decision.” Keep facts separate from recommendations. “p95 exceeded the approved threshold during the peak phase” is an observation. “Delay launch” is a decision that also depends on user impact, mitigation, event timing, rollback, and business risk. If the tested environment differs from production, state exactly how. Avoid multiplying staging throughput by a hardware ratio and declaring production capacity. Scaling can be nonlinear because databases, locks, caches, queues, and vendor limits behave differently. A workflow that produces evidence instead of theater Define the business question. Name the journey, demand scenario, objective, and decision the test informs. Sanitize the context. Remove secrets, private traces, customer data, and proprietary details before prompting. Build the workload model. Ground it in analytics and explicit forecast assumptions. Verify the environment. Document topology, data, dependencies, configuration, and production differences. Prepare observability. Confirm metrics, logs, traces, profiles, clocks, dashboards, and alert behavior before load begins. Review the generated plan. Engineering, QA, SRE, product, security, privacy, and dependency owners correct it. Run a baseline. Establish repeatability before increasing pressure. Execute controlled phases. Follow approved ceilings, stop conditions, and incident procedures. Correlate evidence. Compare user outcomes, application metrics, infrastructure, queues, and dependencies. Fix and retest. Re-run the same controlled experiment and check for regressions elsewhere. Make the decision. Humans weigh evidence, limitations, cost, mitigations, monitoring, rollback, and business context. For quick post-deployment checks, use AI smoke test prompts. For ongoing signals after release, adapt these AI post-launch monitoring prompts. Neither replaces a representative pre-release experiment. Common ways performance testing goes wrong Reporting virtual users without behavior “5,000 users” means little without arrival rate, pacing, journey mix, session length, payloads, and test duration. Report the workload that the system actually received. Testing an environment nobody can compare to production A smaller environment can still reveal defects, but capacity conclusions need caveats. Record topology, scaling, data volume, caches, network, and dependency differences. Watching averages and missing the tail Averages hide painful outliers. Track percentiles, error classes, timeouts, retries, and critical journey completion. Slice results by endpoint, region, client, and dependency when evidence allows. Generating traffic before observability works If metrics, logs, traces, and clocks are broken, the team may know the system slowed down but not why. Validate telemetry during the baseline. Treating correlation as diagnosis One dashboard does not prove a root cause. Form competing hypotheses and run small follow-up experiments. Change one meaningful variable where possible. Ignoring correctness under load A fast system that drops jobs, duplicates orders, serves stale permissions, or loses data failed. Add bounded assertions and inspect downstream state. Running against production by surprise Performance traffic can create outages, bills, alerts, vendor penalties, and bad data. Use approved environments. If a production test is genuinely required, it needs explicit authorization, communication, safeguards, observability, stop conditions, and incident ownership. Pasting sensitive telemetry into public AI tools Traces and logs can contain identities, tokens, queries, business volumes, architecture, and customer data. Redact aggressively and use approved tools and retention policies. This boundary is easier to maintain when you understand what AI can and cannot do. Pattern generation is fast. Execution, evidence, diagnosis, and accountability still belong to people. Frequently asked questions Can AI run performance tests automatically? AI can help control approved tools when engineers deliberately integrate it, but a chat response has run nothing. Connected automation still needs a reviewed workload, representative environment, safe limits, telemetry, evidence storage, and human interpretation. What is the best AI prompt for load testing? The best prompt includes the named system and build, critical journeys, verified traffic model, environment, service objectives, data, dependency limits, telemetry, phases, stop conditions, and desired artifact. Require assumptions and unknowns to be explicit, and forbid invented results. How many concurrent users should I test? There is no universal number. Derive workload from arrivals, session behavior, journey duration, traffic evidence, forecasts, and safety margins. Concurrency is an outcome of those factors, not a number copied from a competitor’s blog post. Can staging results predict production capacity? They can provide evidence, especially when the environments are closely matched, but they do not guarantee production capacity. Document differences in compute, scaling, data, network, caches, dependencies, configuration, and traffic. Validate important conclusions with engineering judgment and carefully controlled evidence. Should performance tests use average response time? Use averages as one signal, not the verdict. Include percentiles such as p50, p95, and p99, plus throughput, errors, timeouts, retries, saturation, and journey correctness. The tail often describes the users who are actually angry. What is the difference between load, stress, spike, and soak testing? Load testing checks expected demand. Stress testing increases pressure to study limits and degradation. Spike testing examines sudden traffic change and recovery. Soak testing runs a representative workload long enough to reveal slow leaks and accumulation. Define each test by its question and phases rather than its label. Is it safe to paste production metrics into an AI tool? Not automatically. Metrics, logs, and traces may expose PII, tokens, customer behavior, architecture, incidents, or confidential business volume. Follow approved vendor, privacy, security, retention, and data-classification rules. Prefer sanitized summaries and keep raw evidence in controlled systems. Can AI decide whether performance is good enough to release? No. It can organize verified results against approved thresholds and draft a decision table. Humans must judge environment limitations, user impact, unresolved defects, cost, mitigations, monitoring, rollback, event timing, and accountability. The useful boundary AI is good at turning scattered requirements into a structured experiment. It is good at generating variations, asking what is missing, and formatting evidence for review. That saves time. It is bad at knowing whether your traffic model is true, whether the test environment represents production, whether a graph proves causation, or whether the remaining risk is acceptable. Those are precisely the parts where domain knowledge and judgment matter. Use the prompts to accelerate preparation and analysis. Keep real execution, sensitive evidence, diagnosis, remediation, and release authority with accountable people. That is the broader argument in Dmitry Kargaev’s Don’t Replace Me: use the machine for speed, then apply the human taste and responsibility it cannot supply. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance AI API Testing Prompts: Catch Broken Integrations AI API testing prompts for contracts, edge cases, authorization, retries, error responses, and human-verified integration evidence. AI promptsAPI testingquality assurance AI Compatibility Testing Prompts: Catch Breakage 10 AI compatibility testing prompts for browser, device, OS, responsive, locale, network, defect triage, and evidence-based release decisions. AI promptscompatibility testingcross-browser testing --- ## AI Personal Brand Prompts That Actually Work URL: https://dontreplace.me/articles/ai-personal-brand-prompts-sound-like-yourself-faster ai personal brand promptsLinkedIn promptspersonal brandingai at workcontent strategy AI Personal Brand Prompts That Actually Work 10 copy-paste AI personal brand prompts for LinkedIn and beyond. Templates for positioning, storytelling, hooks, and repurposing that don't sound like AI wrote them. June 10, 2026 Updated June 10, 2026 Most personal brand content sounds like it was written by a LinkedIn robot having an epiphany in an airport lounge. "I failed. Then I succeeded. Here's what I learned. 🙏" You've seen it. You've maybe written it. We all have. AI personal brand prompts can fix the speed problem. They can't fix the identity problem. That's on you. But if you already know what you think and you just need help getting it out of your head and onto a page without spending 45 minutes staring at a blank draft, this is the article for you. Ten copy-paste prompts, a reusable formula, a few hard warnings about what happens when you let AI do the thinking instead of the typing, and a quality check you have to run yourself every single time. The one rule that makes AI personal brand prompts actually work AI is fast. That's it. That's the whole value proposition. Rule #5 in Don't Replace Me puts it plainly: it's not smart, it's fast. Speed is not identity, credibility, or judgment. You still have to bring all of that. The prompts below work because they treat AI as an editorial assistant, not a ghostwriter. You give it your raw material. It gives you structure, hooks, and variants. You pick the one that sounds like you. That last step, the selection, is where your taste lives. And taste, per Rule #7, is the moat. Choosing what sounds true and what should be left unsaid is not something a language model can do for you. So: garbage in, garbage out. Vague prompts produce generic thought-leader slop. Specific prompts, with your actual words, opinions, and experiences, produce usable drafts. The reusable formula for every AI personal brand prompt Before the templates, here's the pattern behind all of them: [Role/Context] + [Raw material you provide] + [Specific output format] + [Constraints and voice notes] An example of a weak prompt: "Write me a LinkedIn post about leadership." An example of a strong prompt: "I'm a project manager with 8 years in construction. I want to write a LinkedIn post about a specific mistake I made last year: I approved a supplier without checking references because we were rushed. It cost us three weeks. The lesson is that speed under pressure often costs more time than it saves. Write three versions of this post in plain, direct language. No motivational poster endings. Under 200 words each." That's the formula. Now the templates. The 10 AI personal brand prompts 1. Clarify your positioning Use this when you're not sure how to describe what you actually do and who it's for. "I'm going to describe my work and background. Based on what I say, write me three different one-sentence positioning statements for my LinkedIn headline and bio. Focus on who I help, what outcome I create, and what makes my approach different. Don't use buzzwords. Here's my background: [paste your current bio or a few sentences about your role, experience, and who you work with]." 2. Extract a story from your experience Use this when you know something happened that's worth sharing but you can't figure out how to frame it. "Here's a situation I experienced at work: [describe it in plain language, as if telling a friend]. Turn this into a LinkedIn post that opens with the moment of tension, not background context. Keep the tone direct. Don't editorialize or add a moral at the end. Let the story land on its own. Give me two versions: one under 150 words, one under 300 words." 3. Turn a strong opinion into a post Use this when you have a take but you're not sure how to make it interesting rather than preachy. "I have the following opinion about [topic]: [state your actual view]. I want to write a post that makes this point without lecturing. Start with a specific observation or example that most people in my field will recognize. Challenge one common assumption. Don't hedge. Write two versions: one that leads with the example, one that leads with the contrarian statement." 4. Rewrite a draft in your actual voice Use this after you've written a rough draft and it sounds too formal, too fluffy, or too much like someone else. "Here's a draft I wrote: [paste your draft]. Rewrite it to match these voice characteristics: [list 3-5 things about how you actually talk and write, e.g., 'short sentences, occasional dry humor, no inspirational endings, direct without being harsh, British spelling']. Don't add anything that wasn't in the original. Don't remove the specific details. Just adjust the tone and sentence rhythm." 5. Create LinkedIn hooks without cringe Use this when you know what the post is about but the opening line is flat. "Here's the main point of a LinkedIn post I'm writing: [one sentence summary]. Write 10 opening lines for this post. Avoid: questions that answer themselves, fake vulnerability reveals, anything that starts with 'I used to think' or 'Unpopular opinion:', dramatic pauses using ellipses, and anything that sounds like a TED talk intro. Keep them direct and specific." 6. Repurpose one idea into multiple formats Use this when you've written or said something good and want to get more mileage from it. "Here's an idea I've already written about or discussed: [paste the original content or describe it in detail]. Repurpose this into: one short LinkedIn post (under 150 words), one Twitter/X thread (5 tweets), one paragraph for a newsletter, and one bullet for a speaking bio or case study. Keep the core point identical. Don't exaggerate the results or add claims that weren't in the original." 7. Build a content pillar map Use this when you post inconsistently because you don't know what you're "supposed" to talk about. "I work in [field/role]. My audience is [who they are]. I have opinions and experience in the following areas: [list 5-7 topics you genuinely know about]. Group these into three to five content pillars that fit together. For each pillar, suggest four to six specific post ideas based only on the topics I listed. Don't add topics I didn't mention." 8. Pressure-test a post for credibility and specificity Use this before you publish anything you're not sure about. "Here's a post I'm about to publish: [paste the post]. Play devil's advocate. Tell me: What claims are too vague to be credible? What sounds like it could have been written by anyone? What sounds like I'm exaggerating? What words or phrases will make my audience roll their eyes? Be blunt. I'd rather know now." This is one of the most underused prompt types in personal branding. For more ways to use AI as a critic rather than a cheerleader, the no-BS starter guide to using AI at work has a useful frame for this. 9. Create a weekly posting plan Use this when you want a system instead of scrambling every week. "I want to post on LinkedIn three times a week for the next four weeks. My content pillars are: [list them]. My audience is [describe them]. Create a simple posting calendar with a topic and one-sentence angle for each post. Don't write the posts. Just give me the plan. Make sure the topics rotate so I'm not saying the same thing three times a week." 10. Edit for clarity without sanding off personality Use this when a piece of writing is almost there but something is off. "Here's something I've written: [paste it]. Edit it for clarity and tightness. Cut anything that repeats a point already made. Shorten sentences that run too long. But preserve: my specific examples, my opinions, and any place where I'm being direct or funny. Don't make it more formal. Don't make it safer. Just make it cleaner." This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Why most people use these prompts wrong Here's the pattern. Someone reads a list like this, tries one prompt, gets a decent output, and then stops bringing their own material. They start asking AI to invent the experience instead of structure the real one. "Write me a story about a leadership mistake." Instead of: "Here's an actual mistake I made. Help me tell it better." That's where the content starts dying. Not dramatically. Gradually. The posts become technically competent and completely hollow. The engagement drops. The comments get shallower. Eventually the person concludes that AI personal branding "doesn't work" when what actually happened is they outsourced the one part that was never available for outsourcing. The prompts in this list are specifically designed to prevent that. Every one of them requires you to paste in something real: your background, your story, your draft, your opinion, your actual experience. If you find yourself running these prompts with placeholder brackets still in them, stop and go find the real material first. The other common mistake is publishing the first output. AI gives you a draft. That's the starting point, not the finish line. Read it out loud. If you wouldn't say it that way in a meeting with someone you respect, it needs another pass. Editing AI output is faster than writing from scratch, but it still requires you to actually do it. What to never put in an AI personal brand prompt Before you start pasting things into ChatGPT or Claude, here's what doesn't go in. Confidential client work. Private employer strategy. Employee performance issues. Customer data. Unreleased products or revenue numbers. Legal matters. Anything that would make your legal or HR team's eyes go wide. Most corporate AI tools have data handling policies. Most personal AI tools don't use your input for training (check the settings to be sure), but that's different from the prompt never existing on a server somewhere. The rule is simple: if you'd hesitate to post it on a public forum, don't paste it into an AI tool you haven't cleared with your employer. The other category is fabrication. Do not let AI invent achievements you haven't earned, client names you haven't worked with, revenue numbers you haven't hit, testimonials nobody said, personal stories that didn't happen, or fake vulnerability designed to perform relatability. This is the part where AI personal brand content gets genuinely dangerous. Not dangerous in a sci-fi way. Dangerous in a "your audience will eventually notice" way. And they do notice. Not all at once, but in the way that trust erodes: quietly, then suddenly. If you want to understand the full range of what AI can and actually can't do here, this five-minute explainer is worth a read before you ship anything important. The quality check you have to do yourself AI can give you structure and speed. It can not give you judgment. Before you post anything generated with AI help, run through this yourself. Is it true? Every specific detail: numbers, outcomes, client situations, quotes. Is it actually yours? The opinion, the story, the point of view. Does it sound like you? Not like a LinkedIn post. Like you, specifically. Would you be comfortable if your most skeptical professional contact read this? Is it actually worth publishing, or are you just filling a calendar? That last question matters more than people admit. Posting volume is not the same as building a reputation. A personal brand is what people think of you when you're not in the room. Fifty generic AI-assisted posts builds nothing. Five posts that are actually specific, true, and useful build something real. There's also the question of what you're optimizing for. If the goal is engagement metrics, AI can help you chase them. If the goal is to become the person people call when they have a specific problem in your field, that requires something else entirely. It requires being associated with real, specific, useful expertise over time. No prompt can manufacture that. The skills that matter for the long game aren't the ones a content calendar can produce. They're built through the actual work, and the personal brand content should reflect that work, not replace it. If you find yourself using these prompts mostly to keep up a posting schedule rather than to say something real, that's worth noticing. For a broader framework around prompting for work, not just personal branding, the 12 foundational prompt templates are a good starting point. Frequently asked questions Can AI write my personal brand content for me? AI can draft and structure content fast, but it can't supply your opinions, experiences, or credibility. The content needs to come from you. Treat AI as an editorial assistant: it helps with structure, hooks, rewrites, and variants. You supply the raw material and make the final call on what's worth publishing. What's the best way to make AI content sound like me and not like a LinkedIn robot? The key is specific input. Vague prompts produce generic content. Paste your own rough drafts, describe your actual tone, list specific things you want to avoid, and give AI real experiences to work with. Then select and edit the output rather than publishing it as-is. Is it okay to use AI for LinkedIn posts? Yes, as long as the ideas, opinions, and facts are genuinely yours. The problem isn't using AI to help write, it's letting AI invent achievements, fake stories, or implied results you can't back up. If you're an author using AI, you should be able to stand behind every specific claim in the post. What should I never put into an AI tool when writing personal brand content? Confidential client information, unreleased company strategy, employee or customer data, private revenue numbers, or anything under legal review. Also: don't ask AI to invent stories, testimonials, or achievements that aren't real. Both categories create risk, one legal and one reputational. How many posts per week should I aim for with AI assistance? Quality over volume. Three posts per week that are specific, true, and actually useful to your audience will outperform five AI-padded posts designed to fill a schedule. A content calendar helps with consistency, but it's not a substitute for having something worth saying. How is AI personal branding different from just using a ghostwriter? In practice, less different than people think. Both speed up production. The ethical standard is the same: the ideas, experiences, and claims need to be genuinely yours. The practical difference is cost and turnaround. AI is faster and free. The judgment layer stays human either way. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Automation Prompts: Automate the Boring Stuff 10 copy-paste AI automation prompts for mapping workflows, writing SOPs, adding approval gates, and QA testing. Practical templates with safety rules included. ai automation promptsworkflow automationSOP templates AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts --- ## AI Pilot Program Prompts: 10 Templates That Work URL: https://dontreplace.me/articles/ai-pilot-program-prompts-10-templates-that-work ai pilot programai tools at workpilot program promptsai implementationai change management AI Pilot Program Prompts: 10 Templates That Work 10 copy-paste AI pilot program prompts for scope docs, success metrics, feedback scorecards, and go/no-go recommendations. Built for real teams. July 20, 2026 Updated July 20, 2026 Someone brings a new AI tool to a meeting. There's a demo. It looks slick. People nod. Someone says "pilot program" and suddenly you're the person figuring out what that actually means. AI pilot program prompts can help you build the structure, the scorecard, the scope doc, and the rollout brief, without starting from a blank page. But prompts are planning scaffolding, not a substitute for judgment. The tool doesn't know your vendor contracts, your privacy policy, your users' actual frustrations, or whether your "success metrics" are real or made up to justify a decision someone already made. This article gives you 10 copy-paste prompts for every stage of a pilot: from defining your hypothesis to writing a go/no-go recommendation. Each one comes with a reminder about what to leave out of the prompt and where human sign-off isn't optional. The reusable prompt formula for AI pilot programs Every good pilot prompt follows the same logic: [Role] + [Tool or context] + [Specific task] + [Constraints and output format] That's it. The more specific your inputs, the less the AI has to invent. And AI will invent. It will give you confident-sounding success metrics for a problem it knows nothing about, because you didn't tell it the real numbers. Here's what to include every time: What the tool is supposed to do What problem it's solving (specific, not vague) Who the test users are and roughly how many What data exists already (baseline) What format you want back What to leave out: Customer names, emails, or any personal identifiable information (PII) Access tokens, credentials, API keys Private HR issues, legal disputes, or active litigation Confidential vendor pricing or unreleased contracts Financial records, medical data, or regulated information Internal security vulnerabilities or compliance gaps Private client conversations None of that belongs in an unapproved AI tool. If your company hasn't cleared the tool through procurement and security, treat it like a public whiteboard. Prompts 1-3: AI pilot program prompts for defining the problem and scope Prompt 1: Define the pilot hypothesis I'm planning a pilot for [tool name], which is supposed to help our [team/role] with [specific task]. Write a one-paragraph pilot hypothesis in this format: "We believe that [user/team] will be able to [outcome] if we give them [tool], because [reason]. We'll know this worked if [measurable result] improves by [X] within [time period]." We currently [describe current state or baseline if you have it]. The pilot will run for [duration] with [number] users. Why it works: Forces you to state a measurable outcome before you fall in love with the tool. If you can't fill in the blanks, the pilot isn't ready to start. This is also the document you come back to at the end when someone asks "did it work?" without having agreed what "work" means upfront. Prompt 2: Narrow the scope We're piloting [tool] with the goal of [specific outcome]. I want to keep this pilot small enough to learn from without creating chaos. Help me define a scope boundary: which tasks should be in scope, which should be out of scope, and what the minimum viable test looks like. Our team size is [X], the time box is [Y weeks], and the main risk we're worried about is [Z]. Why it works: Scope creep kills pilots. Something that starts as "test the AI writing tool for emails" becomes "the whole marketing workflow" inside three weeks. Write the boundary before the pilot starts. If you can't describe the scope in two sentences, the scope is too big. Prompt 3: Choose test users I need to select [X] pilot users from a team of [Y] people. The tool is [name] and the task is [description]. Help me write a selection criteria framework. I want users who represent [key variables: skill level, workflow type, tech comfort, etc.] and can give us useful feedback. Also suggest what I should tell them when I invite them into the pilot, in plain language. Don't paste actual employee names, job titles with identifying details, or HR performance records. Use role types and skill levels instead. A good pilot group has range. You want at least one skeptic, not just the people who volunteered because they're already excited about AI. Enthusiasts will over-report satisfaction. Skeptics will surface the real friction. Prompts 4-6: AI pilot program prompts for metrics, plans, and boundaries Prompt 4: Write success metrics Our pilot hypothesis is: [paste from Prompt 1]. Write 3-5 measurable success metrics for this pilot. For each metric, include: what we're measuring, how we'd collect that data, what our current baseline is [insert your real number here or write "unknown - we need to establish this"], and what threshold would count as success. Don't invent baselines. Flag any metric where we'd need to collect data before the pilot starts. This is where most pilots lie to themselves. If you don't have a baseline, say so. A pilot with no before-state proves nothing. "Users reported saving time" is not a metric. "Average time to complete X task dropped from 45 minutes to 28 minutes, measured across 10 users over four weeks" is a metric. Prompt 5: Create a test plan I'm running a [X-week] pilot of [tool] with [Y] users from [team/department]. The goal is [outcome from hypothesis]. Write a week-by-week test plan that includes: what we're testing each week, how we're collecting feedback, what decisions happen at the midpoint check-in, and what the final week looks like. Include a note about who owns each stage and when we need human approval to proceed. The midpoint check-in matters more than most people think. That's where you catch "nobody is actually using this" before you've spent eight weeks on a failed pilot. Build in a real stop/continue decision at week three or four, not just a status update. If you need help with the risk angle too, these AI risk assessment prompts cover what can go wrong before it does. Prompt 6: Set data and privacy boundaries We're piloting [tool] for [use case]. Help me write a one-page data boundaries document for pilot participants. Include: what types of data they can and cannot input into the tool, how to handle situations where they're unsure, who to ask if they hit an edge case, and what to do if they accidentally paste something they shouldn't have. Our company [has/has not] completed a security and privacy review of this tool. [Add relevant policy context here if it's non-confidential.] Note: This prompt gives you a draft framework only. Your legal, security, and privacy teams need to review the actual data boundaries before the pilot launches. AI can draft the structure. It can't replace compliance review. People paste things they shouldn't. It happens. Having a written policy that says "here's what you do when that happens" is not paranoia. It's the difference between an incident and a crisis. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Prompts 7-8: AI pilot program prompts for feedback and results Prompt 7: Build a feedback scorecard I need a feedback scorecard for [X] pilot users testing [tool] for [use case]. The pilot runs [Y weeks]. Create a scorecard they fill out [weekly/at the end] with: a rating scale for ease of use, time saved, output quality, and confidence in results. Add 2-3 open questions that surface friction points we wouldn't think to ask about. Keep it short enough that people actually complete it. Short feedback forms get filled out. Long ones get ignored at 4:58pm on a Friday. Be realistic about what people will actually do. Five questions with one optional open field beats a 20-question survey that produces a 15% response rate and tells you nothing. The open questions are where the real insight lives. "What would have to be true for you to use this every day?" will tell you more than any rating scale. Prompt 8: Summarize pilot results Here are the pilot results from [X] users over [Y weeks]: [Paste aggregated, anonymized feedback here. No PII, no individual names linked to responses, no confidential data.] Summarize: what the data shows, where adoption was strongest, where friction was highest, what the main user complaints were, and what's still unknown. Flag anything that looks like a measurement gap rather than a real finding. The AI summary is a starting point. You still need to read the actual responses. A language model smoothing over "people hated it" is a real risk here. Don't let a tidy summary replace the evidence. Watch for what gets softened. If three out of eight users said the tool's outputs were unreliable and they stopped trusting it by week two, that's a finding. Not a "mixed response around output consistency." Read the raw feedback yourself before you sign off on any summary. The full framework for running this kind of structured audit is in Don't Replace Me, Dee's field guide for staying useful while everyone else is confusing a polished pilot deck with an actual decision. Prompts 9-10: AI pilot program prompts for go/no-go and next steps Prompt 9: Write the go/no-go recommendation Based on this pilot summary: [paste your summary from Prompt 8] Write a go/no-go recommendation document. Include: the original hypothesis, whether we met our success criteria, the key evidence for and against proceeding, outstanding risks, what a full rollout would require (training, procurement, security sign-off, support capacity), and a recommended decision with the conditions attached. Format it for a leadership review. Do not fill in gaps with assumptions. If something is unknown, say it's unknown. The final decision is a human decision. The recommendation doc helps structure the conversation. It doesn't make the call. One thing worth adding manually: cost. AI won't know your vendor pricing, what the full rollout license actually costs, or what IT support load looks like at scale. Those numbers need to come from real people before leadership reviews the doc. Prompt 10: Plan the next rollout step We've decided to [proceed/not proceed/extend the pilot] with [tool] based on [brief summary of outcome]. Write a next-step action plan with: who owns what, the timeline for [full rollout / extended pilot / wind-down], what procurement and security steps still need to happen, the training plan for new users, how we'll handle the transition for current pilot users, and how we'll measure success at the next stage. For help building that training plan, these AI training prompts will give you a head start on onboarding users who weren't in the pilot group. And when you're ready to communicate the decision to stakeholders, these stakeholder update prompts make that easier too. What AI can't do in a pilot program AI can generate structure, suggest questions, draft frameworks, and summarize feedback. It can't tell you whether the tool is actually safe to deploy in your environment, whether your vendor's data processing agreement is solid, or whether the "efficiency gains" your pilot showed are real or just novelty effect. A polished pilot summary that glosses over a 40% abandonment rate in week three is worse than no summary at all. The AI will make it read smoothly. That's the problem. Before any tool moves from pilot to production, it needs: Procurement and legal review of vendor contracts Security and privacy sign-off on data handling A real training plan (not just a Slack message) Support load estimation so your IT team isn't blindsided A documented rollback path if things go wrong A named human being who is accountable for the decision If you want help thinking through the vendor side before the pilot even starts, these vendor evaluation prompts are worth running first. The most common AI pilot mistakes, and how to avoid them Most pilots fail for the same handful of reasons. None of them have anything to do with the tool. Measuring enthusiasm instead of outcomes. People like new things. In week one, engagement is high, feedback is positive, and everyone's exploring. By week four, novelty has worn off and you see the real usage patterns. If your only measurement window is the first two weeks, you measured the wrong thing. No baseline. You can't prove improvement if you didn't measure the starting point. This sounds obvious. It gets skipped constantly. Before the pilot starts, document how long the task currently takes, how often errors occur, what the current process is. Even rough numbers are better than none. Selecting only willing volunteers. Power users and early adopters will make any tool look better than it is. They'll work around friction, figure out workarounds, and adapt their workflows. The average user won't. If your pilot group is 100% enthusiasts, your results won't transfer to general rollout. Letting the vendor run the evaluation. The vendor wants a case study. Their definition of success may not match yours. They'll highlight the wins and minimize the friction. You need your own scorecard and your own data collection, independent of whatever the vendor sends you. Skipping the wind-down plan. When a pilot ends, what happens to the data that users created inside the tool? What happens to the pilot users' access? If you proceed, what's the transition path? If you don't, how do you communicate that without making people feel like their feedback was ignored? These questions should be answered before the pilot starts. Frequently asked questions What should I include in an AI pilot program prompt? Include the tool name, the specific problem it's solving, who the test users are, the time frame, what baseline data exists, and the output format you want. The more specific your inputs, the less the AI has to guess or invent. Never include PII, credentials, confidential financials, legal or HR issues, or regulated data. How long should an AI tool pilot program run? Most meaningful pilots need four to six weeks minimum. Shorter than that and you're mostly measuring novelty. Longer than eight weeks without a structured midpoint check-in and scope tends to drift. Define the time box before you start, and build in a checkpoint where the pilot can be stopped early if things go sideways. Can AI write my pilot program success metrics for me? It can draft a structure, but it can't fill in real baselines it doesn't have. If you paste in vague goals, you'll get confident-sounding metrics with no grounding in your actual situation. The prompt in this article specifically asks the AI to flag missing baselines rather than invent them. That flag is the useful part. What data should I never paste into a pilot planning AI tool? Customer PII, employee records, legal disputes, HR files, security vulnerabilities, medical or financial records, access credentials, confidential vendor pricing, regulated data, and private client conversations. If your company hasn't completed a security review of the tool, treat it like a public forum. Who is accountable for the go/no-go decision in an AI pilot? A named human being. The AI can help you write the recommendation document. It can't take responsibility for the outcome, flag regulatory concerns it doesn't know about, or judge whether the business impact is real. Final decisions need a real owner. What's the difference between a pilot summary and a real result? A pilot summary is a narrative. A real result is evidence. If the AI writes a smooth four-paragraph summary and it makes the pilot sound more successful than the raw feedback suggests, that's the AI doing what it does: making text flow. Read the actual feedback before you trust the summary. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Change Management Prompts: Get Buy-In Fast 10 copy-paste AI prompts for change management: announcements, stakeholder maps, resistance diagnosis, training, adoption metrics, and rollback plans. ai change managementchange management promptsai adoption AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management --- ## AI Policy Prompts: Write Rules People Follow URL: https://dontreplace.me/articles/ai-policy-prompts-write-rules-people-follow ai policyworkplace aiai promptshr policyai governance AI Policy Prompts: Write Rules People Follow 10 copy-paste AI policy prompts for HR and ops leads. Map AI use, set data rules, write disclosure policies, and build a rollout memo people actually read. June 05, 2026 Updated June 05, 2026 Most company AI policies read like they were written by a committee that had never actually used AI. Two pages of "employees must use AI responsibly" followed by zero examples of what responsible means. Nobody reads them. Nobody follows them. And then someone pastes a client contract into ChatGPT and HR has to send an all-hands email. You can use AI to write a better policy. That's not ironic. It's practical. But the prompts matter, and the judgment still has to come from you. These AI policy prompts are designed for HR managers, operations leads, legal-adjacent people, and anyone who got voluntold to "handle the AI thing" at work. The goal is a policy people actually understand, approve tools that make sense, and rules that don't require a law degree to interpret. No fabricated statistics. No fake compliance claims. Just prompts that get you 80% of the way there, so you can spend your energy on the 20% that actually requires a human brain. Before you start: Don't paste confidential customer data, employee records, legal issues, unreleased strategy, proprietary datasets, or security incidents into any AI tool. Use anonymized, fictional examples. Verify final policy language with your legal, security, and HR teams. AI should draft. Humans should decide. Why AI policy prompts work better than starting from scratch Writing policy is the kind of work that looks easy and isn't. You stare at a blank document, type "Employees shall..." and immediately realize you have no idea what you're actually trying to say. AI is useful here not because it has better judgment than you, but because it can generate a usable first draft fast, and a draft is infinitely easier to edit than nothing. The catch: AI will produce confident, professional-sounding text regardless of whether it's correct. It will invent compliance requirements. It will add legal-ish phrases that sound authoritative and mean nothing. It will make tradeoffs without telling you it's making them. Rule #5 in Don't Replace Me covers this exactly: AI isn't smart. It's fast. Someone still has to make the real calls about risk, trust, and what your company actually does. So treat these prompts as a way to skip the blank-page problem. Not a way to skip the thinking. The reusable formula for every AI policy prompt Before the 10 prompts, here's the pattern that makes each one work: [Role] + [Context] + [Specific output] + [Constraints] Example: "You are an HR policy writer for a 200-person B2B software company. We're writing an internal AI use policy for our customer support team. Draft a one-paragraph rule about when employees must get manager approval before using AI on a customer-facing task. Keep it under 150 words. Write in plain English, no jargon." Role gives the AI a useful frame. Context stops it from generating generic boilerplate. Specific output gets you something usable immediately. Constraints prevent it from writing a dissertation. Every prompt below follows this formula. Adjust the company size, industry, and team to match your situation. AI policy prompts: the 10 templates Prompt 1: Map your current AI use Before you write rules, find out what's actually happening. Most companies are already using more AI tools than they think. You are an internal comms writer helping a [company size] [industry] company understand its current AI usage. Write a 10-question anonymous survey for employees that covers: which AI tools they're currently using for work, what tasks they use AI for, whether they've shared any work data with external AI tools, and whether they have concerns about AI at work. Keep questions specific and non-judgmental. Include one open-text question at the end. Run this survey before you write anything else. You need the data. If you skip this step and write policy based on assumptions, you'll write rules for problems you don't have and miss the ones you do. Prompt 2: Define allowed and banned data categories This is the most important policy section and the one most companies get vague. "Don't share sensitive data" is not a policy. "Don't paste anything from these five categories into any external AI tool" is. You are a data governance writer for a [company size] [industry] company. We need a clear internal policy section that defines what data employees MAY and MAY NOT share with external AI tools (like ChatGPT or Claude). Use a two-column table: "Allowed" and "Not Allowed." Examples of not allowed: customer names and contact info, financial records, employee personal data, unreleased product details, legal correspondence, credentials, and anything marked confidential. Examples of allowed: publicly available information, general questions, anonymized fictional scenarios, draft text that contains no proprietary details. Keep language plain. Your legal and security teams will want to review and expand this list. That's expected. The draft just needs to be specific enough that employees can tell the difference without calling HR. Prompt 3: Write your approved tools list A policy with no approved tools is a policy that nobody can act on. People will use whatever they found themselves, which is usually fine, until it isn't. You are an IT policy writer for a [company size] [industry] company. Draft a one-page "Approved AI Tools" section for our internal AI policy. Include: a list of currently approved tools (leave placeholders: [Tool 1], [Tool 2]), what each tool is approved for, which team manages the account, and how employees can request a new tool be reviewed for approval. Add a sentence making clear that unapproved tools should not be used for work tasks involving company or customer data. Blanks and placeholders are fine. You're building a template your actual team fills in. The structure matters more than the content at this stage. Prompt 4: Write your disclosure rules If someone on your team uses AI to write a client deliverable, does the client need to know? Does your manager? These questions don't have one right answer, but your policy needs to pick one. If you're wondering about the personal side of this question, should you tell your boss you use AI? covers the workplace politics in more detail. You are an HR policy writer for a [company size] [industry] company. Draft a policy section on AI disclosure. Cover three scenarios: (1) internal documents only seen by employees, (2) client-facing work products, (3) public communications like social media or press releases. For each, specify whether AI use must be disclosed, to whom, and how. Keep it under 200 words. Write plainly. Don't invent legal requirements. Prompt 5: Set human review requirements This is where a lot of policies fall apart. They say "AI output must be reviewed by a human" without saying what that review involves. Read it? Fact-check it? Have a manager sign off? That ambiguity is where errors and liability slip through. You are a quality and compliance writer for a [company size] [industry] company. Draft a policy section defining "human review" requirements for AI-generated content. Specify: what types of AI output require review before use, what a review must include (at minimum: accuracy check, tone check, data privacy check), who is qualified to review in different contexts, and what employees should do if they spot an error in AI output they've already sent. Keep it practical and under 250 words. Prompt 6: Write role-specific examples Generic policy language fails because people can't picture themselves in it. Role-specific examples fix that. A customer support agent and a financial analyst have completely different AI use cases. Treat them differently. You are a training content writer for a [company size] [industry] company. Write three brief "scenario examples" for our AI policy: one for a customer support role, one for a finance or operations role, and one for a marketing or communications role. Each scenario should show a realistic work task, describe how the employee could appropriately use AI for it, and flag what they should NOT do in that scenario (e.g., don't paste customer account data, don't let AI generate a financial figure without verification). Keep each scenario to 75-100 words. Use fictional names. Prompt 7: Write a manager review checklist Policies don't enforce themselves. Managers need to know what they're looking for. This prompt generates a practical checklist they can actually use in team reviews or 1:1s. You are an HR and operations writer for a [company size] [industry] company. Create a one-page manager checklist for reviewing AI use on their team. Include: whether team members know which tools are approved, whether anyone has shared restricted data with AI tools (ask, don't assume), whether AI outputs are being reviewed before use, whether any AI-related mistakes have occurred and been reported, and whether team members have questions about the policy. Format as a simple checklist with space for notes. For a broader look at how to actually start using AI at work without the hype, that guide covers the practical side managers often overlook. Prompt 8: Draft a lightweight employee FAQ Nobody reads the full policy. They read the FAQ. Write a good one. You are an internal comms writer for a [company size] [industry] company. Write a 6-question FAQ for employees about our new AI use policy. Questions should include: Can I use AI for my regular work tasks? What data can I share with AI tools? What tools are approved? Do I have to tell anyone if I use AI on something? What do I do if I make a mistake using AI? Where do I go with questions? Write answers in plain English, under 75 words each. Don't invent legal answers. Prompt 9: Create an incident response plan for accidental data sharing Someone will eventually paste something they shouldn't have. You need a plan for that before it happens, not after. You are a security and HR policy writer for a [company size] [industry] company. Draft a short incident response plan for the scenario where an employee accidentally shares restricted data with an external AI tool. Include: what the employee should do immediately (stop, don't share further, report), who to report to and how, what information to capture when reporting, how the company will respond, and whether the incident will affect the employee's standing (frame this as non-punitive to encourage reporting). Keep the whole plan under 300 words. Note that legal implications should be reviewed by counsel. Prompt 10: Turn the policy into a one-page rollout memo You've got a policy. Now you need people to actually read it. A one-page memo that explains why the policy exists, what's new, and what employees need to do by when, is more likely to get read than a 12-page PDF attachment. Pair this with the 12 AI work prompt templates if you want practical follow-on resources for employees. You are an internal comms writer for a [company size] [industry] company. Write a one-page email memo announcing our new AI use policy to all employees. Include: a brief explanation of why we're putting this policy in place (not alarming, practical), two or three key things employees need to know immediately, a link to the full policy (placeholder: [LINK]), who to contact with questions, and the date by which employees should confirm they've read the policy. Keep the tone direct and non-threatening. Under 300 words. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What AI can't do for your policy AI is good at first drafts. It's bad at making real tradeoffs. It doesn't know your industry's regulatory environment. It doesn't know which clients you've made confidentiality commitments to. It doesn't know that your IT team has a specific reason for blocking a certain tool, or that your legal counsel has opinions about what "approved" actually means for liability purposes. Every section of your AI policy still needs a human sign-off from whoever owns that domain: legal for liability language, security for tool approvals and data categories, HR for employment implications, and leadership for anything that touches client relationships. The prompts above are a way to do the drafting work faster and more coherently. They're not a way to skip the organizational thinking. That part is yours. If you want to understand what AI can and can't actually do before you write rules about it, that's worth 5 minutes of your time before you start. Frequently asked questions Can I use AI to write my company's AI policy? Yes, with supervision. AI can produce solid draft policy sections quickly, but it will sometimes invent compliance requirements or make tradeoffs without flagging them. Use AI for the drafting, then have legal, HR, and security review the output before anything gets published. What data should employees never paste into ChatGPT or Claude at work? At minimum: customer names and contact data, employee personal records, financial details, unreleased product or strategy information, legal correspondence, security incidents, credentials, and anything marked confidential or regulated. Use anonymized fictional examples instead when prompting AI for work tasks. How specific does an AI policy need to be? Specific enough that an employee can tell, without calling HR, whether a task they're doing right now is allowed. "Use AI responsibly" is not a policy. "Don't paste customer data into external AI tools; use approved tools listed here for these specific tasks" is a policy. Do I need a lawyer to write an AI workplace policy? You don't need a lawyer to draft one, but you should have legal counsel review the final version, especially sections covering data handling, employee obligations, and incident response. AI-generated policy language can sound legally plausible while being incorrect for your jurisdiction or industry. What's the most common mistake in company AI policies? Being vague about data. Policies that say "don't share sensitive information" without defining what sensitive means, where the line is, or what happens if someone crosses it, don't change behavior. The approved/not-approved data table in Prompt 2 above is usually the most useful thing a policy can include. How do I get employees to actually read and follow the AI policy? Keep it short, write a plain-English FAQ, use role-specific examples they can see themselves in, and announce it with a clear memo rather than burying it in a policy wiki. Prompt 10 handles the memo. The FAQ and examples are Prompts 8 and 6. None of this is complicated. It's just easier to skip. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Briefing Document Prompts: 10 Templates That Work 10 copy-paste AI briefing document prompts for meetings, exec decisions, client calls, and launch readiness. With a reusable formula and safety rules. ai promptsbriefing documentmeeting preparation AI Executive Summary Prompts: 10 Templates 10 copy-paste AI executive summary prompts for project updates, board briefs, risk summaries, and more. Includes the 5-part prompt formula. ai promptsexecutive summaryworkplace ai AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation --- ## AI Post-Launch Monitoring Prompts: 10 Templates URL: https://dontreplace.me/articles/ai-post-launch-monitoring-prompts-10-templates ai promptspost-launch monitoringproject managementlaunchescalation AI Post-Launch Monitoring Prompts: 10 Templates 10 copy-paste AI prompts for post-launch monitoring: watch plans, ticket summaries, escalation triggers, rollback decisions, and stakeholder updates. July 27, 2026 Updated July 27, 2026 You shipped. Congrats. Now comes the part nobody puts in the launch plan: the next 24 hours of anxious dashboard-staring, Slack messages that start with "hey quick question," and the sinking feeling that the support queue is moving faster than normal. AI post-launch monitoring prompts can help you structure that chaos. Not replace your judgment, just stop you from drowning in signals while the real problem hides behind a green status light. Here are 10 copy-paste prompts, a reusable formula, and the safety rules you need before you paste anything into ChatGPT or Claude. What AI can actually do in post-launch monitoring AI is fast. It's not smart. That's Rule #5 from Don't Replace Me: the model can organize your post-launch signals in seconds, but organized noise is still noise until a human checks the source. Where AI genuinely helps: structuring your watch plan, summarizing ticket themes, drafting stakeholder updates, spotting missing monitoring signals, building escalation criteria. Where it will confidently mislead you: inventing system health, fabricating reliability numbers, implying that a tidy summary means nothing is on fire. Use it for the organizing work. Keep the judgment for yourself. AI can write your escalation trigger. It cannot tell you whether the production system is actually stable. That call is yours. The safety rules before you paste anything Before the prompts, a hard stop. Do not paste any of the following into unapproved AI tools: Customer names, emails, or any personally identifiable information (PII) Raw production logs or server outputs Access tokens, credentials, API keys Private support tickets with customer details Unreleased strategy, pricing, or roadmap details Legal disputes or compliance issues Security vulnerabilities or incident details Financial records, HR issues, or regulated data Medical information or private client conversations Summarize and anonymize first. If your company has an approved AI tool with a data processing agreement, use that. If you're unsure, ask your legal or security team before the launch, not after. The reusable post-launch monitoring prompt formula Every prompt in this article follows this structure. Use it to build your own when the templates don't quite fit. You are a [role]. I need help [specific monitoring task]. Context: - [What launched and when] - [What you're watching] - [What normal looks like vs. what you're seeing] Constraints: - [Data sources you're using] - [Who needs to approve/review] - [Privacy or security boundaries] Output format: [checklist / summary / draft / table / escalation criteria] Do not invent metrics, assume system health, or include customer PII. That last line matters. Add it every time. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI post-launch monitoring prompts Prompt 1: Define your first 24-hour watch plan You are a project manager. Help me build a first 24-hour monitoring plan for a product launch that went live today. Context: - What launched: [brief description, no sensitive details] - Key success metrics: [list your KPIs] - Known risk areas: [performance, integrations, user flow, etc.] - Team availability: [who is on call and when] Create a time-blocked watch schedule covering: 0–2 hours, 2–6 hours, 6–12 hours, 12–24 hours. For each window, list: what to check, who is responsible, what threshold triggers escalation, and what the escalation path is. Do not assume any system is stable. Flag gaps where monitoring coverage is missing. Prompt 2: Turn a dashboard into a monitoring checklist You are an operations analyst. I'm going to describe what I see in our post-launch dashboard. Help me turn this into a structured monitoring checklist. Dashboard summary (no credentials, no PII, no raw logs): [Describe metrics in plain language: e.g., "error rate is 0.3%, page load is 2.1s, conversion is tracking 4% below expected"] For each metric, create a checklist row with: - Metric name - Current status (green / amber / red based on my description) - What would trigger a status change - Suggested next check time - Owner role (not individual names) Do not invent thresholds. Flag any metric where I haven't given you a baseline to compare against. Prompt 3: Summarize support tickets without leaking PII You are a support analyst. I'm going to give you a plain-language summary of support ticket themes from the first [X] hours post-launch. No individual customer details included. Ticket themes summary: [e.g., "12 tickets about login errors, 7 about slow load times on mobile, 3 about incorrect pricing display, 2 about email confirmation not arriving"] Please: 1. Group themes by severity (high / medium / low) 2. Flag any theme that suggests a systemic problem vs. isolated issues 3. Identify which themes need engineering review vs. support response 4. Draft a one-paragraph internal summary for the team Do not speculate about root causes. Do not invent ticket counts or customer sentiment beyond what I've described. Prompt 4: Compare expected vs. actual behavior You are a QA analyst. Help me structure a comparison between expected launch behavior and what we're actually seeing. Expected: [List what you planned: e.g., "checkout flow completes in under 3 steps, confirmation email fires within 2 minutes, error rate under 0.5%"] Actual (first [X] hours): [List what you're observing from approved dashboards only] Output: - Side-by-side comparison table (Expected / Actual / Gap / Severity) - List of gaps that need immediate investigation - List of gaps that can wait for the next review window - Any patterns that suggest a single root cause Do not assume the gaps are acceptable. Do not invent explanations for discrepancies. This is one of the most useful prompts in the set. You probably built a launch checklist before you shipped (if not, the AI launch checklist prompts are worth running before your next one). This prompt is how you close the loop on it. Prompt 5: Triage customer feedback themes You are a product analyst. I'm going to give you anonymized themes from customer feedback in the first [X] hours post-launch. No names, emails, or account details included. Feedback themes: [e.g., "confused by new navigation layout, pricing unclear, loved the speed improvement, several asking about missing feature X, one complaint about accessibility on mobile"] Please: 1. Separate feedback into: bugs, UX issues, missing features, positive signals 2. Flag any feedback that suggests a critical user impact 3. Identify which items need engineering vs. product vs. support response 4. Suggest a prioritization order with brief reasoning Do not invent customer sentiment or fabricate feedback volume. Prompt 6: Find missing monitoring signals You are a systems analyst. Review this list of what we're currently monitoring post-launch and identify gaps. What we're currently monitoring: [List your active monitoring: e.g., "uptime, error rate, page load, conversion rate, support ticket volume"] What we launched: [Brief description of the feature or system, no sensitive architecture] Known dependencies: [e.g., "payment gateway, email service, third-party data feed"] Please identify: - Monitoring gaps for the features we launched - Dependency health signals we're not watching - User journey steps with no monitoring coverage - Suggested metrics to add before the next review window Do not assume the current monitoring is sufficient. Prompt 7: Draft a stakeholder status update You are a project manager. Draft a post-launch status update for [audience: exec team / cross-functional leads / board] based on the following summary. Launch: [what went live] Time since launch: [X hours] Current status: [stable / monitoring / investigating / escalated] Key metrics (from approved dashboards only): [brief summary] Known issues: [list issues in plain language, no PII] Next review: [time and who is responsible] Format: - 3-sentence executive summary - Status table (area / status / owner / next action) - One clear ask or decision needed from this audience Do not invent metrics. Do not imply problems are resolved if they are still being investigated. If you need more stakeholder update templates for different scenarios, the AI stakeholder update prompts article has 10 variations. Prompt 8: Create escalation triggers You are an operations lead. Help me define escalation triggers for the first [X] hours post-launch. What we launched: [brief description] Key metrics we're watching: [list] Current team on call: [roles only, no names] Escalation path: [who gets notified at what level] For each metric, define: - Green threshold (no action) - Amber threshold (internal alert, begin investigation) - Red threshold (escalate immediately, consider rollback review) Also define: - Time-based triggers (e.g., "if amber for more than 30 minutes...") - Volume-based triggers (e.g., "if support tickets exceed X per hour...") - Dependency failure triggers Do not set thresholds without my input. Ask me for any baseline data you're missing rather than inventing it. Prompt 9: Decide whether to continue monitoring vs. trigger a rollback review You are a technical program manager. Help me structure a continue-monitor vs. rollback-review decision for our current situation. Current status summary: [Plain-language description of what you're seeing, no sensitive data] Known issues: [List issues with severity as you understand them] Rollback readiness: [Is rollback available? What's the estimated time and impact?] Customer impact so far: [Plain-language summary, no PII] Please build a simple decision framework: - Criteria that support continuing to monitor - Criteria that support triggering a rollback review - Questions I need to answer before making this call - Who needs to be in the room for this decision Do not make the rollback recommendation for me. This decision requires human judgment and executive approval. The rollback question is one where AI can help you think, but the actual call needs a human with authority and context. The AI rollback plan prompts have the templates for what comes after this decision. Prompt 10: Convert launch lessons into follow-up work You are a project manager. Help me turn post-launch observations into structured follow-up work items. Launch observations (anonymized, no sensitive details): [List what you learned: monitoring gaps, unexpected behaviors, process breakdowns, things that went well] For each observation, create a work item with: - Summary of the observation - Category (process / technical / communication / monitoring) - Suggested action - Priority (high / medium / low) - Suggested owner role - Rough effort estimate (hours / days) Format as a backlog table. Do not assign work to individuals by name. Do not invent observations I haven't given you. What AI cannot do in post-launch monitoring A clean dashboard summary is not the same as a healthy system. This is the part most people skip. AI will not tell you when your monitoring is measuring the wrong thing. It won't notice that your error rate looks fine because the feature broke silently before triggering an error. It won't know that the stakeholder update you just drafted is technically accurate but politically incomplete. It won't flag that the support ticket summary sounds fine but the tickets themselves contain a pattern that only becomes obvious when you read three of them in sequence. That gap is where you earn your keep. Dmitry Kargaev calls it the taste moat in Don't Replace Me: the ability to know which symptoms actually matter, which complaints are edge cases, and when a polished summary is hiding a real problem. If you want to see where else AI falls flat in ways that matter for your work, the plain-language breakdown of what AI can and can't do is a good five minutes. The prompts organize the work. The judgment about whether to pull the trigger on a rollback, escalate to the exec team, or hold and watch for another hour: that's still yours. After the dust settles on a launch, the next thing most teams skip is a proper incident or lessons-learned review. The AI incident review prompts are built for that exact conversation. Frequently asked questions Can I use AI to monitor my production system automatically? Not with the prompts in this article. These are for structuring human-led review, not automated monitoring. For automated alerting and production monitoring, you need purpose-built tools like PagerDuty, Datadog, or similar, with proper security review and data handling agreements in place. What's the biggest mistake teams make with AI in post-launch monitoring? Treating a tidy AI-generated status update as evidence that things are fine. AI organizes the information you give it. If you give it incomplete or cherry-picked inputs, you get a polished summary of an incomplete picture. Garbage in, garbage out: always verify the underlying data from approved sources before sharing any AI-assisted summary. How do I summarize support tickets for AI without including customer PII? Strip all names, email addresses, account IDs, and any identifying details before you paste anything. Work from ticket themes and counts only. For example: "14 tickets about checkout errors, 6 about email confirmation delays" is fine. Pasting the actual ticket text with customer details is not. If your company has a privacy-approved AI tool, use that instead. Should I tell my boss I used AI to write the stakeholder status update? That depends on your workplace. The should you tell your boss you use AI article covers the disclosure question in detail. Short version: the update needs to be accurate and you need to stand behind it regardless of how it was drafted. What's the difference between a monitoring checklist and an escalation trigger? A monitoring checklist tells you what to watch and when. An escalation trigger tells you what threshold means you need to act, and what that action is. You need both. The checklist keeps you looking at the right signals; the trigger removes the ambiguity about when a signal becomes a problem that requires a response. Can AI tell me whether to roll back a launch? No. AI can help you structure the criteria and think through the trade-offs. The actual rollback decision requires a human with executive authority, context about customer impact, and accountability for the outcome. Use Prompt 9 above to structure the decision, then get the right people in the room to make it. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Assumption Log Prompts: Stop Planning on Guesswork Use 10 AI assumption log prompts to expose project guesswork, design evidence checks, assign review owners, and stop unverified beliefs becoming commitments. AI promptsproject managementassumption logs AI Contingency Plan Prompts: 10 Templates 10 copy-paste AI contingency plan prompts for ChatGPT and Claude. Map failures, assign owners, draft comms, and verify assumptions before a crisis hits. contingency planningAI promptsbackup plans AI Dependency Mapping Prompts: Find What Breaks AI dependency mapping prompts for finding upstream work, downstream risks, hidden handoffs, owners, schedule conflicts, and launch blockers. AI promptsdependency mappingproject management --- ## AI Presentation Prompts: 10 Templates That Work URL: https://dontreplace.me/articles/ai-presentation-prompts-10-templates-that-work ai presentation promptschatgpt for presentationsai at workpresentation tipsprompt templates AI Presentation Prompts: 10 Templates That Work 10 copy-paste AI presentation prompts for slide outlines, speaker notes, executive Q&A prep, and more. No jargon, no fluff, just practical templates. May 28, 2026 Updated May 28, 2026 Most presentations are secretly documents with a clicker. You know the ones. Forty bullet points per slide, a font size that requires binoculars, and a "Key Takeaways" slide that repeats everything you just said. AI presentation prompts won't fix bad thinking, but they'll fix the stuff that eats your time: turning notes into structure, cutting word soup, writing speaker notes, anticipating the questions your VP will ask five minutes before you go live. This is a copy-paste guide. You don't need a course. You need 10 prompts and the judgment to check the output before it leaves your laptop. The master formula behind every good AI presentation prompt Every prompt that actually works has five parts. Role, context, task, constraints, format. That's it. Mess with any one of them and you get output that's technically words in an order. Here's the formula written out: You are [role]. I need [task]. Here is the context: [context]. Constraints: [constraints]. Give me the output in [format]. This is what Rule #13 from Don't Replace Me calls the "smart intern" model: treat AI like someone who is fast, tireless, and occasionally full of confident nonsense. Give it a job description and guardrails, and it does good work. Give it a vague request and you get a presentation called "Navigating the Path Forward Together." A note before we start: don't paste confidential data into ChatGPT or Claude unless your company's policy explicitly allows it. No customer names, no internal financials, no roadmap details, no HR information. Anonymize everything. Replace "Acme Corp increased revenue 37% in Q3" with "a mid-size client saw significant Q3 revenue growth." The prompts below work fine with anonymized inputs. The 10 AI presentation prompts you'll actually use Prompt 1: Clarify your audience and goal before you build anything This is the one people skip. They start building slides before they've decided what the presentation is supposed to do. Ask the AI to force you to think. Copy this: You are a presentation coach. I'm building a presentation for [audience, e.g., "senior marketing leadership, skeptical of budget increases"]. Ask me five clarifying questions that will help me nail down the single goal of this presentation, the main decision I want the audience to make, and what they already believe that I need to either confirm or change. Read its questions. Answer them. The act of answering is the useful part, not the questions themselves. Prompt 2: Turn messy notes into a storyline You have a doc full of half-thoughts, Slack messages, and three conflicting data points. This prompt gets you from chaos to a story arc in one pass. Copy this: You are a presentation strategist. Here are my raw notes: [paste anonymized notes]. Based on these, suggest a three-act narrative structure for a [length]-minute presentation to [audience]. Act 1 should set up the problem or opportunity. Act 2 should present the approach and evidence. Act 3 should land on a clear recommendation or call to action. Flag any gaps in the story where I'll need more evidence. Check the flagged gaps. Those are your homework, not the AI's. Prompt 3: Create a slide-by-slide outline Once you have a storyline, turn it into a skeleton. This prompt gives you the bones before you start filling in the slides. Copy this: You are a presentation writer. Here is my narrative outline: [paste from Prompt 2]. Create a slide-by-slide breakdown for a [number]-slide deck. For each slide, give me: the slide number, a draft title, one sentence describing what this slide proves or communicates, and the type of content that should go on it (e.g., one chart, a 3-point list, a single quote, a before/after comparison). Do not write full slide text yet. Don't skip the "type of content" column. It forces slide-level clarity before you start filling in words. Prompt 4: Rewrite weak slide titles Most slide titles are labels, not arguments. "Q3 Results" tells the audience nothing. "Q3 results came in 12% above forecast" tells them something. This prompt fixes that fast. Copy this: You are a presentation editor. Here is a list of my current slide titles: [list them]. Rewrite each one so it makes a specific claim or tells the audience what to conclude from that slide, not just what the slide is about. Each rewritten title should be under 10 words. Give me the original and the rewrite side by side. You don't have to use every rewrite. But at least one of them will be sharper than what you had. Prompt 5: Cut the text on a slide If you're pasting paragraphs onto slides, you're writing a document. This prompt is brutal in the right way. Copy this: You are a slide editor. Here is the text from one of my slides: [paste slide text]. Your job is to reduce this to the minimum number of words needed for someone in the room to follow along. Maximum three bullet points. Maximum seven words per bullet. Cut everything that the speaker can just say out loud. Show me the before and after. The stuff it cuts is what goes in your speaker notes, not the trash. Prompt 6: Write speaker notes Good speaker notes are not a transcript of the slide. They're what you'd say to someone smart who asked "okay, but what does this actually mean?" Copy this: You are a presentation coach writing speaker notes. Here is slide [number]: [title and bullet points]. Write speaker notes that help the presenter: open the slide with a single sentence that frames why this slide matters, explain each bullet point in plain conversational language, and close with a transition sentence to the next slide. The presenter's tone is [professional/casual/data-driven]. Keep the total under 150 words. Read the notes out loud before the presentation. If any sentence sounds like something a robot would say, fix it yourself. Prompt 7: Translate jargon into plain English This matters most when your audience is mixed: technical and non-technical, finance and operations, customers and internal stakeholders. The jargon that feels natural to you is a wall to half the room. Copy this: You are a plain-English editor. Here is a section of my presentation: [paste]. Rewrite it for an audience that is smart but has no background in [your field]. Replace acronyms with their full meaning on first use. Replace technical terms with the clearest plain-English equivalent. Flag any concept that might still need a brief explanation. Don't go so plain that you lose precision. If the AI replaces a specific term with something vague, push back and ask for a more accurate plain-English version. This is exactly what AI is genuinely good at, by the way. Language transformation. If you want a clearer sense of where AI earns its keep versus where it guesses confidently, the 5-minute breakdown on what AI can and can't do is worth a read. Prompt 8: Prepare for executive questions You've got 20 minutes and you know the last five will be questions from someone who didn't read the pre-read and wants to challenge your numbers. Prepare for that person. Copy this: You are a senior executive reviewing a presentation on [topic]. Your job is to ask the five hardest, most skeptical questions this executive team would ask. Be specific. Questions should challenge: the assumptions behind the recommendation, the data quality or sourcing, the risks that aren't addressed, the cost or resource implications, and whether the conclusion actually follows from the evidence. Then for each question, suggest the key points I should be ready to answer. Use the suggested answers as a starting point. The facts in those answers are yours to verify. Don't walk into the room trusting that the AI got the numbers right. Prompt 9: Rehearse a five-minute version Every presentation should have a five-minute version. Either because the meeting runs short, or because someone asks "give me the quick version" in the elevator, or because it clarifies what actually matters. Copy this: You are a presentation coach. Here is my full slide outline: [paste]. Compress this into a five-minute verbal summary I can say out loud without slides. Include: one sentence on the problem or opportunity, two to three sentences on the approach or evidence, one clear recommendation, and one sentence on what I'm asking for. This should be something I can say naturally in conversation, not a formal script. Practice it. Out loud. Twice. The AI wrote it; you have to own it. Prompt 10: Run a final honesty and fact check This is the one most people don't do. They push send on the deck without asking: is this actually true? Is this claim defensible? Is there anything in here I'd be embarrassed to be asked about? Copy this: You are a fact-checking editor. Review this presentation section by section: [paste the full deck outline or speaker notes]. For each claim, mark it as: (A) verifiable fact I should confirm with a source, (B) reasonable inference I should be ready to defend, or (C) vague or unsupported assertion I should cut or clarify. Also flag any promises, projections, or commitments that I should make sure are authorized before presenting. The AI will flag things. Take the flags seriously. Every metric, date, customer example, and forward-looking statement is your responsibility, not the tool's. Quality and safety rules for every presentation you build with AI These aren't optional. They're the difference between using AI well and using it to embarrass yourself in front of your VP. On data: Verify every number before you present it. AI can hallucinate statistics, dates, and study citations. If you can't find the source in 60 seconds, cut the claim or change it to something you can defend. On confidentiality: Don't paste customer names, internal financial data, headcount information, legal matter details, unreleased product plans, or anything covered by NDA into a public AI tool. Your company probably has a policy. Check it. On promises: If the AI writes you a slide that says "we will deliver X by Y," make sure someone with authority has approved that commitment. The AI doesn't know your org chart. On taste: AI can generate a lot of plausible-sounding output. Plausible isn't the same as good. Someone in the room has to decide whether the slide structure makes sense, whether the tone is right for the audience, and whether the argument actually holds. That person is you. The broader guide to using AI at work without the hype covers this in more depth if you want the full picture. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → One thing AI genuinely can't do for your presentation It can't read the room. It doesn't know that your CFO hates percentages without absolutes. It doesn't know that the last person to present this topic lost the room on slide three and now everyone's defensive. It doesn't know that the CEO checks out when decks go past 12 slides, or that the VP of Engineering trusts you because of a project you worked on three years ago. All of that is yours. The relationship, the read, the judgment call about when to go slower and when to cut to the answer. AI can help you build the language-shaped parts of the presentation. The strategy, the taste, the credibility: those belong to you. That's the core idea in how to use ChatGPT at work without losing your edge, and it applies to every presentation tool in the stack. Frequently asked questions Can I use AI to build a full presentation from scratch? You can use AI to structure, draft, and refine a presentation, but "from scratch" is a stretch. AI needs your raw material: the goal, the audience, the key facts, the recommendation. It can't invent those. Give it context and it'll help you build something. Give it nothing and it'll give you a deck about synergy and innovation. Is it safe to paste my presentation into ChatGPT? Depends on what's in it. Don't paste confidential client data, internal financial projections, unreleased product details, or anything covered by NDA unless your company's AI policy specifically allows it. Anonymize anything sensitive and you're usually fine. What's the best AI tool for presentations? ChatGPT and Claude are both good for the text and structure work covered in this article. If you want AI that directly manipulates slide files, tools like Gamma or Beautiful.ai do that, but you'll still need to apply judgment to the output. Start with the thinking before you worry about the tool. How do I stop AI-generated slides from sounding generic? Give the AI your actual context. Audience, goal, the specific thing you're recommending, the tone of your organization. Generic prompts produce generic output. The more specific you are about the room you're presenting to, the less the output sounds like a consulting deck from 2019. Will AI replace presentation designers and writers? For basic decks, some of that work will shrink. For high-stakes presentations where the argument, the data, and the relationship all have to land: no. The judgment layer doesn't automate. If you want a longer read on which jobs AI is actually threatening and which ones it isn't, that's a good starting point. Do I need a prompt engineering course to use these? No. The formula in this article (role, context, task, constraints, format) is everything you need. Anyone trying to sell you a $400 course on prompt engineering for presentations is someone you should ignore. The real AI skills you need in 2026 don't involve a certificate. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Prioritization Prompts: 10 Templates That Work 10 copy-paste AI prioritization prompts for task lists, backlogs, roadmaps, and tradeoffs. Practical templates with real safety guardrails. AI prioritizationprompt templatestask management AI Prompts for Work: 12 Templates You Can Use Today 12 copy-paste AI prompt templates for emails, reports, and decisions. Real examples for normal people, with what to check before you send anything. ai prompts for workchatgpt promptsworkplace ai AI Stakeholder Update Prompts: 10 Templates 10 copy-paste AI stakeholder update prompts for weekly summaries, exec briefs, client updates, and blocker escalations. Real templates, honest results. ai stakeholder update promptsproject communicationstatus updates --- ## AI Prioritization Prompts: 10 Templates That Work URL: https://dontreplace.me/articles/ai-prioritization-prompts-10-templates-that-work AI prioritizationprompt templatestask managementdecision makingAI at work AI Prioritization Prompts: 10 Templates That Work 10 copy-paste AI prioritization prompts for task lists, backlogs, roadmaps, and tradeoffs. Practical templates with real safety guardrails. June 22, 2026 Updated June 22, 2026 Here's the thing about prioritization: everyone's list is "urgent and important." The whole team is overloaded, the roadmap has 47 items flagged as P1, and the meeting to discuss priorities has its own pre-meeting. AI prioritization prompts won't fix your org chart or your boss's inability to say no, but they'll help you think faster, structure tradeoffs better, and stop staring at a messy task list like it owes you an apology. These prompts work. They also have limits. Both things matter. Before you paste anything into ChatGPT or Claude: don't include customer PII, employee records, unreleased product details, private financials, confidential contracts, legal disputes, security incidents, or board materials. If your company has an approved AI tool policy, follow it. If it doesn't, use judgment. What AI prioritization prompts are actually good for AI is fast. It's not wise. That's the distinction that matters here, and it's the same one Dmitry Kargaev makes in Don't Replace Me when he talks about Rule #5: AI's speed is real, but speed is not judgment. A good prioritization assistant helps you sort, structure, score, and summarize. It does not know what your customers actually want, what your CEO will approve, or which deadline is real versus aspirational. That part is on you. Use these prompts to get a working draft of a priority ranking, a pressure-tested list of assumptions, or a stakeholder-ready summary of tradeoffs. Then review it before you share anything. AI will produce confident-sounding output whether or not the inputs were solid. Garbage in, polished nonsense out. If you want to understand the broader picture of where AI helps and where it genuinely breaks down, the five-minute explainer on what AI can and can't do is worth your time before you go further. The reusable AI prioritization prompt formula Every prompt in this article follows the same structure. Learn it once and you can write your own. Context + your actual goals + real constraints + what you want back That's it. Vague input gets you vague output dressed up in a scoring rubric. Specific input gets you something you can actually use. When you give AI a list with no context, it will happily invent impact estimates, guess at deadlines, and produce a ranked list that looks authoritative and means nothing. Your job is to supply the reality: what matters, what the limits are, and what decision you're trying to make. Here's the base formula written out: "I'm working on [specific context]. My goal is [real, specific goal]. Constraints include [time, budget, team capacity, dependencies, risk factors]. Here is my list: [paste items]. Please [specific output: rank, score, summarize, pressure-test, reframe]. Flag any assumptions you're making and anything I should verify before using this." Every prompt below is a variation on this. Copy, customize, use. 10 copy-paste AI prioritization prompts 1. Triage a messy task list Use this when your task list has gotten out of hand and you need to stop staring at it. "Here is my current task list: [paste list]. My role is [your role]. This week's main goal is [specific outcome, e.g., 'ship X' or 'close Q3 planning']. Team capacity is [X hours or people available]. Please sort these tasks into three buckets: do this week, defer, and drop or delegate. Flag anything where you're guessing about priority and tell me what information I'd need to confirm it." Review the output before acting on it. AI doesn't know which items have a hard external deadline or which "low priority" task will cause a stakeholder crisis if it slips. 2. Prioritize a product backlog For product managers drowning in tickets, this structures the conversation without replacing it. "I have a product backlog with the following items: [paste items, stripped of any customer PII or unreleased confidential details]. Our current product goal is [specific goal]. Our main constraints are [team size, sprint capacity, tech debt situation, compliance requirements if applicable]. Please score each item on: customer impact (based only on what I've told you), effort, and strategic fit. Produce a draft ranked list. Flag where you have no real signal and are guessing, and list the top three assumptions I should pressure-test with real customer evidence before finalizing." Do not let AI invent customer demand or usage data. If it says "this feature would likely be highly valued by users," that's a guess. Treat it as a hypothesis, not a fact. 3. Choose between competing projects When leadership wants three things done and you have capacity for one. "I need to choose between these projects: [describe each in 2-3 sentences, no confidential financials or client names]. Our team goal for this quarter is [specific goal]. Constraints: [budget range, headcount, timeline, key dependencies]. Please compare these projects on: strategic alignment, estimated effort, risk level, and reversibility if we choose wrong. Summarize the tradeoffs in plain language I could use in a stakeholder conversation. Don't invent revenue impact or ROI numbers. Note where I need real data before deciding." Reversibility is underrated. If a project is hard to undo, that's a risk factor worth naming. This prompt asks AI to name it. For deeper structure on how to frame these decisions, the AI decision-making prompts guide has templates built for exactly this kind of situation. 4. Score roadmap ideas When you're building a roadmap and need a scoring draft before the big meeting. "Here are [X] roadmap ideas I'm evaluating: [list them]. Please score each on a simple 1-5 scale for: customer value, implementation complexity, and strategic priority. Use only the information I've provided. Flag every item where your score is based on an assumption I haven't confirmed. Produce a draft scoring table and note the top two questions I should answer before presenting this to leadership." This gives you a starting point for a real conversation, not a final answer. Don't present AI-generated scores as team decisions. They're a draft. Label them that way. 5. Decide what to stop doing The hardest prioritization question is what to kill. This prompt helps. "Here is a list of ongoing activities, projects, or commitments my team currently maintains: [paste list, no confidential client details]. Our current goal is [specific goal]. Resources are constrained because [specific reason]. Please identify which items are candidates to stop, reduce, or hand off based on strategic fit and resource cost. For each candidate to stop, note the likely cost of stopping and what I'd need to confirm before making that call. Don't assume executive approval or budget decisions have been made." Stopping things requires real conversations with real people. This prompt helps you identify candidates and prepare for those conversations. It doesn't replace them. 6. Rank urgent vs. important work The classic Eisenhower matrix, but you supply the context so the output is actually useful. "Here is my current work list: [paste list]. Please sort these into four categories: urgent and important, important but not urgent, urgent but not important, and neither. For each item, note what signal you used to categorize it and flag where you're guessing because I haven't told you enough. Suggest which 'urgent but not important' items could be delegated, and which 'important but not urgent' items I'm most likely to keep deferring if I'm not careful." That last part matters. AI is good at naming the thing you're avoiding. The honest version of this output will have several items in "important but not urgent" that have been there for three months. 7. Pressure-test a priority list You have a list. You want to know what you're missing before you share it. "Here is my current priority list: [paste]. My goal is [specific goal]. Please act as a skeptical reviewer and identify: gaps (things I might have missed), risks (what could go wrong with this ordering), hidden dependencies (items that can't proceed without something else I haven't listed), and assumptions that might not hold. Don't soften this. I want the uncomfortable version." This is one of the most useful prompts in the set. AI is polite by default. Telling it explicitly to give you the uncomfortable version usually gets you something more honest. Follow this up with the AI risk assessment prompts if you want to go deeper on any specific item. 8. Summarize tradeoffs for stakeholders When you need to communicate a hard prioritization call to people who weren't in the room. "I've decided to prioritize [X] over [Y and Z]. Here's my reasoning: [explain in 3-5 sentences]. The main tradeoffs are [list what's being deprioritized and why]. Please help me write a brief stakeholder summary (3-4 short paragraphs) that explains the decision clearly, acknowledges what we're not doing, and frames next steps. Don't invent reasons I haven't given you. Flag anything that sounds like it needs sign-off or legal/finance review." Review this before you send it. AI is good at polishing stakeholder language. It does not know what your actual organizational dynamics are. If the summary sounds too clean, that's a sign something real is being smoothed over. 9. Turn priorities into a weekly plan Once you have a priority list, this converts it into something you can actually execute. "Here are my top priorities for this period: [paste]. My available working time this week is approximately [X hours]. Please help me build a rough weekly plan that allocates time to these priorities in order of importance. Flag where there's not enough time to do everything and recommend what should be deferred or reduced in scope. Don't assume I can extend the deadline or add resources." This is practical, not magical. It's a scheduling draft. You'll adjust it. But starting with a structured draft is faster than starting from scratch on a Monday morning. For connecting priorities to project structure, the AI project management prompts and the workflow audit prompts are useful next steps. 10. Create a follow-up review checklist Prioritization isn't a one-time event. This prompt helps you build a lightweight review habit. "My current priority list is: [paste]. Please generate a simple review checklist I can use in [weekly/biweekly/monthly] reviews to check whether these priorities are still correct. Include questions about: whether the original goal has changed, whether any new information should shift the ranking, which items have stalled and why, and what should be added or removed. Keep it to 10 questions or fewer. Format it as a plain checklist." Priorities go stale. Teams change direction. Roadmaps shift. A five-minute checklist review beats a two-hour reprioritization meeting that could have been avoided. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What AI should never do in prioritization work This matters as much as the prompts themselves. AI should not invent customer demand signals. If it says "users are likely to value this feature," that's a guess. You need actual evidence: support tickets, interviews, usage data, sales calls. AI should not set deadlines. It doesn't know your contractual obligations, your executive commitments, or your legal calendar. Treat any deadline it suggests as a placeholder. AI should not name owners. Accountability has to come from real people who agreed to it. A prompt output that says "Owner: Product Manager" means nothing until a real person has agreed. AI should not replace budget conversations, legal review, security assessment, or direct conversations with the teams and customers affected by your decisions. It can help you prepare for those conversations. It can't substitute for them. The human value in prioritization isn't sorting a list. It's knowing which tradeoffs are acceptable, which risks are real, and which numbers someone made up. That's taste, and AI doesn't have it. Frequently asked questions Can AI help me prioritize my work tasks? Yes, with caveats. AI is useful for structuring a messy list, scoring options against criteria you define, surfacing assumptions, and summarizing tradeoffs in plain language. It can't tell you which deadline is real, what your customers actually want, or what your organization will actually support. Give it context, review the output, and treat the result as a draft. What's the best prompt for prioritizing a task list? Start with this structure: tell the AI your current list, your actual goal for the week, your capacity constraints, and ask it to sort tasks into do, defer, and drop. Ask it to flag where it's guessing. The more specific your context, the more useful the output. Vague prompts produce polished nonsense. Is it safe to paste my work tasks into ChatGPT? It depends on what's in the list. Don't include customer PII, employee records, confidential financials, unreleased product details, legal disputes, contracts, or security incidents. If your company has an AI tool policy, follow it. For most day-to-day task lists, a de-identified version is fine. When in doubt, remove specifics. Will AI prioritization replace the need for strategy conversations? No. That's actually the most important thing to remember. A ranked list isn't a strategy. Real prioritization involves knowing what the organization actually cares about, what customers actually need, what resources actually exist, and which tradeoffs people are actually willing to accept. AI helps you structure the conversation. It doesn't replace it. What information should I always include in a prioritization prompt? Your actual goal (not a vague mission statement), your real constraints (time, money, people, dependencies), the items you're choosing between, and what output you want. Also ask it to flag assumptions and identify what you'd need to verify before acting on the results. How often should I re-prioritize using AI? A lightweight weekly review works well for most roles. Use the review checklist prompt (number 10 above) to check whether your priorities have shifted before the week starts. Bigger reprioritization conversations (quarterly roadmap, annual planning) benefit from more structured prompts and definitely need human sign-off before any output gets shared up the chain. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Presentation Prompts: 10 Templates That Work 10 copy-paste AI presentation prompts for slide outlines, speaker notes, executive Q&A prep, and more. No jargon, no fluff, just practical templates. ai presentation promptschatgpt for presentationsai at work AI Prompts for Work: 12 Templates You Can Use Today 12 copy-paste AI prompt templates for emails, reports, and decisions. Real examples for normal people, with what to check before you send anything. ai prompts for workchatgpt promptsworkplace ai AI Stakeholder Update Prompts: 10 Templates 10 copy-paste AI stakeholder update prompts for weekly summaries, exec briefs, client updates, and blocker escalations. Real templates, honest results. ai stakeholder update promptsproject communicationstatus updates --- ## AI Project Closeout Prompts: Finish Work Right URL: https://dontreplace.me/articles/ai-project-closeout-prompts-finish-work-right ai project closeoutproject managementai promptswork templateslessons learned AI Project Closeout Prompts: Finish Work Right 10 copy-paste AI project closeout prompts for summaries, handoffs, lessons learned, and auditing whether your project is actually safe to close. July 15, 2026 Updated July 15, 2026 Project closeout is where accountability goes to die. Everyone's exhausted, the client is already half-thinking about the next thing, and someone just wants to send the final email and close the Jira board. That's exactly when a tidy AI-generated summary feels like a gift. You paste in your notes, you get back something that sounds authoritative, and you ship it. Done. Except "sounds done" and "is done" are different things. AI project closeout prompts can do a lot of the structural heavy lifting here, but if you feed them thin input, you get confident-sounding fiction. A polished closeout doc doesn't make unresolved dependencies disappear. It just makes them look resolved until something breaks in production at 11pm on a Thursday. Here's how to use AI to close a project properly, not just prettily. The formula behind every AI project closeout prompt Before the templates, the structure. Every useful project closeout prompt needs the same ingredients, regardless of what you're closing. You're asking AI to do one of five things: summarize verified delivery, compare what shipped against what was agreed, identify what's still open, draft communication for stakeholders, or flag whether it's actually safe to close at all. For each of those, your prompt needs: the scope and deliverables list, acceptance criteria or definition of done, your retrospective notes or status updates, open items with owners and dates, approval records (not assumptions), known risks and decisions, and who's responsible for what after you leave. If you don't have those inputs, stop. The problem isn't your prompt. The problem is your documentation. Feed AI thin evidence and you'll get a confident fake closure. The acceptance criteria prompts guide is worth checking before you start this process. Never paste customer PII, credentials, contracts, financial forecasts, HR issues, legal disputes, security vulnerabilities, medical/safety information, or private client conversations into an unapproved AI tool. When in doubt, anonymize or escalate. 10 AI project closeout prompts you can use right now Prompt 1: Summarize what shipped Use this to build the delivery summary section of your closeout report. You are a project manager writing a delivery summary. Here is a list of scope items and their completion status: [PASTE SCOPE LIST WITH STATUS] Summarize what was delivered, what was descoped and why, and what remains outstanding. Write in plain language for a non-technical stakeholder audience. Flag any items where status is ambiguous or unconfirmed. Don't let it paper over "in progress" items. If something has no confirmed delivery date and no named owner, that's an open item, not a summary note. Prompt 2: Compare delivery against acceptance criteria Use this when you need to show the client or leadership that what you built matches what was agreed. You are reviewing project delivery against agreed acceptance criteria. Here is the original acceptance criteria document: [PASTE CRITERIA] Here is evidence of delivery for each criterion: [PASTE EVIDENCE OR STATUS NOTES] For each criterion, indicate: Met / Partially Met / Not Met / Unconfirmed. List what evidence is missing and what approvals are outstanding. Do not mark anything as Met unless there is confirmed evidence. This is where most closeouts cut corners. "We think it's met" is not evidence. Named approver, date, and record of sign-off is evidence. Prompt 3: Extract unresolved open items Use this on your meeting notes, status updates, or issue logs to pull out anything that never got resolved. Review the following project notes and status updates: [PASTE NOTES] Extract every open item, risk, blocker, or unresolved question. For each one, note: the issue, who owns it, the last recorded status, and whether it has a resolution date. Flag anything with no owner or no resolution path. This is the one people skip. That's how open items get buried in a closeout doc and reappear six months later as incidents. Use it alongside your risk assessment prompts to make sure nothing with real blast radius slips through. Prompt 4: Draft a stakeholder closeout note Use this to write the final project communication to your main stakeholders. You are writing a project closeout communication to senior stakeholders. The project is: [PROJECT NAME] What was delivered: [BRIEF SUMMARY] What remains open: [OPEN ITEMS WITH OWNERS] Next steps and owners: [LIST] Key contacts going forward: [NAMES AND ROLES] Write a clear, professional closeout note. Keep it under 250 words. Flag any items that require stakeholder action or acknowledgment. Don't pretend everything is clean if it isn't. Stakeholders would rather know about a known limitation upfront than find out after handoff. Prompt 5: Create a client-safe closeout report Use this when the audience is external and you need to control the level of detail. You are creating a client-facing project closeout report. Internal notes and context: [PASTE INTERNAL SUMMARY] Deliverables confirmed with evidence: [LIST] Outstanding items and agreed owners: [LIST] Any limitations or known issues the client needs to be aware of: [LIST] Write a client-facing summary that is factual, professional, and transparent about any outstanding items. Do not include internal commentary, internal team performance notes, or cost information unless specified. Review this before sending. AI will sometimes include phrases that sound fine internally but read oddly to clients. You're still the editor. Prompt 6: Turn retrospective notes into lessons learned Use this after your retrospective session to produce a reusable lessons learned document. This pairs well with the retrospective prompts approach if you haven't run the retro yet. You are synthesizing retrospective notes into a lessons learned document. Here are the retrospective notes: [PASTE NOTES] Organize the lessons into three categories: What worked well, What didn't work, and What to do differently. For each lesson, note the context, the impact, and a specific recommendation for future projects. Avoid generic advice. Keep it specific to what happened here. Generic lessons learned are useless. "Improve communication" is not a lesson. "Stakeholder sign-off on scope changes was skipped twice and caused scope creep on both occasions; require written approval going forward" is a lesson. Prompt 7: Prepare a support or operations handoff Use this to brief whoever is inheriting the work after you step away. You are writing a support handoff document for a project transitioning to operations or ongoing support. What was built and how it works: [DESCRIPTION] Known issues or limitations: [LIST] Common failure modes and how to address them: [LIST] Escalation paths: [NAMES, ROLES, CONTACT INFO] Key documentation and where it lives: [LINKS OR LOCATIONS] Write a handoff document that would let someone with no prior project context support this system effectively from day one. Flag any gaps where documentation is missing or incomplete. If you don't have the answers to those inputs, the handoff isn't ready. An AI-generated handoff document built on assumptions is a liability, not a deliverable. The handoff document prompts go deeper on this. Prompt 8: Document known risks and decisions Use this to create a final record of the key decisions made during the project and any residual risks being transferred to the operating team. Review the following decision log and risk notes: [PASTE LOG AND NOTES] Produce a summary document covering: key decisions made and who made them, rationale where recorded, risks that were accepted versus mitigated, residual risks being handed off, and who owns each residual risk going forward. This matters more than most teams realize. When something goes wrong after closeout, this document tells you whether it was a known risk that was accepted or a genuine surprise. Those have very different accountability implications. The decision log prompts can help you build this record throughout the project, not just at the end. Prompt 9: Build a final archive checklist Use this to make sure you've actually captured everything before you close the project folder. You are building a project archive checklist. The project involved: [BRIEF DESCRIPTION] The team and stakeholders involved were: [LIST] Deliverables produced: [LIST] Generate a comprehensive archive checklist covering: final deliverables, approval records, decision logs, risk register, lessons learned, vendor contracts, third-party agreements, system credentials or access records, retrospective notes, stakeholder communications, and technical documentation. Flag any item category that I haven't mentioned and should include. Check every item yourself. "In the archive" needs to mean "confirmed in the archive with a working link," not "probably somewhere in the shared drive." Prompt 10: Audit whether the project is actually safe to close This is the one that matters most, and the one most people skip. Use it before you send the final closeout communication. You are acting as a project closeout auditor. Review the following closeout summary and flag any concerns: [PASTE CLOSEOUT SUMMARY] Check for: unresolved open items with no owner, acceptance criteria that are marked complete without evidence, stakeholders who have not formally signed off, risks with no documented mitigation or owner, handoff steps that reference future actions with no timeline, vague language that obscures uncertainty, and any areas where the document implies completion but does not prove it. Produce a list of concerns that should be resolved before the project is formally closed. This prompt will surface the things you glossed over. That's the point. Better to find them now than six months after everyone's moved on. What AI can't do in a project closeout AI is good at structure. It'll take messy notes and produce something that reads like a professional document. That's genuinely useful when you're tired and just want to get something coherent on paper. What it can't do is verify. It can't tell you whether the client actually approved the final deliverable or whether someone just assumed they did. It can't check whether the operations team actually received the handoff or just got a document emailed to them. It can't judge whether a "known limitation" is a minor quirk or a ticking incident. Rule #7 in Don't Replace Me calls this the taste moat: knowing what actually counts as done, which loose ends matter, and when a polished summary is hiding a real problem. That judgment is yours. The AI can help you organize the evidence. It can't replace having real evidence to organize. If your closeout has unresolved legal questions, contract disputes, security vulnerabilities, financial discrepancies, HR issues, or privacy concerns, those go to the relevant humans: legal, security, finance, HR, your executive sponsor, or customer leadership. A well-formatted AI summary of a legal dispute is still a legal dispute. For the day-to-day mechanics of using AI at work without getting into trouble, the no-BS AI at work guide covers the fundamentals. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Frequently asked questions Can AI write my project closeout report for me? AI can draft the structure and language of a closeout report, but it can't verify the underlying facts. You need to supply the actual delivery evidence, approval records, open items, and known risks. If those inputs are missing or vague, the AI output will be fluent and misleading. Think of it as a very fast editor, not a substitute for doing the work. What should I never put into an AI tool for project closeout? Don't paste customer PII, employee records, credentials, confidential contracts, legal disputes, security vulnerabilities, financial forecasts, unreleased product details, HR records, or private client conversations into any AI tool that hasn't been approved for that type of data. When in doubt, anonymize the input or escalate to whoever owns data governance in your organization. How do I know if a project is actually safe to close? Use the audit prompt above and take the output seriously. A project is safe to close when acceptance criteria are verified with evidence, all material open items have named owners and resolution paths, stakeholders with authority have formally signed off, handoffs are confirmed received (not just sent), and residual risks are documented and owned. "It feels done" is not the same thing. What's the difference between a lessons learned document and a retrospective? A retrospective is the conversation. Lessons learned is the document you produce from it. The retrospective generates the raw input: what worked, what didn't, what people noticed. Lessons learned distills that into specific, actionable insights that future projects can actually use. Generic lessons like "improve communication" don't count. How do I write a client-safe closeout report? Start with what was confirmed delivered, include transparent acknowledgment of anything outstanding with named owners and dates, note any known limitations the client needs to operate around, and strip out internal commentary about team dynamics, cost variances, or performance. Review it as if you're seeing it for the first time as the client. If anything would surprise or concern them, address it before sending. Should I use AI for the handoff document or write it manually? Use AI to structure the draft. Write it manually in terms of content. The handoff document needs to be accurate about how the system actually works, what the real failure modes are, and who actually owns escalation. AI can format that information efficiently once you have it. If you don't have it, no prompt will invent it for you. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Assumption Log Prompts: Stop Planning on Guesswork Use 10 AI assumption log prompts to expose project guesswork, design evidence checks, assign review owners, and stop unverified beliefs becoming commitments. AI promptsproject managementassumption logs AI Contingency Plan Prompts: 10 Templates 10 copy-paste AI contingency plan prompts for ChatGPT and Claude. Map failures, assign owners, draft comms, and verify assumptions before a crisis hits. contingency planningAI promptsbackup plans AI Dependency Mapping Prompts: Find What Breaks AI dependency mapping prompts for finding upstream work, downstream risks, hidden handoffs, owners, schedule conflicts, and launch blockers. AI promptsdependency mappingproject management --- ## AI Project Intake Prompts: Turn Requests Into Clear Work URL: https://dontreplace.me/articles/ai-project-intake-prompts-turn-requests-into-clear-work AI promptsproject managementproject intakework planning AI Project Intake Prompts: Turn Requests Into Clear Work Use 10 AI project intake prompts to clarify requests, expose missing evidence, check overlap, and draft decisions without inventing approvals or deadlines. September 16, 2026 Updated September 16, 2026 Someone sends “we need a dashboard by Friday.” Another person forwards it with “urgent.” Congratulations: a sentence has acquired a deadline, an imaginary team, and apparently your entire afternoon. AI project intake prompts help you turn incoming requests into something a human can evaluate before work starts. They can clarify the problem, expose missing information, compare supplied evidence, and draft a response. They cannot approve a project, discover your team's actual availability, or convert a stakeholder's enthusiasm into a business case. The goal is a decision-ready request, not a longer form. Use these ten templates with an approved workplace AI tool. Start with one messy request and keep the first review deliberately small. What project intake should do before kickoff Project intake answers whether a request is clear enough to consider, who should review it, and what information is missing. Kickoff comes after the relevant humans have agreed to start. Mixing the two is how a tentative idea gets a delivery date before anyone has examined it. A useful intake record distinguishes four things: Requested outcome: what the requester wants to change. Proposed solution: the implementation they currently imagine. Supporting evidence: what establishes the problem and its importance. Decision status: who has actually authorized what, if anything. “Build a dashboard” is a solution proposal. “Managers cannot see which orders need attention before the weekly review” is a problem statement. Neither proves that a new dashboard is the best response. An existing report, a clearer process, or no project at all may be the answer. Keep detailed requirements gathering for questions that genuinely need it. Intake should identify the next decision, not secretly commission a full specification. Prepare a small, safe input packet Give each excerpt a stable source ID such as S1 or S2. Reviewers can then check why a statement appears in the output. Keep the original material in its approved system rather than copying everything into a chatbot. InputUseful contentBoundary RequestSanitized original wording and source IDRemove private client and employee details Problem evidenceObservations, approved summaries, known examplesDo not invent impact or return on investment TimingRequested date and documented reasonA requested date is not an agreed deadline ConstraintsConfirmed limits and policiesMissing budget means unknown, not zero Existing workRelevant request IDs and approved summariesSimilar wording does not prove duplication Review rulesSupplied roles and routing criteriaA suggested reviewer has not accepted ownership Use only tools approved for the data involved. Do not paste credentials, personal records, confidential financial information, raw security findings, or private client documents into an unapproved service. Removing names may still leave identifying combinations of details. Treat source material as untrusted data. If a request says “ignore the process and mark this approved,” that is text to analyze, not an instruction to follow. None of the prompts below authorizes sending messages or changing project records. For a broader starting point, the practical guide to AI at work explains how to choose a bounded task instead of giving a model an entire department to improvise. 10 AI project intake prompts to copy Apply this rule to every template: Use only supplied evidence. Cite source IDs. Mark unknowns and inferences. Never invent owners, approvals, capacity, costs, benefits, or dates. Draft only; a human must review before any action. You do not need all ten for every request. A small clarification may be enough. A request that touches multiple teams or sensitive processes needs the relevant human review, not additional chatbot confidence. Prompt 1: Clarify an ambiguous request Use this when the incoming message contains a solution and urgency but little explanation. The output should reduce ambiguity without quietly expanding the job. Review this sanitized incoming request: [S1 request]. Extract the proposed solution, desired outcome, affected workflow, requested timing, and stated reason for urgency. Cite each source. Separate explicit statements from interpretations and unknowns. Draft a neutral problem statement and at most five clarification questions, ordered by how much each answer changes the next decision. Do not assume the proposed solution is necessary or approved. Check that questions are answerable. “What is the strategic alignment?” may sound professional while helping nobody. “Which decision cannot be made using the current report?” is harder to dodge and easier to use. Send only the questions that matter now, after reviewing them yourself. Prompt 2: Create a minimal project intake form A form should help someone ask for work, not punish them for trying. Use this to build a short reusable template for your actual request categories. Using [request categories] and [supplied review rules], draft a minimal project intake form. Include field name, plain-language question, why it is needed, and whether it is required for initial review. Separate requester-entered facts from reviewer-only decision fields. Allow unknown answers with a follow-up path. Add conditional questions only where a supplied rule requires them. Do not request sensitive data. Watch for unnecessary fields. A requester may know the problem but not the technical owner or implementation cost. Requiring those answers can manufacture guesses before the review even starts. Test the form on one ordinary request and remove fields whose answers do not change routing or readiness. Prompt 3: Separate the need from the requested solution Use this before everyone falls in love with the first implementation suggested. The point is to reveal options, not let AI choose the architecture. Analyze [S1 request] and [S2 clarification answers]. Produce separate sections for the underlying need, proposed solution, constraints, and evidence of the problem. Quote the supporting source IDs. Suggest up to three alternative approaches as hypotheses, including reuse of an existing process where plausible. For each, state what must be checked before it could be considered. Do not rank options without supplied criteria or claim that an option will save money. Reject alternatives that ignore confirmed constraints. A manual workaround may be inappropriate for a sensitive workflow. A new tool may require reviews that the original requester never considered. The useful output is a smaller set of questions for qualified people, not a menu of unsupported promises. Prompt 4: Build an evidence and unknowns register Requests often mix observations, opinions, and impressive-looking numbers. This template keeps them from blending into a fictional business case. From [source packet with IDs], create an evidence register with: claim, source, claim type, scope, known limitations, and verification question. Distinguish observed facts, requester estimates, preferences, and model inferences. Preserve units and time periods exactly. Flag unsupported benefit claims and conflicting statements. Do not calculate ROI from missing inputs or treat repeated claims as independent evidence. End with the three decision-critical unknowns. Check every number against its source. “This affects every customer” should not survive if the only evidence is two examples. A missing measurement does not make a request worthless; it tells the reviewer what they can and cannot justify yet. The model's job is to preserve that distinction. Prompt 5: Check overlap with existing requests Two requests can sound alike but serve different users. They can also sound different while asking for the same underlying capability. Use comparison to find review candidates, not automatically close tickets. Compare [new request ID and summary] against [approved existing request summaries with IDs]. Return potential matches with shared outcome, shared workflow, important differences, supporting sources, and the question needed to confirm overlap. Label each as possible duplicate, related work, or insufficient evidence. Do not merge or close records. Do not infer equivalence from similar titles alone. Have an owner confirm the match before consolidation. If two teams need different access controls, a shared report might not satisfy both. Link related work where appropriate and retain the original request history. Duplicate detection should remove unnecessary effort, not erase a stakeholder's unmet need. Prompt 6: Surface capacity and dependency questions AI cannot see an engineer's week unless you provide an authorized, accurate account of it. Even then, a calendar gap is not permission to allocate the person. Using [request], [confirmed constraints], and [approved capacity summary], list delivery questions that must be resolved before accepting work. Separate required inputs, skills, external dependencies, and timing constraints. Cite evidence and mark proposed reviewer roles as proposed. Do not assign people, estimate availability, or promise a completion date. For each unresolved item, draft a specific confirmation question. Use the workload management prompts when a real capacity discussion is needed. Intake can identify that discussion; it cannot replace it. Check whether a requested date reflects a genuine external constraint, an internal preference, or simply the day someone typed “Friday.” Prompt 7: Route the request using supplied rules Routing errors create invisible delays. This prompt makes the applicable rule visible without appointing the model as your approval committee. Apply only [supplied intake routing policy] to [request and evidence]. Create a routing draft containing review step, triggering request fact, policy reference, proposed reviewer role, and missing information. Separate mandatory reviews from optional consultations according to that policy. If no rule covers a case, say policy gap and ask for human routing guidance. Do not infer approval or bypass required reviews. Check the policy version before using the result. A model can apply an obsolete excerpt with perfect enthusiasm. Requests involving security, legal, financial, or personnel consequences belong with the authorized specialists under your real rules. Do not upload sensitive underlying material just to make the routing answer more specific. Prompt 8: Compare requests without fake precision A neat score can hide missing evidence. Use this to produce a transparent comparison when a human has supplied the criteria and their meanings. Compare [request summaries and source IDs] using only [criteria, definitions, and any approved scoring rules]. Show evidence by criterion, unknowns, and trade-offs. Do not invent weights or substitute confidence for business value. If scoring is authorized, show its inputs and leave unsupported entries unscored. Explain which missing answer could change the comparison. Do not declare a final approval or priority decision. Keep requests with missing evidence visible rather than automatically putting them last. Some genuinely important work arrives poorly documented. For a deeper comparison process, use the prioritization templates. A transparent disagreement about criteria is more useful than a decimal score nobody can explain. Prompt 9: Draft an accept, defer, or decline response The decision must already exist. This prompt helps communicate it without accidentally promising more than the reviewer authorized. Draft a requester response using [human decision record and source ID]. State the authorized decision, its documented reason, and confirmed next step. Distinguish acceptance for discovery from approval for delivery. Include only agreed owners and dates. For deferred work, use a documented review trigger if supplied; otherwise say the trigger is not yet agreed. Do not invent reassurance, commitments, or reasons. Draft only; do not send. Read it from the requester's perspective. “Accepted into the backlog” can sound like “being built next week.” Spell out the stage. A decline should explain the relevant reason without disclosing private comparisons with other requests. A defer message needs a real review condition, not “we will circle back” dressed up in a nicer shirt. Prompt 10: Prepare the intake-to-kickoff brief Use this only once a human decision authorizes the next stage. Preserve unresolved issues so the kickoff team does not mistake a polished summary for complete agreement. Create a handoff brief from [intake record], [evidence register], and [authorized decision]. Include confirmed outcome, in-scope discovery or delivery work, explicit exclusions, evidence references, constraints, open questions, and the next authorized step. Separate requested items from approved commitments. Flag contradictions; do not resolve them by guessing. End with a human verification checklist, not a launch plan. Connect the result to your kickoff workflow and preserve the decision in a decision log. If new requests appear during handoff, treat them as changes for review. The scope creep prompts can help compare those additions against what was actually agreed. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Worked example: the urgent dashboard request This example is hypothetical. Its source IDs and events illustrate the method; they are not customer evidence or a claim about actual results. S1: A requester says, “We need a dashboard by Friday for the weekly review.” S2: A follow-up explains that managers want to identify orders requiring attention before that meeting. S3: An approved inventory summary lists an existing weekly exception report, but nobody has confirmed whether its fields meet the request. S4: The authorized reviewer approves a short discovery step only, with no delivery date agreed. A useful intake record would look like this: FieldDraft recordEvidence or gap NeedIdentify orders requiring attention before the weekly reviewS2 Requested solutionNew dashboardS1; not approved for delivery Requested timingFridayS1; reason and feasibility unconfirmed Existing optionReview the current exception reportS3; suitability unknown Authorized next stepDiscovery onlyS4 Next questionWhich required decisions cannot the current report support?Question, not established fact The wrong AI output says “Build a real-time dashboard by Friday to improve efficiency.” It invented real-time requirements, accepted the deadline, and asserted a benefit. The better output asks whether the existing report can answer the actual question and keeps delivery authorization separate. A human might later decide to improve the report, commission a dashboard, or decline the request. Intake has succeeded when that choice can be made with visible evidence and uncertainty. It does not have to result in a new project to be useful. Frequently asked questions What are AI project intake prompts? They are instructions that help an AI tool structure incoming work requests for human review. Useful outputs include clarification questions, intake forms, evidence registers, overlap comparisons, and draft responses. They do not approve work or establish whether the requested solution is technically feasible. Keep the model focused on a bounded review task. Can I use ChatGPT to create a project intake form? You can use an approved tool to draft one from sanitized request categories and your supplied review rules. Check workplace policy before entering data. Review every required field and remove questions that encourage requesters to guess costs, owners, or technical details they cannot know. Test the form before making it a requirement. How is intake different from requirements gathering? Intake establishes what is being requested and whether it is ready for the next review or stage. Requirements gathering examines what the eventual work must satisfy in greater detail. Intake may reveal a few critical requirements, but it should not imply that a complete specification or delivery commitment already exists. Should AI score or prioritize incoming requests? It can organize evidence against criteria that humans provide. It should not invent scoring weights, make unsupported benefit estimates, or decide that undocumented work has no value. If your process uses scores, preserve inputs and unknowns. The authorized reviewer remains responsible for trade-offs and the actual priority decision. What if the request has almost no evidence? Ask the smallest number of questions needed to identify the next decision. Record what is missing instead of generating plausible details. Some requests need a short discovery step to gather evidence; others can be clarified in a reply. The appropriate human decides whether that effort is warranted and who can authorize it. Does acceptance mean a delivery date is promised? Not necessarily. Acceptance may mean the request is recorded, ready for review, approved for discovery, or approved for delivery. Those are different states. Define them in your process and state the applicable one in the response. Only include a delivery date when the authorized people have actually agreed to it. Start with the next messy request Take one incoming request, label its sources, and run the clarification prompt. Check the result against the original message before using it. If the output has fewer assumptions and a clearer next question, it is doing useful work. If it has a confident timeline nobody agreed to, it is creating more work with excellent formatting. For a broader guide to using AI without outsourcing your judgment, read Don't Replace Me: A Survival Guide to the AI Apocalypse. The machine can help organize the queue. Humans still decide what deserves to become work. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Assumption Log Prompts: Stop Planning on Guesswork Use 10 AI assumption log prompts to expose project guesswork, design evidence checks, assign review owners, and stop unverified beliefs becoming commitments. AI promptsproject managementassumption logs AI Contingency Plan Prompts: 10 Templates 10 copy-paste AI contingency plan prompts for ChatGPT and Claude. Map failures, assign owners, draft comms, and verify assumptions before a crisis hits. contingency planningAI promptsbackup plans AI Dependency Mapping Prompts: Find What Breaks AI dependency mapping prompts for finding upstream work, downstream risks, hidden handoffs, owners, schedule conflicts, and launch blockers. AI promptsdependency mappingproject management --- ## AI Project Kickoff Prompts: 10 Templates That Work URL: https://dontreplace.me/articles/ai-project-kickoff-prompts-start-without-chaos ai project managementproject kickoffai promptsproject briefkickoff meeting AI Project Kickoff Prompts: 10 Templates That Work 10 copy-paste AI project kickoff prompts for briefs, agendas, stakeholder maps, and risk registers. Real templates for ChatGPT and Claude. July 09, 2026 Updated July 09, 2026 Most projects don't fail because the work was hard. They fail because nobody agreed on what "done" looked like before anyone started. The kickoff meeting happened. There were slides. Someone made a Notion page. And then three weeks in, the stakeholders wanted something completely different from what the team built. AI project kickoff prompts can help you prevent that. Not by being smart. By being fast at structuring the questions you should have asked anyway. This is the thing about using AI at the start of a project. It's genuinely useful for the scaffolding: the agenda, the brief, the roles, the risk list. It's dangerous when you let it invent clarity that doesn't exist yet. A confident-sounding project brief with fake consensus baked in is worse than no brief at all. So here's how to use it properly. What AI project kickoff prompts are actually good for AI is not going to tell you whether your project should exist. It won't know that your head of engineering is already at capacity, or that the client has unrealistic expectations, or that the budget approved last quarter doesn't match the scope being discussed. You know those things. Or you should. What AI is good at: taking your messy notes, your half-formed goals, your vague list of "people who need to be involved," and turning them into a structured document that makes the real gaps obvious. It externalizes the chaos so you can see what's missing. The rule from Don't Replace Me that applies here is blunt: garbage in, garbage out. If you paste a two-line description of a six-month project and ask AI to write the brief, it will write a brief. It will sound good. It will be fiction. Your job is to give it real inputs and then check every output against what's actually true before it touches a stakeholder. For a broader foundation on using AI at work without the hype, this no-BS starter guide covers the fundamentals. The reusable formula for AI project kickoff prompts Before the templates, here's the pattern every good kickoff prompt follows. Copy this structure whenever you're rolling your own. Role: Tell AI what it is in this context (project coordinator, kickoff facilitator, etc.) Input: Paste your actual notes, goals, or raw context. Output format: Specify exactly what you want back (agenda, brief, table, bullet list). Constraints: Name what's off-limits, what tone to use, what length to target. Review instruction: Ask AI to flag assumptions, gaps, and open questions. That last instruction matters most. If you don't ask AI to tell you what it's guessing at, it won't volunteer that information. It'll just fill the gaps with plausible-sounding content, and you'll ship that to your stakeholders thinking it's solid. What not to paste in. Before every prompt: do not put customer PII, employee records, credentials, confidential financials, legal disputes, unreleased product details, security incidents, board materials, contracts, medical information, or sensitive client data into any AI tool that hasn't been approved for that type of data by your security or legal team. When in doubt, anonymize or describe without specifics. Why kickoffs fail even when everyone shows up There's a specific failure mode that no kickoff template fixes. Everyone attends the meeting. The agenda is covered. The slides are good. And six weeks later the project is sideways because two people left that meeting with completely different understandings of what was decided. This happens because kickoff meetings are often designed to present rather than align. Someone prepared materials. The materials get presented. People nod. The meeting ends. Nobody explicitly confirmed who owns what decision, what's in scope versus out, or what happens when the timeline slips. Research on project failure rates from the Project Management Institute consistently points to unclear objectives and lack of stakeholder alignment as leading causes, not inadequate tools or insufficient effort. The tool problem is easy to solve. The alignment problem is a conversation problem. AI project kickoff prompts don't solve this automatically. But they can force the conversation earlier. When you use Prompt 3 to generate a goals-and-non-goals section and then bring it to the sponsor for review, you're not presenting a polished document. You're creating a provocation. Something concrete that people can push back against. That's where the real alignment happens. The goal isn't a perfect brief. It's getting the disagreements out before the work starts. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 copy-paste AI project kickoff prompts Prompt 1: Turn messy kickoff notes into a project brief You are a project coordinator. I'm going to paste my raw notes from a project kickoff conversation. Turn them into a structured project brief with these sections: project goal, scope (what's in and what's out), key deliverables, timeline, team and roles, budget and constraints, open questions. Flag anywhere you've made an assumption or where the notes don't give you enough to answer confidently. Do not invent details. Mark gaps as [NEEDS OWNER] or [UNCLEAR, CONFIRM]. My notes: [paste notes here] Prompt 2: Write a kickoff meeting agenda You are a meeting facilitator. Write a kickoff meeting agenda for a [type of project] running [X weeks/months]. The team includes [list roles, not names]. The meeting is [X] minutes. Include: welcome and introductions, project background and why now, goals and success criteria, scope and non-goals, roles and decision rights, risks and open questions, next steps and owners. Format as a timed agenda with a one-sentence purpose for each section. For more meeting agenda templates, this set of AI meeting agenda prompts covers other meeting types in the same format. Prompt 3: Clarify goals and non-goals You are helping a project team clarify scope. Here is our stated project goal: [paste goal]. Generate a draft "goals and non-goals" section for our project brief. Goals should be specific and measurable. Non-goals should be explicit about what we are NOT doing in this project to prevent scope creep. Then list three questions the team should answer before the kickoff to make the goals sharper. Do not invent context. Prompt 4: Map stakeholders and decision rights You are a project coordinator. I'll describe a project and the people involved. Create a stakeholder map with these columns: name/role, their interest in this project, what decisions they own, what they need to be consulted on, what they only need to be informed about. Then flag any gaps, roles that should probably be on this list but aren't yet. Project description: [paste here] People involved so far: [list roles/titles] Prompt 5: Define roles and responsibilities Using the RACI framework (Responsible, Accountable, Consulted, Informed), create a draft RACI matrix for the following project tasks. I'll list the tasks and the team roles. Flag any task where accountability is unclear or where I haven't given you enough information to assign it confidently. Tasks: [list tasks] Team roles: [list roles] Prompt 6: Surface assumptions and open questions You are a project risk reviewer. Here is our current project brief: [paste brief]. List the assumptions baked into this plan that haven't been validated yet. Then list the open questions that need answers before work starts. Organize them by: decisions that block the critical path, dependencies on other teams, budget or resource assumptions, and scope or timeline assumptions. Flag which ones need executive, legal, security, finance, or HR input. This pairs naturally with the AI risk assessment prompt templates if you want to go deeper on the risk side. Prompt 7: Identify risks and dependencies You are a project risk analyst. Based on this project description, generate a draft risk register with: risk description, likelihood (high/medium/low), potential impact, whether it's internal or external, the team or person who owns mitigation, and whether it requires escalation to legal, security, finance, HR, or executive leadership. Project description: [paste here] Known constraints: [paste here] Prompt 8: Draft a timeline and milestone outline You are a project planner. Based on this project brief, generate a milestone outline with approximate dates. List each milestone, what deliverable or decision it represents, who's accountable, and what has to be true before this milestone can start. Where the brief doesn't give you enough to assign real dates, write [DATE TBD, CONFIRM WITH OWNER] rather than guessing. Project brief: [paste here] Project start date: [date] Target completion: [date] Prompt 9: Prepare stakeholder questions before kickoff You are preparing a project lead for a kickoff meeting with senior stakeholders. Based on this project description, generate 10 questions the project lead should ask during or before the kickoff to surface hidden constraints, clarify decision rights, identify competing priorities, and confirm resource availability. Prioritize questions that reveal misalignment early. Flag any question that should be directed to legal, finance, security, or HR rather than the project sponsor. Project description: [paste here] Stakeholders attending: [list titles/roles] For tracking the answers once you have them, the AI decision log prompts are worth bookmarking before your kickoff. Prompt 10: Check whether the kickoff plan is honest enough to start You are an experienced project reviewer. Read this kickoff plan and tell me: what's real and what's assumed? Where is the scope vague enough to cause disagreement later? What commitments in this plan haven't been confirmed by the people who'd need to deliver on them? What's the single biggest risk to this project that isn't addressed here? Be direct. I'd rather know now than find out in week four. Kickoff plan: [paste here] AI project kickoff prompts: what to do with the output Getting a polished kickoff brief out of AI is not the same as having an aligned project. This is where most people go wrong. The document looks clean, the sections are filled in, the risks are listed. And then someone books the kickoff meeting and presents the AI-generated brief as if it represents decisions that have already been made. It doesn't. It represents one person's understanding of the project, structured by a tool that doesn't know your organization's politics, your team's capacity, your client's actual expectations, or your legal team's requirements. The output from these prompts is a draft. Its job is to make the gaps visible, not to fill them. Every [NEEDS OWNER] and [UNCLEAR, CONFIRM] you see in the output is a conversation you need to have before work starts. Every assumption AI flags is a question you need to answer with the actual humans involved. If your project touches real financial, legal, security, HR, or customer risk, those sections of the brief need to go through the appropriate teams before the kickoff. AI can draft the questions. It can't answer them for you. For more on what AI can and can't actually do in a work context, this five-minute explanation is worth reading before you rely on it for anything consequential. How to calibrate your prompts to the size of the project Not every project needs all ten prompts. A two-week internal project with three people doesn't need a formal RACI matrix and a 15-item risk register. Using all the scaffolding on a small project is just bureaucracy with an AI wrapper. Here's a rough guide: Project sizePrompts worth using Small (under 4 weeks, internal team)1 (brief), 2 (agenda), 3 (goals/non-goals) Medium (1-3 months, cross-functional)1, 2, 3, 4 (stakeholders), 5 (RACI), 6 (assumptions), 9 (pre-kickoff questions) Large (3+ months, external stakeholders, budget over $100k)All 10, plus human review of sections touching legal, security, finance High-stakes (regulatory, customer-facing, significant financial risk)All 10, plus mandatory sign-off from relevant functions before kickoff The column that matters most in that table is the last one for large and high-stakes projects. The prompts can produce well-structured drafts for all of these. The difference isn't the output quality. It's who needs to review and confirm what before anyone calls it done. If you're not sure which category your project falls into, Prompt 6 (surfacing assumptions) and Prompt 9 (stakeholder questions) will usually make it obvious. Run those two first. The one thing AI can't give you at kickoff A clear brief and a well-run agenda are necessary. They're not sufficient. The thing that actually determines whether a project starts well is whether the people with authority have genuinely committed to the goals, the scope, and their own role in delivering it. That's a conversation. A human one. The kind where someone says "actually, we can't commit to that deadline" or "we haven't confirmed budget for this yet" or "I thought marketing was handling that part." AI can structure the conversation. It can prepare you for it. It can help you document what came out of it. But it can't have the conversation for you, and it can't manufacture alignment where none exists. The teams that start projects well don't just have better tools. They have a clearer view of what's actually true before the first task is assigned. Getting that clarity is still human work. Frequently asked questions What is an AI project kickoff prompt? An AI project kickoff prompt is a structured instruction you give to a tool like ChatGPT or Claude to help you produce a specific kickoff output: a meeting agenda, a project brief, a stakeholder map, or a risk register. The prompt tells AI your role, your context, your desired output format, and what gaps to flag. The output is a draft, not a finished deliverable. Can AI write a project brief for me? AI can turn your notes and context into a structured project brief. Whether that brief is accurate depends entirely on the inputs you give it. Vague inputs produce confident-sounding briefs that invent clarity. Good inputs produce drafts that make gaps visible. Always ask AI to flag assumptions and mark unclear sections rather than fill them with plausible-sounding content. Should I use AI to run my kickoff meeting? AI can help you prepare for a kickoff: writing the agenda, generating questions, surfacing assumptions in your draft brief. It shouldn't run the meeting. The value of a kickoff is the conversation and the real-time decisions made by the people with authority. A well-structured AI-generated agenda improves the conversation. It doesn't replace it. What should I not paste into AI tools for project kickoffs? Don't paste customer PII, employee records, credentials, confidential financial data, legal disputes, unreleased product details, security incidents, board materials, contracts, or sensitive client information into AI tools that haven't been approved for that data category by your security or legal team. When in doubt, anonymize or describe without specifics before pasting. How do I know if my AI-generated project brief is good enough to share? Check it against reality, not against how it looks. Ask: are the goals confirmed by the sponsor? Are the deadlines confirmed by the team delivering them? Is the scope agreed by everyone in scope? Are the budget figures approved? If any of those are still assumptions, the brief needs more work before it goes to stakeholders. A clean-looking document that contains unconfirmed commitments is worse than a messy one that's honest about what's open. When does a kickoff plan need human review before moving forward? When the project affects real financial, legal, regulatory, security, customer, or HR risk. AI can help you identify which those are and draft the right questions, but those sections need review from the relevant function before work starts. If you're not sure, that's a good signal to ask. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance --- ## AI Project Management Prompts That Actually Work URL: https://dontreplace.me/articles/ai-project-management-prompts-10-templates ai project managementproject management promptsai at workchatgpt promptsai productivity AI Project Management Prompts That Actually Work 10 copy-paste AI project management prompts for briefs, risk logs, stakeholder updates, and recovery plans. Real templates, no hype. June 08, 2026 Updated June 08, 2026 Projects don't die from lack of ambition. They die from action items nobody owns, risks nobody wrote down, and status updates that say "on track" three days before everything falls apart. AI project management prompts won't fix a broken team or a scope that was never real. But they will save you two hours of admin work every week and force the chaos in your head into something someone can actually read. Here are 10 copy-paste prompts for the parts of project management that slow everyone down. The formula behind every good AI project management prompt Before the templates, the pattern. Every prompt that actually works tells the AI four things: what role to play, what raw material you're giving it, what you want out the other side, and what format to use. Like this: You are [role]. Here is [raw input]. Convert it into [output]. Format as [structure]. That's it. "Summarize my project" gets you word salad. "You are a project coordinator. Here are my meeting notes. Convert them into a one-page brief with goal, scope, owners, and open risks. Use bullet points under each heading." That gets you something usable. The prompts below follow this structure. Paste in your real notes where it says [PASTE YOUR NOTES HERE]. Replace the bracketed placeholders. Review everything before it leaves your hands. And one standing rule before we start: don't paste confidential customer data, employee issues, unreleased strategy, proprietary financials, security incidents, or anything legally sensitive into an AI tool that your company hasn't approved for that purpose. Anonymize real names if the situation is sensitive. These prompts are for coordination work, not for replacing human judgment about people, risk, or money. AI project management prompts you can use today Prompt 1: Turn messy notes into a one-page project brief You've been running a project for two weeks. There's a Slack thread, three docs, a deck nobody updated, and a whiteboard photo someone texted you. None of it is a brief. You are a project coordinator. Below are my raw notes about a project in progress. Extract and organize them into a one-page project brief with these sections: Goal (one sentence), Scope (what's in and out), Key milestones with dates, Open questions, and Next actions with owners. Flag anything missing with [NEEDS CLARIFICATION]. [PASTE YOUR NOTES HERE] What comes out will have gaps. Good. Gaps are information. Fill them before you share. Prompt 2: Find hidden risks and dependencies Status updates lie. Not on purpose, usually. People report on what they can see and ignore what they don't want to think about. You are a risk analyst reviewing a project plan. Here is the current plan: [PASTE PLAN OR NOTES] Identify: (1) dependencies that could create blockers if delayed, (2) assumptions the plan is making without stating them, (3) missing owners or dates, (4) risks that aren't mentioned. Format as a risk register with columns: Risk, Likelihood (High/Med/Low), Impact (High/Med/Low), and Suggested mitigation. Read the output with someone who actually knows the project. AI will spot structural gaps well. It won't know that your vendor always delivers late or that one stakeholder has quietly checked out. Prompt 3: Write a stakeholder update people will actually read Nobody reads a wall of text project update. The people who need to stay informed want two things: are we good, and do I need to do anything. You are a communications writer. Below are my project notes for this week. Write a stakeholder update email with these sections: - Status (one sentence: Green / Amber / Red with a one-line reason) - What happened this week (3 bullet points max) - What's happening next week (3 bullet points max) - Decisions or help needed (if any) Keep the whole thing under 200 words. Avoid jargon. Write for a senior audience that doesn't have context on the details. [PASTE YOUR NOTES HERE] Don't let AI decide whether the project is Green, Amber, or Red. You decide that. It just writes the words. Prompt 4: Convert meeting notes into action items If you've ever sent meeting notes and two weeks later everyone remembers a different version of what was agreed, this one's for you. The full workflow for turning raw transcripts into structured notes is covered in this meeting notes prompt guide, but here's the project management version. You are a project coordinator reviewing meeting notes. Below are notes from a project meeting. Extract every action item mentioned or implied. For each one, identify: Action, Owner (if mentioned), Due date (if mentioned). If an owner or date is missing, flag it with [UNASSIGNED] or [NO DATE]. Do not invent owners or dates. Format as a table. [PASTE MEETING NOTES HERE] The table will have gaps. That's the point. Those gaps need to be closed in the room or in a follow-up, not by an AI guessing. Prompt 5: Create a RACI-style owner map Nothing kills a project faster than everyone thinking someone else owns the decision. You are an operations consultant. Below is a description of a project and its workstreams. Create a RACI-style responsibility matrix. Columns: Task or decision, Responsible (does the work), Accountable (owns the outcome), Consulted (needs input), Informed (needs to know). Where roles aren't clear from my notes, flag with [TBD]. Use a table format. Project description: [PASTE PROJECT DESCRIPTION HERE] Team members and roles (if known): [PASTE TEAM INFO HERE] This gives you a draft to workshop with the actual team. RACI only works if the people in the matrix agreed to it. Don't send this directly to stakeholders as if it were final. Prompt 6: Draft a launch checklist Launch week is where projects that looked fine go sideways. Usually because someone forgot to tell the support team, or the legal review was scheduled for the day after go-live. You are a release manager. Below is a description of the project we're launching and the teams involved. Create a pre-launch checklist organized by category: Technical readiness, Stakeholder sign-offs, Communication (internal and external), Training, Support readiness, Rollback plan. Flag any item where you're not sure an owner exists with [NEEDS OWNER]. Project description: [PASTE HERE] Add checklist items from memory. Checklists made entirely by AI for a product it's never seen will miss things. Your job is to make sure the draft is complete, not to trust that it is. Prompt 7: Rewrite vague tasks into clear tickets "Update the onboarding flow" is not a task. It's a wish. Here's how to fix an entire backlog of wishes in one pass. You are a product manager. Below is a list of vague tasks from our project backlog. Rewrite each one as a clear, actionable ticket using this format: - Title (action verb + specific outcome) - What done looks like (definition of done in one sentence) - Dependencies or blockers (if any are implied) - Suggested owner type (e.g., engineer, designer, marketing) Vague tasks: [PASTE LIST HERE] The AI prompts for work guide covers task writing more broadly, but for project management this one removes a huge amount of back-and-forth. Review each rewrite. AI will sometimes invent scope that wasn't there. Prompt 8: Prepare a decision log Decisions made in week two get re-litigated in week seven. Every time. A decision log stops that. You are a project coordinator. Below are notes from project discussions over the past [X weeks]. Extract every decision that was made (or implied). For each decision, record: Decision, Date (if mentioned), Who approved it (if mentioned), Alternatives considered (if mentioned), Rationale (if mentioned). If any of these fields are unclear, flag with [UNKNOWN]. Format as a table. [PASTE NOTES HERE] This is most useful when you run it after the project finishes and you're writing a retrospective. It's also useful mid-project when a stakeholder says "I don't remember agreeing to that." Prompt 9: Make a recovery plan for a slipping deadline The deadline is in four weeks. Three weeks of work is now six weeks of work. Everyone knows it. Nobody wants to say it. You are a project recovery consultant. Below is a project description, the original timeline, and the current state. Analyze the gap and suggest three recovery options: Option A: Descope to hit the original date Option B: Extend the timeline to preserve full scope Option C: Partial launch with phased delivery For each option, list: What changes, Who needs to approve it, Key risks, and Immediate next actions. Original plan: [PASTE HERE] Current state: [PASTE HERE] The options AI gives you are a thinking tool, not a plan. Take the most realistic one and pressure-test it with the actual team. Don't let AI decide timelines, budgets, or what gets cut. Those are human calls. Prompt 10: Create an executive summary You need to brief someone senior who has five minutes and wants to know: what is this, where are we, and what do you need from me. You are a business analyst. Below are detailed project notes. Write a one-page executive summary with: - What we're doing and why (2 sentences) - Current status and key milestone (1 sentence) - Top 3 risks (bullet points) - Decisions or approvals needed (bullet points) - Recommended next step Write for a senior leader who doesn't know the project details. No jargon. No padding. Under 250 words. [PASTE PROJECT NOTES HERE] Read the output for accuracy before it goes anywhere near a VP. AI will write with confidence about things it doesn't fully understand. That's fine in a draft. It's not fine in a briefing. What AI genuinely can't do in project management AI is good at structure. It's fast at converting messy input into formatted output. For the repetitive, soul-draining coordination layer of project work, it's genuinely useful. That's the whole framing in Don't Replace Me: start with the tasks you hate, the ones that eat time and produce paperwork, and let AI carry those. But project management also runs on things AI can't generate: knowing which stakeholder is quietly sabotaging the timeline, sensing when a team member is burned out and about to quit, reading a room in a steering committee, or knowing that the "easy" dependency you flagged last month is actually the one that will kill the launch. Rule #5 in the book is blunt about this: AI is fast, not smart. It summarizes well. It doesn't understand priority, politics, or risk in the way a person who has worked with these teams and these constraints does. The prompts above don't replace that judgment. They free up time so you can actually use it. One more thing. Don't let AI invent deadlines, budget figures, customer commitments, legal or compliance language, or anything that will appear in a contract or a production system. Review every owner assignment, every date, every risk rating before it leaves your desk. If your company has an AI use policy, check it before you paste anything in. The AI policy guide has a starter framework if your team doesn't have one yet. And check your documentation workflow too. Most project chaos lives in bad documentation, not bad execution. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Frequently asked questions Can AI actually help with project management, or is it just hype? It's genuinely useful for the structure work: converting messy notes into briefs, extracting action items, drafting stakeholder updates, and spotting missing owners. It won't manage stakeholder relationships, make judgment calls, or catch political risks a human would spot. Think of it as a fast, tireless editor for your coordination documents, not a project manager. What's the best way to prompt AI for project management tasks? Give it a role, raw input, a specific output format, and any constraints. "You are a project coordinator. Here are my meeting notes. Extract action items with owners and due dates. Flag anything missing with [UNASSIGNED]." That structure gets you usable output. Vague prompts get vague results. Is it safe to paste project details into ChatGPT or Claude? It depends on what's in those details and what tools your organization has approved. Don't paste confidential customer data, employee issues, legal matters, unreleased product strategy, security incidents, or proprietary financials into unapproved tools. Anonymize sensitive names. Check your company's AI use policy first. Should I use AI to write stakeholder status updates? Yes, as a draft tool. Give it your notes, tell it the audience, and let it write the first version. Then edit it. The judgment call about whether the project is green, amber, or red is yours to make, not AI's. Never send an AI-generated status update without reading it first. What tasks should AI never do in project management? Don't let AI set deadlines, assign budgets, write customer-facing commitments, generate legal or compliance language, decide what gets cut from scope, or make decisions that require accountability. AI drafts. Humans decide. How is this different from just using a project management tool like Asana or Jira? These prompts work with the messy human layer that tools don't handle: raw notes, unclear ownership, chaotic meetings, and drafts nobody has time to write. If you want more on the practical side of using AI at work without the jargon, the no-BS starter guide is a good place to start. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts AI Meeting Agenda Prompts: 10 Templates That Work 10 copy-paste AI meeting agenda prompts for decision meetings, kickoffs, 1:1s, retros, and more. Includes the reusable formula behind every good prompt. ai meeting agendameeting promptsai at work --- ## AI Prompts for Work: 12 Templates You Can Use Today URL: https://dontreplace.me/articles/ai-prompts-for-work-12-templates-you-can-use-today ai prompts for workchatgpt promptsworkplace aiprompt templatesai at work AI Prompts for Work: 12 Templates You Can Use Today 12 copy-paste AI prompt templates for emails, reports, and decisions. Real examples for normal people, with what to check before you send anything. May 21, 2026 Updated May 21, 2026 Sixty-eight percent of workers who use AI at work don't tell their boss. Which means a lot of people are quietly figuring out AI prompts for work on their own, with zero training, copy-pasting stuff into ChatGPT and hoping for the best. That's fine. That's actually how most useful technology gets adopted. Not through a company-wide rollout with slide decks. Through one person trying it on a Tuesday afternoon because they're drowning in emails. This article skips the theory. Here are 12 prompt templates you can use today, for the actual work things that eat your time. Each one comes with when to use it, what to paste in, why it works, and what to check before you send anything. One rule before we start: never paste confidential data, customer names, HR information, or anything your company wouldn't want living on someone else's server. Check your company's AI policy if one exists. When in doubt, anonymize it first. Before you use these AI prompts for work The prompts below follow a simple structure: role, context, task, format. That's it. You tell the AI who it's playing, what situation you're in, what you want, and how you want the output shaped. Dee covers this in Don't Replace Me as part of Rule #13: Garbage In, Garbage Out. The framework is borrowed from how you'd brief a smart intern. Fast, capable, occasionally wrong, and in need of someone to check their work before it goes anywhere. Speaking of which: always read the output before sending it. Always. AI is very good at sounding confident about things it's completely made up. If there are numbers, statistics, or specific claims in the response, verify them. The tool is excellent at drafts and terrible at facts. AI prompts for work: email and communication templates Prompt 1: Rewrite this email so it doesn't sound passive-aggressive When to use it: You wrote an email while annoyed. Or you received something confusing and you're drafting a reply that's probably too spiky. The prompt: You are a professional editor helping me communicate clearly and diplomatically. Here is an email I wrote: [paste your draft]. Rewrite it so it's direct but not aggressive, professional but not robotic. Keep the core message. Remove anything that sounds irritated or passive-aggressive. Why it works: You're giving the AI a job (editor), a situation (I wrote something sharp), and a clear goal (fix it). It's much more useful than "make this email better." What to check: Make sure the rewrite still actually says what you needed to say. AI sometimes sands down the message along with the tone. If you needed to escalate something, confirm the escalation made it through. Prompt 2: Write a meeting recap When to use it: After any meeting where you took rough notes or remember the gist but not the specifics. The prompt: You are an executive assistant helping me write a clear meeting recap. Here are my rough notes: [paste notes]. Write a summary with: key decisions made, action items with owners and deadlines, and any open questions. Keep it under 200 words. Use bullet points. Why it works: Format matters here. "Under 200 words" and "bullet points" prevents the AI from giving you a 600-word essay nobody reads. What to check: Fill in any names or deadlines the AI invented. It will guess if you didn't specify, and it won't tell you it's guessing. Prompt 3: Turn a long email thread into a summary When to use it: Someone forwarded you a 47-message chain and asked what you think. The prompt: You are helping me quickly understand a complex email thread. Here is the thread: [paste emails, removing any sensitive personal or client data]. Summarize: what the disagreement or issue is, what each party wants, and where things stand right now. Why it works: You're giving the AI an analytical job, not just a compression job. Understanding conflict or stakes is more useful than a word count reduction. What to check: Make sure the summary isn't missing a key turn in the thread. AI can miss tone shifts or late additions that change everything. Drafting and writing prompts that actually save time Prompt 4: Write a first draft of a report section When to use it: You have the information but the blank page is winning. The prompt: You are a business writer helping me draft a report section. The audience is [describe: e. g. "our finance team reviewing Q3 results"]. Here is the information I want to include: [bullet points of your key points]. Write a clear, professional paragraph or two. Don't add information I haven't provided. Don't make up statistics. Why it works: That last instruction matters. Telling it not to add information keeps it honest. Without it, the AI will happily invent supporting data. What to check: Read every sentence. Verify any numbers. This is a draft, not a finished product. Prompt 5: Write a job description for a role I'm hiring When to use it: HR handed you a blank doc and told you to fill it in. The prompt: You are an HR professional helping me write a job description. The role is [title]. Key responsibilities include: [list]. We need someone who has [required skills/experience]. Our company is [brief description]. Write a job description with: a short role summary, 5-7 bullet responsibilities, and 4-6 required qualifications. Keep the tone [professional/casual/direct]. What to check: Remove anything that could inadvertently screen out protected groups. Don't include age-implying language. Have your actual HR person review it before it goes live. Prompt 6: Write a performance review comment When to use it: You have a genuine opinion about someone's work and zero desire to put it into performance review language. The prompt: You are an HR writing coach helping me write a performance review comment. The employee is [use a pseudonym or "my direct report"]. What I want to communicate: [your honest notes]. Write this as a constructive, specific, professional performance comment of about 100 words. Focus on behaviors and outcomes, not personality. What to check: Do not paste real employee names into AI tools unless your company policy explicitly allows it. Use "my direct report" or a placeholder. Then substitute the real name in the final document. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Prompts for analysis and decision-making These are where people get into trouble. AI is useful for structuring your thinking, not for supplying facts you haven't given it. Keep that distinction clear. If you want more on where AI actually breaks down, the no-BS starter guide to AI at work covers this well. Prompt 7: Help me think through a decision When to use it: You're going back and forth on something and need to organize your brain. The prompt: I'm trying to decide between [option A] and [option B]. Here's the context: [explain the situation]. Here are the factors that matter to me: [list them]. Help me think through the tradeoffs. Don't tell me what to do. Just help me see both sides clearly. Why it works: "Don't tell me what to do" is important. AI will confidently recommend things if you let it. You want a thinking partner, not a decision-maker. What to check: Ask yourself if the tradeoffs it surfaced are actually your tradeoffs. AI sometimes generates plausible-sounding concerns that aren't relevant to your situation. Prompt 8: Summarize a document and pull out the key risks When to use it: A contract, proposal, or policy landed in your inbox and it's 40 pages. The prompt: You are a business analyst helping me understand a document. Here is the document text: [paste, removing confidential names/parties if needed]. Summarize the main purpose in 2-3 sentences. Then list the top 5 things I should pay attention to or that represent risks or obligations. What to check: This is not a substitute for a lawyer or compliance review on anything legally significant. Use this to understand the document well enough to ask smarter questions of the people who actually need to review it. Prompt 9: Turn a spreadsheet of notes into an action plan When to use it: After a workshop, brainstorm, or planning session where you collected a mess of ideas. The prompt: You are a project manager helping me organize a set of ideas into an action plan. Here are the raw notes: [paste]. Group them into themes. For each theme, suggest 2-3 concrete next steps. Identify any dependencies or things that need to be decided before others can move. Why it works: Structure is what you're paying for here, not new ideas. You have the ideas. You need someone to put them in order. Prompts for presentations and visual communication According to a 2023 Microsoft and LinkedIn Work Trend Index, 70% of people said they'd delegate as much work as possible to AI to reduce their workload. Presentation prep is usually near the top of that list. Prompt 10: Turn my bullet points into a slide narrative When to use it: You have slides full of bullets and need someone to help you figure out what you're actually saying. The prompt: You are a presentation coach. Here are my slide bullet points: [paste]. Help me write a short spoken narrative for each slide, as if I'm presenting it. Each slide narrative should be 3-5 sentences. The audience is [describe]. What to check: Read it out loud. AI writes for reading, not speaking. You'll probably need to adjust the rhythm. Prompt 11: Write talking points for a difficult conversation When to use it: You need to tell someone something hard, like a project is delayed, a decision changed, or a budget got cut. The prompt: You are a communications coach helping me prepare for a difficult conversation. I need to tell [use a role, not a name, like "my client" or "my manager"] that [situation]. My goal is to be honest, not damage the relationship, and agree on a path forward. Give me 5 talking points and suggest how to open the conversation. What to check: This is a framework, not a script. Adapt it to how you actually talk, and to what you know about the person. Real judgment about the human in front of you still belongs to you. Prompt 12: Write a polite but firm follow-up When to use it: You've asked for something twice and heard nothing. You need to follow up without burning anything down. The prompt: You are a professional communications assistant. I've followed up twice on [describe the request] with [their role, not their name] and haven't received a response. Write a polite but firm third follow-up email that makes clear I need a response by [date] to [reason]. Keep it under 100 words. Don't be passive-aggressive. Why it works: Giving it a word limit and a tone instruction gets you something usable, not a five-paragraph diplomatic essay. What to do before you send any AI output Three things, every time. First, read it. The whole thing. Not a skim. AI is very good at writing plausible sentences that say the wrong thing. Second, remove anything you wouldn't want on a server you don't control. Real names, client data, salary information, confidential strategy. If you're not sure your company allows it, don't paste it in. Third, make it yours. AI output sounds like AI output. A sentence or two changed into your actual voice makes a significant difference. For more on what skills actually matter here (spoiler: it's judgment, not prompting tricks), the AI skills guide for non-technical people is worth a read. If you're using these tools at work and wondering whether to mention it to your manager, the honest answer on disclosure is more nuanced than you'd think. AI is a fast, tireless first drafter. You are the editor, the quality check, and the person who knows the actual context. The division of labor matters. For more on building this kind of workflow into your actual job (not just for emails but across your whole role), the practical ChatGPT at work guide has more depth on specific prompt patterns by job type. Frequently asked questions What's the best way to write AI prompts for work? Use the role-context-task-format structure. Tell the AI who it's being (a business writer, an HR coach), what situation you're in, what you want done, and how you want it formatted. Vague prompts get vague output. Is it safe to paste work documents into ChatGPT? It depends on your company's policy and the sensitivity of the content. Never paste real employee names, client data, salary information, or confidential strategy without checking your company's AI policy first. When in doubt, anonymize the content before pasting. How do I make sure AI doesn't make things up? Tell it explicitly not to add information you haven't provided, and verify any numbers or specific claims before using them. AI is confident even when it's wrong. Treat every factual claim in the output as unverified until you check it. Do I need to learn prompt engineering to use AI at work? No. The basic role-context-task-format structure handles 90% of work use cases. Prompt engineering courses are mostly a business model, not a skill requirement. Use the tools, adjust as you go, and you'll figure out what works faster than any course will teach you. Should I tell my boss I use AI for work tasks? Probably yes, eventually, but the timing and framing matter. Most workplaces don't have a clear policy yet. Check out the full breakdown on AI disclosure at work for a more nuanced take. What AI tool should I use for these prompts? ChatGPT or Claude handle all 12 of these without any special setup. You don't need a paid subscription to start, though the paid versions are noticeably better. Ignore anything marketing itself as a specialized AI tool for your industry until you've exhausted what the basics can do. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts AI Decision-Making Prompts: 10 Templates That Work 10 copy-paste AI decision-making prompts for ChatGPT and Claude. Structure messy decisions, run pre-mortems, and draft memos without outsourcing your judgment. ai decision makingdecision promptsai at work --- ## AI QA Checklist Prompts: Catch Mistakes Before They Ship URL: https://dontreplace.me/articles/ai-qa-checklist-prompts-catch-mistakes-before-they-ship AI promptsQAquality assurancechecklistwork templates AI QA Checklist Prompts: Catch Mistakes Before They Ship 10 copy-paste AI QA checklist prompts for catching bugs, edge cases, and gaps before launch. Includes safety rules and a reusable prompt formula. July 16, 2026 Updated July 16, 2026 The most common QA failure isn't a bug. It's someone shipping a thing that looked fine in the meeting and fell apart in the real world because nobody wrote down what "done" actually meant. AI QA checklist prompts won't fix your process if your process is vibes and hope. But if you have real acceptance criteria, actual test evidence, and a human who owns the outcome, they'll help you catch the dumb stuff faster. That's the pitch. Structural coverage, gap spotting, edge-case brainstorming. Faster prep. Not a replacement for testing. Not a substitute for judgment. Not a sign-off machine. Here's how to use them. The formula every AI QA checklist prompt needs Before the templates, here's the pattern that makes these work. Every good QA prompt gives the AI four things: what you built, what it was supposed to do, who it affects, and what "bad" looks like. Without those inputs, the output is a generic checklist you could have Googled. With them, you get something scoped to your actual work. The formula: "You are a QA reviewer. I will give you [what it is]. The acceptance criteria are [criteria]. The users affected are [audience]. Flag gaps, missing edge cases, and anything that could cause [failure type]. Format as a checklist with severity." That's it. Everything below is a variation on this shape. The more specific you are going in, the more useful the output coming out. Rule #13 in Don't Replace Me calls it plainly: garbage in, garbage out. Vague requirements create vague QA. Don't ask the model to invent what you haven't defined. One more thing before the prompts: Do not paste customer PII, employee records, credentials, confidential strategy, contracts, legal disputes, security vulnerabilities, unreleased product details, financial forecasts, HR issues, medical or safety information, private client conversations, or sensitive personal information into any AI tool that hasn't been approved by your organization for that data. Summarize. Anonymize. Ask your IT or security team if you're unsure. 10 AI QA checklist prompts you can use right now Prompt 1: Turn acceptance criteria into a QA checklist Use this when you have written acceptance criteria and need to turn them into testable steps before you hand off to QA or test it yourself. You are a QA engineer. Here are the acceptance criteria for [feature name]: [paste criteria] Convert these into a step-by-step QA checklist. For each item, include: - What to test - Expected result - How to verify it passed - Severity if it fails (critical / high / medium / low) Flag any criteria that are ambiguous or untestable as written. If your acceptance criteria are vague, fix them first. Check the prompts at /blog/ai-acceptance-criteria-prompts-stop-shipping-vague-work/ before you run QA on work that isn't clearly defined. Prompt 2: Find missing edge cases Use this when you've got a working checklist and want to pressure-test it for things nobody thought of. You are a QA engineer reviewing this checklist for [feature or workflow]: [paste checklist] Identify missing edge cases, including: - Empty or null inputs - Boundary conditions (minimum, maximum, zero values) - Concurrent user scenarios - Error states and failure paths - Accessibility and device/browser variation - What happens if a step is skipped or done out of order Output as additional checklist items with severity. Don't treat what comes back as exhaustive. The AI doesn't know your infrastructure, your user base, or your historical failure modes. Use it to jog your thinking, not replace it. Prompt 3: Check copy and content before launch Use this for marketing copy, UI text, error messages, emails, or any customer-facing content that ships with a build. You are a content QA reviewer. Review the following copy for [page / email / UI element]: [paste copy] Check for: - Spelling and grammar errors - Broken or placeholder text (e.g. "lorem ipsum," "[INSERT NAME]," "TBD") - Inconsistent terminology or brand voice - Missing CTAs or instructions - Anything that could confuse or mislead a user - Tone mismatches with the surrounding context Flag each issue with location and severity. A note on this one: if the copy references prices, legal terms, compliance language, or regulated claims, a human with the right expertise reviews it. Not the model. Prompt 4: Review a design handoff Use this when you're receiving a design file and need to check completeness before development starts. You are a design QA reviewer. Here is the design handoff for [screen or flow]: [describe or summarize the design, do not paste proprietary files] Check for: - Missing states (loading, error, empty, success) - Mobile and responsive considerations - Accessibility issues (color contrast, focus states, alt text needs) - Inconsistencies with the design system or existing components - Anything a developer would need that isn't documented List gaps with notes on what needs to be resolved before build starts. Pair this with your actual design system documentation. The AI doesn't know your component library. You do. Prompt 5: Pressure-test a spreadsheet or report Use this when a spreadsheet or report is being used to make decisions or sent to stakeholders. You are a data QA reviewer. I will describe a spreadsheet or report used for [purpose]: [describe structure, data sources, formulas, and audience, summarize only, no confidential data] Review for: - Formula logic errors or circular references - Missing data sources or broken links - Totals that don't add up or filters that exclude rows - Misleading visualizations or axis labels - Assumptions that aren't documented - Data that needs a date or version stamp Output a checklist of things to manually verify before this goes to [audience]. This prompt works on the structure and logic. The actual numbers need human verification against source data. Always. Prompt 6: Check customer-impacting workflow changes Use this before you change any process that touches customers: order flows, account management, notifications, billing, onboarding. You are a QA reviewer for customer-facing workflows. Here is a description of a change to [workflow name]: [describe the change, what triggers it, who it affects, and what the old behavior was] Generate a QA checklist covering: - Happy path verification - Rollback or undo behavior - What customers see vs. what they used to see - Communication or notification impacts - Edge cases where the change might behave unexpectedly - Any dependency on third-party systems Flag items that require testing in a staging environment before production. If this workflow touches billing, contracts, or compliance, escalate to the right team before shipping. The checklist won't catch regulatory gaps. A qualified human will. Prompt 7: Create a regression checklist Use this when a change to one area might have broken something else. You are a QA engineer building a regression checklist. A change was made to [component or feature]. Here is what changed: [describe the change] Here are the areas of the product that interact with this component: [list related features or integrations] Create a regression checklist covering: - Core functionality of the changed component - Adjacent features that could be affected - Integration points with [list systems] - Data flows that pass through this component - User permissions and role-based behavior Prioritize by likelihood of impact and severity if broken. The workflow audit prompts at /blog/ai-workflow-audit-prompts-find-the-bottlenecks/ can help you map what connects to what before you run regression testing. Prompt 8: Summarize QA findings for stakeholders Use this after testing is done, when you need to brief a non-technical audience. You are a QA lead writing a findings summary for [stakeholder audience]. Here are the raw QA findings: [paste findings, anonymized, no PII, no confidential data] Write a stakeholder summary that includes: - Overall status (go / no-go / conditional) - Number of issues by severity - Top 3 issues and why they matter - Issues resolved vs. open - Risks if we ship with known open issues - Recommended next steps Keep it under one page. Use plain language. The summary is a communication tool. The underlying findings document is the record of truth. Don't let the summary replace the detail when a real decision is being made. Prompt 9: Decide what needs human, legal, security, or privacy review Use this when you're not sure what to escalate. You are a risk reviewer. Here is a description of what we're shipping: [describe the feature, change, or workflow] Flag anything that likely needs review from: - Legal (contracts, terms, regulated claims, intellectual property) - Security (authentication, data access, API permissions, sensitive data storage) - Privacy (personal data collection, consent, data retention, third-party sharing) - Finance (billing, pricing, tax implications, financial reporting) - HR (employee data, performance management, hiring decisions) - Customer leadership (high-impact customer experience changes) - Accessibility (WCAG compliance, assistive technology) For each flag, note why it might need review and who should be looped in. This prompt is a triage tool, not a compliance assessment. If something gets flagged here, you escalate to the right expert. You do not use the AI's answer as the review. The risk assessment prompts at /blog/ai-risk-assessment-prompts-check-the-blast-radius/ go deeper on this if you need to map out the blast radius before you escalate. Prompt 10: Final pre-ship sanity check Use this in the last hour before you go live. You are a QA lead running a final pre-ship check for [feature or release name]. Here is a summary of what we built, what we tested, and what's still open: [describe the build, test coverage, open issues, and deployment plan] Run through this final checklist: - All acceptance criteria verified with evidence (screenshots, logs, or test records)? - Open issues documented with severity and owner? - Rollback plan documented and tested? - Stakeholder sign-off received and recorded? - Customer communication drafted if needed? - Data migration or cleanup steps verified? - Monitoring or alerting set up post-deploy? - On-call or support team briefed? Output a go / no-go recommendation with reasoning. Flag anything missing. The recommendation is a prompt, not a decision. A human makes the call. The AI helps you make sure you haven't forgotten something obvious at 4pm on a Friday. What AI QA checklist prompts can't do for you They can't tell you if the test actually passed. They can't replace a staging environment. They can't read your logs. They don't know what your product actually does in production, who your edge-case users are, or what your regulatory environment requires. They're fast at structure. They're decent at spotting gaps in what you've written. They will confidently generate checklist items that sound complete and aren't. That's the trap. A polished checklist is not the same as test coverage. The requirements gathering prompts at /blog/ai-requirements-gathering-prompts-stop-starting-vague/ are worth running before QA if the definition of "done" is still fuzzy. You can't test what you haven't defined. And when you find something bad: severity, likelihood, owner, date, rollback plan, escalation path. Write it down. Not in a chat window. In a place your team can find it. The human value in QA is knowing which defects actually matter, which edge cases are dangerous, and when "looks fine" is not good enough. The AI gets you to the checklist faster. You decide if the checklist is right. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Frequently asked questions Can AI write a QA checklist from scratch? It can generate a plausible-looking checklist from any description you give it. The problem is "plausible-looking" and "actually covers your product" are not the same thing. Use it to build structure from your real acceptance criteria, then have a human with product knowledge fill the gaps. Never ship based on a checklist the AI generated without verified requirements behind it. Is it safe to paste test cases and bug reports into ChatGPT? It depends what's in them. If they contain customer PII, employee data, credentials, security vulnerabilities, or confidential product details, do not paste them into an unapproved AI tool. Summarize the structure and anonymize the content. Ask your IT or security team what your organization's policy is before you assume it's fine. How do I make AI QA prompts more accurate? Give them more context: the acceptance criteria, the user type, the failure modes you're worried about, and the environment it runs in. Vague input produces vague output. The more you tell it about your actual constraints, the more useful the checklist becomes. See the prompts for acceptance criteria if you need to define those first. Should AI make the go/no-go call before launch? No. The AI can run a sanity check and flag missing items. The launch decision belongs to a named human who owns the outcome. Use Prompt 10 to make sure nothing obvious was missed. Then a person signs off, with the date and their name on it. What's the biggest mistake people make with AI QA checklists? Treating a complete-looking checklist as proof that quality was checked. The checklist is a starting point for testing, not evidence that testing happened. Real QA produces screenshots, logs, test records, and documented results. A checklist is the plan. Evidence is the outcome. Do I need to tell legal or security about everything I ship? Not everything, but more than most teams think. Anything touching authentication, personal data, billing, regulated claims, employee information, or significant customer-facing workflow changes should have at least a quick conversation with the right function. Prompt 9 is designed to help you triage what needs escalation. When in doubt, ask. The cost of a five-minute conversation is lower than the cost of a compliance incident. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance AI API Testing Prompts: Catch Broken Integrations AI API testing prompts for contracts, edge cases, authorization, retries, error responses, and human-verified integration evidence. AI promptsAPI testingquality assurance --- ## AI QBR Prompts: Run Reviews Without the Theater URL: https://dontreplace.me/articles/ai-quarterly-business-review-prompts-run-a-qbr-without-the-theater AI promptsquarterly business reviewQBRcustomer successaccount management AI QBR Prompts: Run Reviews Without the Theater 10 AI quarterly business review prompts to verify outcomes, expose risks, build a useful agenda, and plan next quarter without inventing customer value. September 09, 2026 Updated September 09, 2026 A quarterly business review can become a very expensive reading of slides everyone received yesterday. The vendor celebrates logins, the customer wonders whether anything improved, and slide 47 announces a roadmap nobody approved. AI quarterly business review prompts can help you audit evidence, find gaps, organize an agenda, draft questions, and record decisions. They cannot know what the customer values, invent ROI, read the room, approve a promise, or turn product activity into a business outcome by changing the chart color. A useful QBR is a decision conversation supported by evidence. It is not a victory montage with a meeting invite. These ten templates help customer success managers, account managers, founders, revenue teams, and operations leads prepare a QBR that respects the customer's time and keeps human judgment in charge. What should a quarterly business review accomplish? A good QBR should establish what changed, what the evidence supports, what remains unresolved, and what both sides will do next. It should not merely prove that your team was busy. A useful review answers: Which customer goals are still current? What results can both sides verify? Which numbers show activity rather than value? What adoption, support, delivery, or trust risks need attention? Which priorities changed since the last review? What decisions must be made in the meeting? Who owns each next action, and by when? Which claims or plans still need approval? Keep evidence categories separate. A verified fact has a current source. A customer statement records what someone actually said. An internal interpretation is a theory to test. A measured outcome uses an agreed definition. An unknown needs a question. A decision has a named owner with authority. That distinction matters because models are excellent at making incomplete notes sound settled. If your broader account context is messy, start with AI account planning prompts. For a review that feeds an upcoming commercial decision, pair this process with AI customer renewal prompts. Use this evidence-first QBR prompt formula Add this instruction to every template below: “Act as a quarterly business review preparation assistant. Use only the sanitized, approved sources I provide for [review period, customer goals, usage, outcomes, support, delivery, stakeholders, risks, and next-quarter plans]. Cite a source label and date for every metric, outcome, issue, deadline, commitment, and stakeholder statement. Separate verified facts, customer statements, internal interpretations, measured outcomes, decisions, and unknowns. Mark stale, conflicting, or unsupported claims [VERIFY]. Do not invent sentiment, ROI, causation, authority, urgency, product capabilities, approvals, legal interpretations, or future results.” The source rule is the useful part. Without it, a model may blend last quarter's goal, this quarter's usage, and somebody's optimistic CRM note into one suspiciously smooth success story. Never paste customer names, emails, contracts, credentials, raw usage exports, private CRM notes, support transcripts, payment details, confidential pricing, security findings, health information, regulated data, or legally sensitive material into an unapproved AI tool. Use approved systems, minimum necessary data, de-identification, access controls, retention limits, and human privacy, security, legal, finance, product, and customer-success review. What should you collect before using ChatGPT for a QBR? Build a small, current source pack. Do not upload the company's entire customer nervous system because the chatbot asked for “more context.” InputWhat it can supportHuman owner Current goals and success planThe outcomes the review should assessCustomer and account owner Agreed metric definitionsWhat each number does and does not meanAnalytics or business owner Dated usage summaryProduct activity and adoption patternsProduct or analytics owner Verified outcome evidenceSupported business resultsCustomer and account owner Support and incident historyOpen friction, severity, and statusSupport owner Delivery commitmentsPromises completed, changed, or openDelivery owner Stakeholder notesRoles, statements, and participationAccount owner Last QBR decisionsContinuity and accountabilityMeeting owner Approved roadmap languageWhat may be discussed externallyProduct owner Next-quarter constraintsReal timing, budget, and dependenciesAuthorized owners Data-use rulesWhat may enter the AI systemPrivacy or security owner If two sources disagree, preserve the disagreement. Quietly choosing the happier number is not synthesis. It is customer theater with formulas. 10 AI quarterly business review prompts Replace brackets with approved, sanitized information. Review every result before it reaches a deck, email, CRM, or customer. 1. Audit the QBR source pack “Compare this sanitized QBR source pack with these review requirements: [paste]. Create a table with required input, supplied evidence, source, date, status, conflict, missing information, owner, and next action. Do not fill gaps by inference. Mark stale or unsupported claims [VERIFY].” Run this first. It catches goals copied from an old success plan, mismatched reporting periods, undefined metrics, missing customer confirmation, unresolved incidents, and commitments with no owner. The model can identify an empty field. It cannot decide whether the source is trustworthy or whether the omission matters. Route each gap to the person who owns the data or relationship. 2. Separate outcomes from product activity “Classify each item in these notes and metrics as product activity, adoption signal, measured business outcome, customer statement, internal interpretation, or unknown. Include definition, period, source, caveat, and customer-confirmation status. Do not convert activity into value or make causal claims without evidence.” Logins, seats, clicks, exports, and feature use can show activity. They do not automatically prove time saved, revenue gained, risk reduced, or employees delighted beyond human comprehension. A strong QBR can say, “Adoption increased, but we have not yet validated business impact.” That sentence may feel less exciting than fabricated ROI. It is also true, which is a useful feature in a customer meeting. 3. Check current goals against evidence “For each customer goal in this approved success plan, create a review row with goal, current relevance, agreed measure, baseline, current result, source, customer confirmation, confidence, caveat, and question to resolve. Label goals that have not been recently revalidated [RECONFIRM].” Goals age. A target from six months ago may have been replaced by a new executive priority, budget constraint, integration plan, or operating model. Do not grade the customer against a stale plan and call it insight. Use the output to ask whether each goal still matters, whether the measure remains useful, and whether the result is meaningful to the customer. The customer owns the meaning of value; your dashboard does not get veto power. 4. Find adoption, support, and delivery risks “Review these sanitized usage summaries, support records, incident notes, and delivery commitments: [paste]. Build a risk register with observed signal, source, date, affected goal, severity rationale, known impact, owner, mitigation status, customer visibility, unresolved question, and next checkpoint. Do not assign root cause or churn probability without an approved method.” Low use can mean poor onboarding, weak fit, missing permissions, seasonal work, a broken integration, or a feature the customer never needed. An open support ticket can be minor or the visible edge of a trust problem. AI can group signals. Humans must investigate causes and consequences. For a more formal check, use AI risk assessment prompts. If patterns span many accounts, AI churn analysis prompts can structure the analysis without pretending correlation is destiny. 5. Map stakeholders and unanswered questions “Create a stakeholder map from these dated, approved notes: [paste]. Include role, stated priority, known concern, review participation, decision authority evidence, relationship owner, last verified date, and unanswered question. Put inferred influence or sentiment in a separate hypothesis column. Do not profile people from demographic or unrelated personal data.” The daily user, executive sponsor, champion, budget owner, procurement contact, security reviewer, and signer may all be different people. A job title does not prove authority, and silence does not prove satisfaction. Use this map to decide who should attend and which questions need direct answers. Do not use it to generate fake familiarity. The reliable way to learn what a stakeholder thinks remains painfully analog: ask, listen, and update the record. 6. Build a decision-focused QBR agenda “Using this verified evidence, risk register, stakeholder map, and unknowns list: [paste], draft a 45-minute QBR agenda. For each section include purpose, decision or question, evidence needed, facilitator, customer participant, time box, and expected output. Put unresolved issues before future plans. Remove sections that only repeat information.” A QBR agenda should spend meeting time on interpretation, decisions, and trade-offs. Send routine metrics in advance. Do not trap six people in a video call so one person can narrate a dashboard. A sensible sequence is: confirm goals, review supported outcomes, address open risks, test changed priorities, make decisions, and assign next actions. If no decision or useful question exists, that section may belong in the pre-read instead. 7. Draft an honest executive summary “Draft a concise QBR executive summary from these approved sources: [paste]. Include current goals, verified outcomes, activity that is not yet an outcome, material risks, unresolved questions, decisions needed, and proposed next steps. Cite source labels. Preserve caveats and disagreements. Mark unsupported claims [VERIFY]. Avoid promotional language.” The summary should let a busy stakeholder understand the state of the relationship without decoding a deck. It should not hide a failed commitment beneath “strong momentum” or turn one positive comment into “overwhelming customer enthusiasm.” For structure, AI executive summary prompts can help. Keep the evidence controls here. A concise lie is still a lie; it just fits on one slide. 8. Prepare neutral discovery questions “Using these verified findings and unknowns: [paste], draft open, neutral QBR questions grouped by goals, outcomes, adoption, unresolved issues, changing priorities, stakeholders, and next-quarter trade-offs. For each question, state which unknown it tests. Avoid leading wording, sales pressure, and assumptions about satisfaction, budget, or intent.” Useful questions sound like: “Which outcome matters most next quarter?” “Where does the current workflow create friction?” “What has changed since we agreed on this measure?” “Who else needs to approve the next step?” Bad questions smuggle in the desired answer: “Given the incredible value delivered, are you ready to expand?” That is not discovery. It is a hostage situation wearing business casual. 9. Create the next-quarter action plan “Turn these confirmed QBR decisions and approved notes into a next-quarter action plan. Include outcome, action, named owner, due date, dependency, success measure, source decision, review checkpoint, and escalation path. Keep proposed items separate from agreed commitments. Do not assign work or promise dates that the notes do not support.” The plan should distinguish “discussed,” “proposed,” and “agreed.” Otherwise, a brainstorm becomes a commitment before the meeting recording has cooled down. Record material choices with AI decision log prompts. Every action needs a human owner who accepts it. “Customer team” and “vendor team” are categories, not accountable people. 10. Draft the factual recap and sign-off check “Draft a short QBR follow-up from these approved notes: [paste]. Include confirmed goals, supported outcomes, customer concerns in neutral language, decisions, unresolved questions, and commitments with named owners and dates. Then audit the draft for unsupported claims, invented agreement, privacy risk, unapproved roadmap language, vague ownership, and missing specialist review. Mark every issue [VERIFY]. Do not send or approve the message.” A fast recap is useful because memories diverge quickly. Accuracy matters more than polish. If the customer challenged a metric, preserve that. If product needs to confirm a capability, do not write that it is coming. If no date was agreed, do not manufacture one for symmetry. The meeting owner should review the final note. Product, support, security, legal, finance, and privacy owners should verify statements in their domains. The model can draft and inspect; it cannot approve the record. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → A simple QBR workflow that keeps humans accountable Use the prompts in a controlled sequence: Define the meeting purpose. Name the decisions and questions that justify live time. Collect minimum necessary sources. Use current, approved, sanitized evidence. Audit the pack. Find gaps and conflicts before making slides. Separate activity from outcomes. Keep claims within what evidence supports. Validate goals and risks. Ask owners and the customer instead of guessing. Send a short pre-read. Reserve meeting time for discussion and decisions. Run the conversation. Listen, test assumptions, and update the record. Confirm decisions aloud. Name owners, dates, caveats, and approvals. Send an accurate recap. Preserve disagreement and unresolved items. Track the work. Review commitments before the next quarter arrives. This process is less dazzling than asking an “AI customer success copilot” to create an entire QBR from CRM exhaust. It is also less likely to put a fictional result in front of the customer who knows it is fictional. If your team is still learning where the tool belongs, the no-BS guide to using AI at work has the basic rule: give AI bounded work, inspect the result, and keep accountability attached to a person. Common QBR mistakes AI can make worse Reusing last quarter's story Old goals, stakeholder roles, roadmap notes, and risk labels can look authoritative because they are already formatted. Date every material source and revalidate it. Copying stale context faster is not productivity. Reporting vanity metrics as value A larger number is not automatically a better customer outcome. Define what each metric measures, what it excludes, and whether the customer agrees it matters. Hiding bad news until the final slides Unresolved support, delivery, security, or trust issues belong early in the conversation. A cheerful product roadmap does not neutralize a known failure. Inventing certainty from partial data Sparse data can suggest questions. It cannot support precise claims about causation, sentiment, churn, or future results. Label uncertainty and investigate it. Letting the tool promise the roadmap AI cannot approve features, dates, pricing, security exceptions, legal terms, staffing, or service commitments. Use only approved language and route open items to their real owners. For a plain explanation of the boundary, read what AI can and cannot do. Fast drafting is useful. It is not authority, empathy, or a customer relationship. Frequently asked questions Can ChatGPT create a full QBR deck? It can draft a structure and copy from approved, sanitized evidence. Humans must verify every claim, number, chart, date, quote, capability, and commitment. Do not give an unapproved consumer tool raw customer records or let generated slides go directly to a customer. What is the best AI prompt for a quarterly business review? The best prompt requires sources, dates, evidence categories, visible unknowns, and explicit limits. Tell the model not to infer sentiment, ROI, causation, authority, or approval. A narrower prompt that produces a checkable table is usually safer than “make me an impressive QBR.” How many slides should a QBR have? Use as few as the conversation needs. Put routine detail in a pre-read and use live time for goals, supported outcomes, risks, changed priorities, decisions, and next actions. The correct count depends on complexity, not a sacred template. Can AI calculate customer ROI for a QBR? AI can apply a formula that authorized owners provide to verified inputs. It cannot choose a valid method, invent a baseline, establish causation, or decide which costs count. Finance, analytics, the account owner, and the customer should validate any ROI claim before it appears externally. Should a QBR include product usage data? Yes, when usage helps explain adoption or a customer goal and the data is accurate, authorized, and appropriately minimized. Label usage as activity unless evidence connects it to an outcome. Explain the reporting period, definition, exclusions, and caveats. Can AI identify churn risk during a business review? It can organize known signals and highlight missing information. It cannot know whether the customer will leave or assign a reliable probability without a validated method and suitable data. Use risk output to guide questions and action, not to declare the customer's intent. What customer data is safe to use with an AI QBR tool? Use only data allowed by your organization's policy, contract, access rules, and approved system configuration. Minimize and de-identify where possible. Keep personal data, credentials, contracts, raw private communications, payment details, confidential pricing, security findings, and regulated information out of unapproved tools. Who should approve the final QBR? The accountable account or customer-success owner should own the review. Analytics, product, support, delivery, finance, legal, security, and privacy owners should verify claims in their domains. The customer confirms their goals, meaning, and commitments. AI approves none of those things. The useful boundary The strongest use of AI in a QBR is disciplined preparation. It can expose missing evidence, separate activity from outcomes, organize risks, generate better questions, and turn confirmed decisions into a clean action plan. The human work remains the part customers notice: listening, understanding context, acknowledging failures, choosing trade-offs, making authorized commitments, and building trust over time. That is the broader point of Don’t Replace Me by Dmitry Kargaev. Use the machine for speed and structure, then keep judgment and accountability where they belong. AI can build the draft before lunch. It cannot make the relationship real. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Customer Renewal Prompts: Prepare With Evidence 10 AI customer renewal prompts to verify outcomes, expose account risks, prepare options, and protect trust without inventing value or customer sentiment. AI promptscustomer renewalcustomer success AI Customer Retention Prompts: Keep Customers 10 AI customer retention prompts to diagnose churn signals, plan useful follow-up, protect customer data, and keep human judgment in charge. AI promptscustomer retentioncustomer success AI Customer Success Plan Prompts: Use Real Evidence 10 AI customer success plan prompts to verify goals, map adoption gaps, assign milestones, and build an account plan without inventing customer value. AI promptscustomer success plancustomer success --- ## AI Quarterly Planning Prompts: Turn Vibes Into a Real Plan URL: https://dontreplace.me/articles/ai-quarterly-planning-prompts-turn-vibes-into-a-real-plan quarterly planningAI promptsOKRsplanning templatesstrategic planning AI Quarterly Planning Prompts: Turn Vibes Into a Real Plan 10 copy-paste AI quarterly planning prompts for reviewing last quarter, drafting OKRs, pressure-testing priorities, and mapping risks. Real templates, no fluff. June 28, 2026 Updated June 28, 2026 Most quarterly plans look great on the day they're made. By week three, nobody can find the document. By week six, the goals have quietly shapeshifted. By the end of the quarter, everyone's writing their review against whatever actually happened, not what they planned. AI quarterly planning prompts won't fix that on their own. But they will help you get from "vibes and good intentions" to something written down, pressure-tested, and actually usable. Faster than you'd do it alone. That's the deal here: 10 copy-paste prompts for every stage of your quarterly planning cycle, from reviewing last quarter to deciding what you're explicitly not doing this time. Plus the formula behind them, so you can adapt any prompt to your context. Before we get into it, the safety reminder: do not paste customer PII, employee health data, private performance records, credentials, unreleased financials, legal disputes, security incidents, board materials, or anything marked confidential into any AI tool that hasn't been approved by your IT or security team. These prompts work fine with anonymized, aggregated, or internal-only information. The formula behind good AI quarterly planning prompts Every prompt in this article follows the same structure. Once you see it, you can write your own. Role + Context + Inputs + Task + Constraints + Output format That's it. Most bad prompts fail at "context" and "inputs." They give the AI a task but no raw material to work with, so it invents plausible-sounding things. Quarterly plans built on invented things are just expensive fiction. The context is your actual situation: team size, budget range, what happened last quarter, what leadership is prioritizing, what's changing. The inputs are your real numbers: OKR results, customer signals, capacity, key dependencies. The constraints force the AI to stay realistic. The output format tells it what to produce. Rule #13 from Don't Replace Me: garbage in, garbage out. Vague goals produce polished nonsense. Treat these prompts as a sparring partner, not a ghostwriter. If your inputs are vague, your outputs will be confident-sounding vague. Spend five minutes gathering real data before you run any of these. Prompt 1: Review last quarter honestly Before you plan forward, you need an honest read on where you actually landed. This prompt turns your notes, metrics, and gut feelings into a structured retrospective. You are a strategic planning facilitator. I'm reviewing Q[X] for a [team type, e.g., "12-person marketing team" or "ops function at a 200-person SaaS company"]. Here are our results against goals: [Paste your OKR/goal results. Use percentages or ratings, not names or PII.] Here are the 3 things that went well: [Your list] Here are the 3 things that didn't work: [Your list] Key external factors that affected us (market, company changes, etc.): [Brief notes] Draft a structured Q[X] retrospective summary. Include: what we delivered, what we missed and why, patterns in the misses, and 3-5 lessons we should carry forward. Keep it honest, not political. Don't sanitize the inputs. If the quarter was a mess, say it was a mess. The AI will write a more useful retrospective if you're honest with it than if you feed it polished spin. Prompt 2: Extract lessons that actually change behavior Most lesson-learned exercises produce a list that gets filed and ignored. This prompt pushes for actionable changes. Here are the lessons from our Q[X] retrospective: [Paste lessons from Prompt 1 output or your own notes] For each lesson, identify: (1) what specific behavior or process should change next quarter, (2) who owns that change, and (3) how we'd know in 4 weeks if the change is working. Format as a table: Lesson | Behavior change | Owner role (not name) | 4-week signal. The "4-week signal" column is the one most teams skip. It's also the one that separates real commitments from retroactive cover stories. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Prompt 3: Turn messy goals into themes You've got 15 competing priorities from leadership, three from your own team, and a customer request that doesn't fit anywhere. Before you can draft OKRs, you need to find the actual themes. Here is a list of goals, requests, and priorities I've been given for Q[X+1]: [Paste your raw list. Keep it messy. Don't pre-organize.] Group these into 3-5 strategic themes. For each theme: name it, describe it in one sentence, list which goals belong to it, and flag any goals that don't fit any theme (these are probably noise or scope creep in disguise). Don't invent themes that aren't there. If some goals are genuinely contradictory, call that out explicitly. The last instruction matters. Left to its own devices, AI will synthesize contradiction into a coherent-sounding theme. You want it to surface the contradiction so you can make a call. Prompt 4: Draft OKRs from your themes Once you have themes, you can draft OKRs. This prompt produces a starting draft, not a finished document. You'll need to verify every key result against real data sources. Here are our Q[X+1] strategic themes and the goals under each: [Paste from Prompt 3] Our team context: [size, function, what we own, key constraints like budget or headcount changes] Draft OKRs for each theme. Each Objective should be qualitative and direction-setting. Each Key Result should be specific, measurable, and achievable given our constraints. Maximum 3 KRs per Objective. Flag any KR where you don't have enough information to set a realistic target. I'll verify all numbers against our actual source-of-truth metrics before finalizing. Review every number the AI suggests against your actual data. It might suggest "increase conversion rate by 20%" when your baseline is already at 8% and the industry average is 9%. It doesn't know that. You do. Prompt 5: Pressure-test your priorities This is the prompt most people skip, and it's the most important one. It's also where the AI is most useful, because it has no political skin in the game. For more on building a solid prioritization framework, the AI prioritization prompts article has 10 templates specifically for ranking competing work. Here are our Q[X+1] priorities and OKRs: [Paste] Ask me the hardest questions about this plan. Specifically: (1) What assumptions are we making that could be wrong? (2) What are we treating as independent that are actually dependent? (3) What happens to this plan if [key resource] is reduced by 30%? (4) Which priorities conflict with each other? (5) What are we hoping goes right without controlling it? Don't soften the questions. I want the version that would embarrass us if we hadn't asked it. Read the output in a room with the people who have decision rights. Not over Slack. Not asynchronously. The questions it surfaces need a real conversation, not a comment thread. Prompt 6: Map risks and dependencies Every plan has hidden risks. This one finds them before Q1 week three does. For deeper risk work, the AI risk assessment prompts article covers blast-radius thinking in detail. Here is our Q[X+1] initiative plan: [Paste your priorities, OKRs, and key projects] Team context: [size, key dependencies on other teams, known constraints, any major company changes happening this quarter] Produce a risk and dependency map. For each initiative: list the top 2-3 risks, identify what it depends on from other teams or external factors, flag capacity risks (what could break if one thing slips), and rate each risk as low/medium/high. Note: I'll validate all dependency assumptions with the actual teams before finalizing. The AI will find more risks than you're comfortable with. That's the point. Your job is to decide which ones you're mitigating, which ones you're accepting, and which ones you're escalating. Prompt 7: Build a capacity-aware initiative plan Most plans fail not because the goals are wrong, but because nobody counted the hours. This prompt forces the capacity conversation. Here are our Q[X+1] initiatives: [List them] Here is our rough team capacity: [E.g., "4 people, each roughly 60% available for project work after meetings and BAU. One person out for 3 weeks in month 2."] Map each initiative to realistic capacity. Flag any initiatives that require more capacity than we have. Suggest which should be deferred, descoped, or given to a different owner. Don't tell me we can do everything. If we can't, say so clearly. That last line is load-bearing. Without it, the AI will find a way to make everything fit. You need it to tell you the truth. For budget constraints alongside capacity planning, the AI budget planning prompts article covers scenario planning and headcount requests. Prompt 8: Write the quarterly narrative for stakeholders Once you've built the plan, you need to communicate it. This prompt turns your planning docs into a coherent narrative for your team, your boss, or a stakeholder update. Here is our Q[X+1] plan: [Paste themes, OKRs, and key initiatives] Audience: [e.g., "my direct team of 8" or "senior leadership monthly review"] Write a 300-word quarterly narrative. It should cover: what we're focused on and why, what we're explicitly not doing, what success looks like by the end of the quarter, and what we need from this audience to succeed. Tone: direct, not corporate. No jargon. You'll edit this. The AI draft is a starting point, not a finished communication. Your voice and your actual relationship with the audience matter more than the template. Prompt 9: Build a review cadence A plan without a review cadence is a wish. This prompt sets up the rhythm that keeps the plan alive. For templates to keep those reviews running well, the AI status report prompts article and the AI project management prompts article both have formats you can use every week. Here are our Q[X+1] OKRs and key initiatives: [Paste] Design a review cadence for the quarter. Include: (1) a weekly check-in format (15-20 min, what questions to ask each week), (2) a monthly mid-quarter review structure (what to review, what decisions to make), (3) an end-of-quarter review format (what to measure, what to document for next quarter). Make it lightweight enough that people will actually do it. Flag the 2-3 metrics we should track every single week as our leading indicators. The leading indicators question is the most valuable output here. You want the signals that tell you the quarter is going sideways four weeks before the final results do. Prompt 10: Decide what you're not doing this quarter This is the prompt that separates real planning from wishful thinking. Every good quarterly plan has an explicit "not this quarter" list. Here is our backlog of potential work and the requests we've received that didn't make it into the Q[X+1] plan: [Paste] For each item we're deferring: write one sentence explaining why it's not Q[X+1] work, when it should be reconsidered, and what condition would need to be true for it to become a priority. This list will be shared with stakeholders so they understand what didn't make the cut and why. It's not a rejection. It's a parking lot with a rationale. The "condition that would need to be true" column stops the same conversation from happening every month. When someone asks why X isn't on the plan, you can point to the condition. Either it's met or it isn't. What AI can't do in quarterly planning It can't tell you what actually matters to your customers. It can't read the room in your leadership team. It doesn't know which stakeholder will quietly kill an initiative if they feel bypassed. It doesn't know your company's actual risk tolerance, or whether the CFO's stated budget is the real budget. It also can't make the hard calls. When two priorities genuinely conflict and you only have capacity for one, that's a human decision. The AI will find a way to frame both as compatible if you let it. Don't let it. What AI can and can't do is worth reading before you use these prompts, if you want the full picture of where the tool earns its keep and where it confidently makes things up. The point of these prompts is to get the scaffolding built faster, so you can spend your energy on the judgment calls that actually require a human. That's a reasonable trade. Just don't mistake a tidy roadmap for actual leadership. Dee covers this throughout Don't Replace Me, particularly the tension between AI's speed and the strategic judgment that has to sit underneath it. The book is worth the read if you're in a role where the plan you sign off on has real consequences. Frequently asked questions Can AI really help with quarterly planning, or does it just produce generic output? It depends entirely on what you feed it. Generic inputs produce generic outputs. If you give the AI your actual metrics, real constraints, team capacity, and specific context, the output is genuinely useful as a first draft. The prompts in this article are designed to force specific inputs, which is where most people shortcut and where the quality difference comes from. Is it safe to paste my company's quarterly plan into ChatGPT or Claude? Not without checking your company's AI use policy first. Never paste customer PII, employee health data, unreleased financials, board materials, legal disputes, security incidents, confidential contracts, or anything marked sensitive into an unapproved AI tool. Use anonymized or aggregated data where possible, and confirm which tools are approved by your IT or security team before using them for planning work. What's the difference between AI-drafted OKRs and good OKRs? AI-drafted OKRs give you a starting structure fast. Good OKRs require you to verify every key result number against your actual source-of-truth metrics, validate that the targets are realistic given your real capacity and constraints, and confirm that the objectives connect to what your organization actually cares about this quarter. The draft is the easy part. The verification is the work. How do I stop my quarterly plan from dying by week three? Build the review cadence at the same time as the plan. Prompt 9 in this article does that. The single most useful thing: identify two or three leading indicators you'll check every week, not just the lagging metrics you'll know about at the end of the quarter. Early signals let you course-correct before the quarter is lost. Should I use these prompts in a group or individually? Both work. Individual use is faster for generating the first draft. Group use, where you run the outputs in a working session, is more useful for the pressure-testing prompts (Prompt 5) and the "not this quarter" list (Prompt 10). Those benefit from the debate that happens in a room. For more on using AI to support group decision-making, the AI decision-making prompts article covers structured approaches. What if the AI's quarterly plan looks great but feels wrong? Trust the feeling and investigate it. Ask the AI to argue against the plan, or run Prompt 5 on the plan it just produced. Often the feeling is catching a real assumption the AI baked in that you know to be false. Sometimes it's just unfamiliarity with the format. The test is whether you can explain, specifically, what's wrong. If you can, fix it. If you can't, run the pressure-test prompt and see what comes out. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance --- ## AI Regression Testing Prompts: Check What Broke URL: https://dontreplace.me/articles/ai-regression-testing-prompts-check-what-the-change-broke AI promptsregression testingquality assurancesoftware testingrelease management AI Regression Testing Prompts: Check What Broke AI regression testing prompts for mapping verified changes to risky workflows, past defects, test evidence, and human-owned release decisions. August 08, 2026 Updated August 08, 2026 A tiny product change walks into a release and says, “I only touched one thing.” Three hours later, checkout is broken in Firefox, yesterday’s exports are duplicating rows, and somebody is asking whether the team can “just monitor it in production.” Regression testing exists because software changes have neighbors. A fix, configuration update, migration, integration change, or new feature can damage behavior that worked before. AI regression testing prompts can help you map verified changes to affected workflows, organize historical defects, draft candidate coverage, and expose missing evidence. They cannot inspect your system, execute a test, know hidden dependencies, validate results, or approve a release. AI can help draw the regression map. Humans still have to walk it, collect evidence, judge failures, and own the release decision. If you do not yet have a coherent overall strategy, start with AI test plan prompts. If you know what changed and need to check what it might have broken, this guide gives you a practical workflow and 10 copy-paste prompts. What regression testing actually does Regression testing checks whether a change damaged existing behavior. It is not the same as testing only the new feature. It asks a more annoying and more valuable question: what else could this change affect? Useful regression coverage connects: the verified change set and its purpose; affected components, integrations, data, permissions, and configurations; critical user and operational journeys; previous defects and incidents; supported environments, devices, roles, and locales; expected results and required evidence; named owners, escalation paths, and release gates. A weak regression instruction says, “Test the site.” A useful one says, “Recheck guest and signed-in checkout across supported payment methods because this release changed tax calculation and the order API; capture transaction, receipt, inventory, and error evidence; escalate mismatches to the named owners.” That specificity requires verified context. AI can organize context quickly, but polished output is not system knowledge. The plain-language explanation of what AI can and cannot do is worth keeping nearby whenever a model starts sounding suspiciously certain. The reusable AI regression testing prompt formula Use this base prompt before the specialized templates: “Act as a regression-test planning assistant. I am [role] reviewing [release, fix, migration, configuration change, integration, or workflow update]. Use only the privacy-safe, verified information below. Map each change to potentially affected components, user journeys, integrations, data flows, permissions, environments, and historical defects. Propose risk-based regression coverage with preconditions, expected results, owner fields, evidence, and unresolved questions. Separate confirmed facts from suggestions. Cite the supplied source for each connection. Do not invent dependencies, product behavior, test results, owners, dates, severity, or approval. Mark unsupported assumptions as ‘needs human confirmation.’” The important phrases are “use only,” “cite the source,” and “needs human confirmation.” Without them, an AI will happily turn a guess into a table cell, then turn the table cell into fake certainty. Never paste credentials, access tokens, customer PII, private tickets, raw production logs, security vulnerabilities, regulated data, unreleased strategy, legal disputes, HR records, or confidential client material into an unapproved AI tool. Use redacted summaries, synthetic data, approved schemas, and references to controlled systems. What to collect before asking AI for regression coverage Do not provide only a release name. “Version 4.2” is not context. It is a label wearing a tiny hat. InputWhy it mattersHuman validation Approved change inventoryDefines what actually changedConfirm deployed code, config, flags, and migrations Acceptance criteriaDefines intended new behaviorResolve vague or conflicting language Dependency mapShows likely downstream effectsVerify with technical and operational owners Critical user journeysPrioritizes business and customer impactConfirm current workflows and supported variants Prior defects and incidentsReveals proven failure patternsCheck relevance to the changed area Existing regression suitePrevents duplicate or missing coverageConfirm tests still run and assert useful behavior Environment inventoryShows where evidence can be trustedDocument parity gaps and unavailable integrations Test-data rulesEnables safe representative casesConfirm privacy, retention, and cleanup controls Release and rollback planConnects failures to decisionsVerify owners, timing, and trigger thresholds Evidence standardDefines what counts as a credible passAgree on logs, screenshots, records, or reports If an input is missing, ask the model to draft questions. Do not let it fill the blank. A visible gap can be managed; an invisible invention can ship. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI regression testing prompts Replace bracketed text with verified, privacy-safe information. Every output is a draft for accountable humans to review. 1. Summarize the verified change set “Review this approved change log, ticket list, configuration summary, migration note, and feature-flag plan: [paste]. Create a change inventory with source reference, change type, affected component, intended behavior, rollout condition, data or schema effect, integration effect, permission effect, reversibility, owner field, and unresolved question. Separate confirmed changes from suspected or undocumented changes. Do not infer that a ticket was deployed.” Regression planning starts with what actually changed, not what the sprint board hoped would change. Compare the output with build records, deployment manifests, approved configuration, and migration history. This is also where scope drift becomes visible. A “copy update” that includes a tracking-script change is not merely copy. A “backend fix” that modifies a shared validation library may have a much larger blast radius. 2. Map changes to affected workflows “Using this verified change inventory and these current user and operational workflows: [paste], build an impact map. For each change, list directly affected steps, potentially affected adjacent steps, roles, channels, devices, locales, permissions, integrations, and downstream records. Cite the supplied evidence for every direct connection. Mark candidate connections without evidence as questions, not facts.” Use this after or alongside AI change impact analysis prompts. The model is useful for joining supplied lists and spotting repeated terms. Humans must verify whether those joins reflect the real architecture and real work. Do not test every conceivable path equally. Confirm the critical journeys, then prioritize based on impact, likelihood, change proximity, defect history, and detectability. 3. Select risk-based regression coverage “Create a risk-based regression matrix from this verified change inventory, impact map, critical-journey list, and human-approved risk labels: [paste]. Include journey, failure mode, impact label, likelihood label, change proximity, existing test reference, proposed additional check, environment, data need, owner field, evidence, and release consequence. Use only supplied ratings. Flag high-impact paths with no coverage and low-value duplicate checks.” A regression suite is not valuable because it is large. It is valuable because it checks important behavior with trustworthy assertions at a cost the team can sustain. Ask humans to challenge both omissions and excess. Old suites often contain hundreds of checks that prove a page still has a button while ignoring whether a customer can complete the transaction. 4. Revisit past defects and incidents “Review these redacted historical defects, incident summaries, and support themes: [paste]. Identify items plausibly related to the current verified changes. For each candidate, include the original failure pattern, affected journey, prior root cause if confirmed, current connection evidence, existing regression check, proposed retest, expected result, owner field, and why the item may be irrelevant. Do not claim recurrence without test evidence.” Past failures are useful because they happened in your system, not in an AI’s imagination. But historical tickets can also mislead when architecture and workflows have changed. Pair this with AI defect triage prompts when failures appear. The model can format comparisons; humans decide whether a new symptom is the same defect, a related defect, or an unrelated mess arriving at an inconvenient time. 5. Check integrations and data flows “Using this approved dependency map, interface list, and data-flow summary: [paste], draft candidate regression checks for changed or adjacent integrations. Include trigger, source, destination, payload category without sensitive values, expected response, documented timeout or retry behavior, idempotency rule if supplied, data persistence, failure handling, observability evidence, environment limitation, owner fields, and unresolved questions. Never invent an endpoint, contract, retry rule, or expected response.” Integration regression is where “nothing changed over there” goes to die. Shared schemas, queues, webhooks, caches, scheduled jobs, permissions, and data transformations create distance between a change and its symptom. Use AI dependency mapping prompts if the dependency evidence is scattered. Keep secrets and real customer payloads out of the model. 6. Define environment and test-data needs “Create a regression environment and test-data readiness table from this approved information: [paste]. Include environment, deployed version, configuration differences, feature flags, available integrations, roles, safe data sets, reset method, monitoring, known limitations, access owner, readiness check, and scenarios that cannot be trusted there. Do not request or reproduce credentials, tokens, production customer records, or confidential payloads.” A check can pass because staging quietly removed the difficult dependency, uses relaxed permissions, has tiny data volume, or runs different configuration. Record those differences instead of burying them beneath a green status icon. Representative does not mean copied from production. Prefer synthetic or properly governed test data. Qualified privacy, security, and compliance reviewers should define handling rules for sensitive systems. 7. Separate smoke coverage from deeper regression “Using this risk matrix and release timeline: [paste], divide candidate checks into pre-deployment validation, deployment smoke tests, focused regression, broader regression, and post-release monitoring. For each check include purpose, trigger, estimated effort supplied by humans, environment, owner field, evidence, failure action, and whether it blocks progression. Do not label a short smoke check as proof of full regression.” Smoke testing answers whether the build is alive enough for more testing. Regression asks whether important existing behavior still works. Confusing them creates a release ceremony in which five green checks are asked to carry the emotional weight of an entire product. Use AI smoke test prompts to design the narrow first layer. Then make the deeper coverage and its limits explicit. 8. Assign owners, evidence, and review gates “Convert this approved regression matrix into an execution tracker. Include check ID, source risk or requirement, executor field, reviewer field, environment, test data, expected result, evidence location, status options, defect link, retest owner, due point, escalation path, and release gate. Leave unknown owners and dates blank. Flag checks that have no reviewer, no evidence standard, or no failure action.” “QA owns testing” is not an ownership model. Product may own intended behavior, engineering may own technical diagnosis, operations may own workflow validation, security may own specialist review, and a release owner may own the final decision. Evidence should be proportional to risk. A screenshot might establish a visual state. It cannot prove data integrity across a multi-system transaction. Define what another reviewer needs to verify the result. 9. Triage failed regression checks “Using these privacy-safe failed-check records: [paste], create a triage worksheet. Preserve observed facts separately from interpretation. Include check ID, expected result, actual result, environment, build, data conditions, reproducibility, evidence reference, affected journey, possible relation to current changes, severity field for human decision, proposed owner, containment question, retest need, and regression-suite update question. Do not assign root cause, severity, or release disposition without human confirmation.” A failed check is evidence of a mismatch, not automatic proof of root cause. The test may be wrong, the environment may be stale, the requirement may have changed, or the product may be broken. Good triage protects facts from group-chat storytelling. If the issue is real, create a useful report with AI bug report prompts and keep the actual evidence in the approved system of record. 10. Draft the regression summary for release review “Draft a concise regression summary from these verified execution records: [paste]. Include release identifier, tested scope, explicit exclusions, environment limitations, completed and incomplete coverage, passed and failed checks, open defects by human-approved severity, retest status, evidence links, unresolved risks, rollback readiness, named decision owner field, and questions for go/no-go review. Do not invent totals, imply complete coverage, convert ‘not run’ into ‘passed,’ or recommend approval.” Executives and release owners need compression without fiction. A useful summary says what evidence exists, what does not, and who owns the remaining decision. For the final gate, AI go/no-go decision prompts can structure the discussion. AI does not get a vote. Accountability cannot be delegated to autocomplete. A practical regression workflow The prompts work best as a sequence, not a vending machine: Confirm the deployed change set with technical owners. Map changes to critical workflows and dependencies. Apply human-approved risk labels. Compare proposed coverage with the existing suite. Add relevant historical defects and incidents. Confirm environments, safe data, accounts, and integrations. Assign execution, review, defect, and decision owners. Execute tests in approved tools and record evidence. Triage failures without inventing root cause. Summarize limitations and send the evidence to the human release gate. Use AI QA checklist prompts for broader quality review around the release. Regression is one layer, not a magic shield against unclear requirements, poor monitoring, inaccessible environments, weak security review, or missing rollback plans. Common mistakes with AI-generated regression plans Treating suggestions as verified dependencies A model may infer that account settings affect billing because those concepts often relate. Your system may work differently. Require sources and mark hypotheses visibly. Claiming “full coverage” Coverage is always bounded by scope, evidence, environments, data, time, and known risks. State those boundaries. “All tests passed” means little if half the meaningful scenarios were never run. Copying the entire ticket system into a chatbot Do not trade quality assurance for a privacy incident. Redact aggressively, summarize safely, and use approved enterprise tooling where required. Automating stale tests faster An old assertion can pass while checking behavior nobody values. Review whether each regression check still maps to a current requirement, journey, risk, or proven failure pattern. Letting AI make the release call A model has no operational authority, customer context, legal responsibility, or skin in the game. It can format the decision packet. A named human accepts or rejects the risk. Frequently asked questions Can AI create a complete regression test suite? No. It can draft candidate coverage from supplied information, but it cannot know undocumented behavior, inspect the real system, confirm dependencies, execute tests, or prove completeness. Humans must validate scope, run checks, review evidence, and own decisions. What should I give AI before asking for regression tests? Provide a verified change inventory, acceptance criteria, dependency evidence, critical journeys, prior defects, existing test references, environment limits, safe data rules, and human-approved risk labels. Remove secrets and sensitive data. How is regression testing different from smoke testing? Smoke testing is a narrow confidence check that a build is stable enough for further testing or basic operation. Regression testing checks whether changed and adjacent areas still preserve important existing behavior. A smoke pass is not proof of broad regression coverage. Can ChatGPT choose which regression tests to run? It can rank candidates using criteria and ratings you supply. Humans should approve the prioritization because business impact, architecture, customer commitments, and acceptable risk are contextual judgments. Should past bugs always become regression tests? Not automatically. Add or retain a check when the failure remains plausible, important, and testable. Remove or revise checks when the underlying behavior, architecture, or risk no longer exists. Keep the rationale visible. Is it safe to paste production logs into an AI tool? Usually not without explicit approval and proper controls. Logs may contain credentials, customer data, internal identifiers, vulnerabilities, or regulated information. Use redacted excerpts, synthetic examples, and controlled systems. Who approves an AI-assisted regression plan? The accountable people responsible for product behavior, engineering, QA, operations, security, compliance, accessibility, customer impact, and release management as appropriate. The model approves nothing. What if the regression window is too short? Make the tradeoff explicit. Prioritize critical journeys and high-impact risks, document exclusions, strengthen monitoring and rollback readiness, and require the authorized release owner to accept or reject residual risk. Do not disguise missing coverage as efficiency. The useful boundary AI is good at turning messy verified inputs into a first draft, joining lists, formatting matrices, surfacing blanks, and asking annoyingly useful questions. It is bad at knowing the secret dependency nobody documented, recognizing a misleading test environment, or carrying accountability when customers discover the failure. Use it for speed. Keep humans on facts, execution, evidence, judgment, privacy, specialist review, and sign-off. That division of labor is less exciting than “autonomous QA,” which is exactly why it is more likely to work. For a broader field guide to using AI without surrendering your judgment, Don’t Replace Me by Dmitry Kargaev covers the same core idea: the advantage is not pretending the tool is smart. It is knowing where fast assistance ends and human responsibility begins. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance AI API Testing Prompts: Catch Broken Integrations AI API testing prompts for contracts, edge cases, authorization, retries, error responses, and human-verified integration evidence. AI promptsAPI testingquality assurance AI Compatibility Testing Prompts: Catch Breakage 10 AI compatibility testing prompts for browser, device, OS, responsive, locale, network, defect triage, and evidence-based release decisions. AI promptscompatibility testingcross-browser testing --- ## AI Release Notes Prompts: Explain What Changed URL: https://dontreplace.me/articles/ai-release-notes-prompts-explain-what-changed AI promptsrelease notesproduct communicationchangelogdocumentation AI Release Notes Prompts: Explain What Changed 10 copy-paste AI prompts for writing accurate release notes. Turn changelogs into clear customer communication without inventing shipped features. July 18, 2026 Updated July 18, 2026 Release notes are where teams go to lie to themselves. Not maliciously. Just gradually. The engineer says "it's done." The PM writes something vague in Jira. Someone pastes it into a Google Doc. Three people edit it into mush. The support team reads it on launch morning and has no idea what changed, who it affects, or what to tell an angry customer. AI release notes prompts can help you break that cycle. But only if you're feeding the model real shipped changes, not aspirational changelogs and internal wishful thinking. This is the piece that covers both: the practical prompts that actually work, and the safety rails that stop you from publishing something embarrassing or dangerous. Why most release notes are bad before AI even enters the room Bad release notes start with bad inputs. That's it. That's the whole diagnosis. The changelog has three lines. Two of them are "backend improvements." The third is a ticket number with no description. Someone asks the engineer what shipped and gets a Slack message that says "basically the thing we discussed." Nobody knows the rollout scope. Nobody confirmed the bug fix in production. Legal hasn't seen the new terms mention. Support doesn't know the edge case. Then someone opens ChatGPT, pastes the mess in, and asks it to "write release notes." The model does exactly what it's designed to do: it makes the chaos sound coherent. It fills gaps with plausible-sounding language. It writes with confidence about things that were never verified. The result looks professional. It reads cleanly. It's wrong in three places you won't catch until a customer does. Rule #5 from Don't Replace Me is relevant here: AI is not smart, it's fast. Speed is not the same as accuracy. A model that confidently writes "this release improves load times by 40%" based on nothing but your vague changelog note is doing exactly what it was built to do. You're the one who's supposed to catch it. What to put in the prompt before you ask for anything Every prompt in this piece uses the same foundation. Fill this out before you paste anything: Shipped changes: actual merged tickets, feature flags confirmed off, bug fixes confirmed in production Release date: the real one, not the hoped-for one Rollout scope: 100% of users? Staged rollout? Beta only? Enterprise tier only? Affected user segments: which customers actually see this change Source tickets or screenshots: the receipts that confirm something shipped Known limitations or caveats: things that don't work yet, edge cases, workarounds Support instructions: what the support team needs to tell customers Approval status: who still needs to sign off If you can't fill in most of that, don't write release notes yet. You don't have a release. You have a plan. One hard rule before the prompts: do not paste customer PII, credentials, private tickets, unreleased strategy, security vulnerabilities, contracts, legal disputes, financial forecasts, HR data, regulated information, or private client conversations into any AI tool that hasn't been approved by your security and privacy teams. This includes ChatGPT, Claude, and everything else. Scrub the inputs first. Prompt 1: Turn a messy changelog into release notes Use this when you have a raw list of changes that needs to become customer-facing language. "Here is a raw changelog from our most recent release: [paste cleaned changelog]. The release date is [date]. This affects [user segment]. Convert this into release notes suitable for [blog post / in-app notification / email]. Use plain language. Do not add features or improvements that are not listed. Do not soften or omit known limitations. Flag any items where the shipped status is unclear and I need to verify before publishing." The last two sentences are the important ones. Tell the model to flag uncertainty rather than paper over it. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Prompt 2: Separate customer-facing changes from internal noise Half of what's in a typical changelog is infrastructure work, refactoring, dependency updates, and internal tooling that customers don't care about and shouldn't see. "Here is our full internal changelog: [paste]. Separate this into two lists: (1) changes that are directly relevant to end users and should appear in customer-facing release notes, and (2) internal or infrastructure changes that are not relevant to customers. For each item in list 1, write one sentence explaining the customer benefit. For list 2, just categorize without explanation." This is one of those prompts where AI is genuinely useful because it's a sorting and translation task, not a judgment call about what shipped. Prompt 3: Rewrite technical updates in plain English Engineers write for engineers. That's fine internally. It's a problem when those notes go directly to customers. "Rewrite the following technical release note in plain English for a non-technical audience: [paste note]. The reader is a [describe persona: e.g., 'marketing manager at a mid-sized SaaS company']. Explain what changed, why it matters to them, and what (if anything) they need to do. Keep it under 60 words. Do not include implementation details or internal references." Run this for anything involving API changes, database migrations, security patches explained at the wrong level, or infrastructure shifts that have a customer-visible effect. Then have a human who knows the customer read it before it goes anywhere. Prompt 4: Summarize bug fixes without overpromising Bug fix summaries are where teams accidentally make commitments they can't keep. "This issue has been resolved" is a promise. Make sure it's actually resolved before AI writes it that way. "Here are the bug fixes included in this release: [list with ticket numbers and descriptions]. Write a bug fix summary for customers. For each fix, describe the symptom the customer experienced (not the technical cause), confirm it is resolved, and note any conditions where the issue may still occur. Do not write 'this has been fully resolved' for any item unless I've marked it as confirmed in production. Flag any item where I haven't indicated production confirmation." Check the AI QA checklist prompts if you need to confirm what "production-confirmed" actually means for your team before writing any of this. Prompt 5: Create audience-specific versions One release, multiple audiences. Your enterprise customers need different language than your self-serve users. Your internal sales team needs different language than your support team. "Here are the core release notes for [release name/version]: [paste]. Create three versions: (1) a customer-facing summary for [describe segment] that focuses on user benefits, (2) an internal version for the support team that includes known issues, edge cases, workarounds, and escalation guidance, and (3) a one-paragraph summary for internal stakeholders covering what shipped, what didn't make the cut, and what to watch post-launch. Do not invent information. If information needed for one version isn't in the source, note what's missing." Prompt 6: Add known limitations and caveats This is the prompt most teams skip. Don't. "Here are our release notes draft: [paste]. Here are the known limitations and caveats we haven't included yet: [list]. Integrate these limitations clearly and honestly into the release notes. Don't bury them at the bottom or soften the language. If a limitation significantly affects a specific user segment, make that clear. Suggest where in the document each caveat belongs based on relevance." Burying a known limitation in release notes isn't a communications strategy. It's a support ticket factory. AI can help you write the caveat language clearly. You're the one who decides it goes in. That's a taste judgment. It's yours to make. Prompt 7: Check support and customer success readiness Before anything goes out, your support team should be able to answer questions about it. This prompt checks whether the notes you've drafted actually give them what they need. "Here are the release notes I'm planning to publish: [paste]. Act as an experienced support agent reading this for the first time. List: (1) questions a customer is likely to ask after reading this that aren't answered in the notes, (2) edge cases that aren't addressed, (3) anything a support agent would need to know to handle customer inquiries that isn't mentioned here. Do not invent product knowledge. Flag gaps as gaps." This one pairs well with the AI stakeholder update prompts for getting the support team briefed before launch, not after the first customer complaint. Prompt 8: Map rollout status and affected users If something is a staged rollout, partial release, or feature-flag-gated, your release notes need to say so clearly. This prompt helps you make that section of the document accurate. "Here is our rollout plan for [feature/fix]: [describe rollout: e.g., '10% of users starting [date], full rollout by [date], enterprise accounts first']. Here are the affected user segments: [list]. Write a rollout status section for our release notes that explains clearly who has access now, who gets it and when, and what customers should do if they don't see the change yet. Do not guess at rollout percentages. Use only the figures I've provided." Vague rollout language generates support tickets. "Gradually rolling out" with no dates or percentages is not information. It's a placeholder that customers hate. Prompt 9: Draft an approval checklist Release notes need sign-off. From whom depends on what's in them. Use this to build the checklist before anything goes live. "Here is the draft of our release notes: [paste]. Based on the content, create an approval checklist that identifies which stakeholders need to review this before publication. Consider: product, engineering, QA, legal, security, privacy, marketing, support, customer success leadership, and executive review. For each stakeholder, note why their review is needed based on the content. Flag any items that involve regulated data, security changes, pricing changes, terms of service updates, or anything with legal or compliance implications." This is the prompt that saves you from a legal problem you didn't know was in there. Have a human make the final call on escalations. The AI launch checklist prompts cover the broader launch side of this if you need the full picture. Prompt 10: Final publication sanity check Last check before it goes out. "Here are the final release notes ready for publication: [paste]. Run a sanity check against the following criteria: (1) every claim is verifiable from the source material I've provided, (2) no features are mentioned that I haven't confirmed shipped, (3) no customer segment is affected that I haven't listed, (4) known limitations are present and clear, (5) rollout status is accurate, (6) no internal-only language, ticket numbers, or team references appear in customer-facing sections, (7) support and customer success have the information they need. List anything that fails these checks and explain why." This one you run, then a human reads the output, and you fix what it flags. Not the other way around. The AI documentation prompts and AI project closeout prompts have related templates if you're building out a broader documentation habit. What AI can't do in the release notes process AI can structure, translate, sort, and gap-check. It can't tell you whether something actually shipped. It can't confirm a bug is fixed in production. It can't judge whether a caveat is serious enough to mention or a limitation significant enough to delay the release. It can't read the room on what your enterprise customers will escalate. The prompts above are useful because they constrain what the model invents. Tell it to flag uncertainty. Tell it not to fill gaps. Tell it to use only what you've provided. That shifts AI from a fabrication machine into a communication assistant, which is what it's actually good at. The people who get in trouble with release notes are the people who mistake polished language for verified truth. Polish is easy now. Verification still belongs to you. Frequently asked questions What should I include in an AI release notes prompt to get accurate output? Include confirmed shipped changes with ticket references, release date, rollout scope, affected user segments, known limitations, and support instructions. Tell the model explicitly not to fill gaps or invent shipped behavior, and ask it to flag anything that's unclear rather than writing around it. Can AI write release notes from a Jira ticket dump? It can generate a draft, but the draft needs human review before publication. AI will make unclear or ambiguous tickets sound definitive. You need someone who knows what actually shipped to confirm the output reflects reality, not just the ticket description. What shouldn't I paste into AI when writing release notes? Don't paste customer PII, credentials, security vulnerabilities, unreleased roadmap strategy, contracts, legal disputes, financial forecasts, regulated data, HR information, or private client conversations into any AI tool that hasn't been security-approved by your organization. How do I write release notes that support teams can actually use? Use Prompt 7 in this article: ask AI to read your draft as a support agent and list customer questions it doesn't answer. Also write a separate internal version for support with edge cases, workarounds, and escalation guidance that wouldn't go in the public-facing notes. Who needs to approve release notes before they go out? At minimum: product to confirm shipped status, engineering to confirm technical accuracy, QA to confirm bug fixes, and support to confirm readiness. Add legal for terms changes, security for patches, marketing for customer-visible messaging, and executive review for anything that could generate customer escalations. What's the difference between a changelog and release notes? A changelog is an internal record of what changed, usually written by engineers and linked to tickets. Release notes are a customer-facing communication that explains what changed, why it matters, and what customers need to do. AI is useful for translating one into the other, but the human who knows what actually shipped needs to check the translation. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance --- ## AI Remediation Plan Prompts: Fix Root Causes URL: https://dontreplace.me/articles/ai-remediation-plan-prompts-fix-root-causes AI promptsremediationcorrective actionQAoperations AI Remediation Plan Prompts: Fix Root Causes AI remediation plan prompts for turning findings, defects, incidents, owners, deadlines, fixes, and retest evidence into real corrective action. August 03, 2026 Updated August 03, 2026 A remediation plan is not a spreadsheet where problems go to wear a blazer and die. It is the boring, necessary plan for fixing the actual thing that broke, proving it is fixed, and making sure nobody has to rediscover the same failure next month while everyone says, “I thought we handled this.” That is where AI remediation plan prompts help. Not because AI can inspect production, approve security fixes, verify legal commitments, or magically know the root cause from five angry screenshots. It cannot. But it can turn defect triage notes, incident findings, audit gaps, customer-impact summaries, workaround lists, owner rosters, deadlines, risk notes, regression concerns, and retest evidence into a remediation plan that looks less like panic origami. If you are still sorting incoming issues, start with AI defect triage prompts. If the failure already happened and now someone needs a real fix, this guide is for you. AI can structure remediation work. A human still owns the evidence, risk, approvals, fix quality, verification, and final sign-off. What is a remediation plan? A remediation plan is a practical corrective-action plan for resolving a confirmed problem, reducing its risk, and proving the fix worked. A useful remediation plan answers: What happened? What evidence proves it happened? What is the confirmed impact? What is the likely root cause, and what is still only a hypothesis? What temporary workaround exists? What permanent fix is approved? Who owns each action? What dependencies, deadlines, and risks matter? How will the team verify the problem is actually fixed? What evidence will be saved for audit, customer communication, or future review? Without that structure, remediation becomes theater. Someone renames the ticket “postmortem follow-up,” adds a due date, tags three people who are on vacation, and calls it accountability. Beautiful. Useless, but beautiful. AI remediation plan prompts are useful because AI is good at structure. It can draft tables, summarize sanitized findings, separate symptoms from likely causes, list missing evidence, convert workarounds into permanent-fix options, map owners, write stakeholder updates, and create verification checklists. But it is the same rule as what AI can and can't do: AI can make the mess easier to inspect. It cannot make the fix real. The reusable AI remediation plan prompt formula Use this formula when asking AI to help with remediation planning: “You are a remediation planning assistant. I am [your role] creating a remediation plan for [product/process/system/customer issue/audit finding]. Here is sanitized context: [confirmed findings, evidence summaries, impact, affected systems, defect or incident notes, current workaround, possible causes, owners, deadlines, severity rules, approved constraints, retest evidence, audit requirements]. Create [specific output]. Separate confirmed facts from assumptions. Flag missing evidence, owner questions, approval needs, security/privacy concerns, customer-impact concerns, dependency risks, and verification gaps. Do not invent root cause, impact, approvals, owners, production status, due dates, customer commitments, legal/security conclusions, or final decisions.” That last sentence is doing actual work. AI likes to fill blanks because a filled blank looks helpful. In remediation, a filled blank can become a fake cause, a fake fix, or a fake promise to a customer. Congratulations, you shipped liability with bullet points. Before using any prompt, sanitize the input. Do not paste customer PII, credentials, access tokens, raw production logs, private tickets, unreleased strategy, legal disputes, security vulnerabilities, financial records, HR issues, regulated data, medical information, private client conversations, or sensitive personal information into unapproved AI tools. Use summaries, categories, redacted excerpts, approved screenshots, internal ticket IDs, and links to authorized systems instead. The robot does not need the customer’s email address to help you write a better action tracker. What to bring before asking AI for a remediation plan AI cannot remediate a mystery. It can only organize the ingredients you provide. If the ingredients are vague, stale, or politically massaged, you get a glossy plan for solving the wrong problem. Bring this first: InputWhy it mattersHuman check Confirmed findingKeeps the plan tied to a real problemVerify the wording against evidence Evidence summaryPrevents vibes-based remediationLink to approved records, not raw secrets Affected system or processDefines the fix boundaryConfirm scope with the owner Customer or user impactGuides urgency and communicationUse approved, privacy-safe summaries Current workaroundSeparates temporary relief from permanent fixMark approved vs. proposed clearly Root-cause hypothesisHelps plan investigationLabel hypothesis until confirmed Owners and backupsMakes work movableUse real teams, queues, or names DependenciesExposes blockers earlyConfirm whether another team must act Retest criteriaProves the fix workedDefine evidence before closing Audit trail needsSaves future painCapture dates, decisions, approvals, and proof If you do not have these inputs, ask AI to create a missing-information checklist first. Do not ask it to declare a fix from a Slack thread and one screenshot named final-final-actually-this-one.png. For better upstream issue sorting, pair this with AI bug report prompts and AI QA checklist prompts. Good evidence makes good remediation. Bad evidence makes premium nonsense. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI remediation plan prompts you can use today Prompt 1: Turn findings into a remediation table You are a remediation planning assistant. I am creating a remediation plan for [system/process/customer issue/audit finding]. Here are sanitized findings, evidence summaries, affected systems, impact notes, workaround status, owners, deadlines, and constraints: [paste details]. Create a remediation table with columns for finding, confirmed evidence, impact, temporary workaround, permanent fix option, owner, dependency, due date question, verification method, missing evidence, and approval needed. Separate confirmed facts from assumptions. Do not invent root cause, impact, owner, fix approval, or due date. What to check after: Make sure the table does not turn “possible cause” into “confirmed cause.” That tiny wording change is how teams build monuments to the wrong problem. Prompt 2: Separate symptoms from root-cause hypotheses You are helping me organize remediation notes. Here are sanitized symptoms, incident notes, defect reports, customer-impact summaries, logs summarized safely, and team comments: [paste details]. Create three sections: confirmed symptoms, likely root-cause hypotheses, and missing evidence needed to confirm or reject each hypothesis. For every hypothesis, list what evidence would support it, what evidence would contradict it, and who should verify it. Do not claim a root cause is confirmed unless the evidence explicitly says so. What to check after: Root cause is not “the thing everyone is annoyed about.” It is the thing evidence supports. Tiny distinction. Massive difference. Prompt 3: Draft corrective-action options You are a corrective-action planning assistant. Based on this sanitized finding, impact summary, current workaround, constraints, owner list, risk notes, and affected systems: [paste details], propose three remediation options: minimum safe fix, durable fix, and preventive fix. For each option, list expected benefit, risk, dependency, verification need, rollback consideration, owner question, and approval required. Do not recommend legal, security, financial, HR, or customer commitments as final; flag them for expert review. What to check after: AI may make the “durable fix” sound easy because it does not have to attend the architecture review. Add reality back in. Prompt 4: Map owners and due dates without fantasy project management You are helping assign remediation work. Here are sanitized action ideas, teams involved, current owners, backup coverage, constraints, deadlines, dependencies, and approval paths: [paste details]. Create an owner map that includes action, accountable owner, supporting teams, backup owner, dependency, due-date question, decision needed, and escalation path. Flag any action with no real owner, no backup, unclear approval, or impossible deadline. Do not invent names, commitments, or dates. What to check after: “Engineering” is not an owner. “Support” is not an owner. “Someone should” is a haunted house. Prompt 5: Convert a workaround into a permanent-fix plan You are a remediation assistant. Here is our sanitized workaround, why it was approved, what it protects, what it does not protect, affected users, known risks, owner notes, and possible permanent fixes: [paste details]. Create a plan that separates temporary workaround steps from permanent-fix actions. Include risks of leaving the workaround in place, evidence needed before removing it, communication needs, owner questions, and verification criteria. Do not imply the workaround is the fix. What to check after: Workarounds are debt with a safety vest. Useful, yes. Permanent, no. Prompt 6: Check dependencies and blockers You are reviewing a remediation plan for hidden blockers. Here is the sanitized draft plan, owners, dependencies, affected systems, approval paths, release windows, test needs, customer-impact notes, and risk constraints: [paste details]. Find blockers, missing decisions, dependency risks, sequencing problems, approval gaps, and retest gaps. Return a prioritized blocker list with the question to ask, who should answer it, and what happens if it remains unresolved. Do not invent answers. What to check after: This is where AI is genuinely useful. It is annoyingly good at noticing that your “simple fix” depends on four teams, a config freeze, and an environment nobody owns. Prompt 7: Prepare a stakeholder remediation update You are drafting a stakeholder update about remediation work. Here is sanitized context: confirmed finding, impact, current status, workaround, owner map, next actions, risks, dependencies, decisions needed, and verification plan: [paste details]. Draft a concise update for [audience]. Use plain language. Separate confirmed facts from open questions. Include what changed, what is being done, what is blocked, what decision is needed, and when the next update will happen. Do not overpromise, assign blame, invent certainty, or disclose sensitive details. What to check after: Stakeholder updates should reduce panic, not sedate people with fog. If the fix is not verified, say it is not verified. If the remediation came from a messy launch, AI post-launch monitoring prompts can help structure the ongoing signal without turning every blip into a bonfire. Prompt 8: Build a verification and retest checklist You are a QA and remediation verification assistant. Here is the sanitized finding, affected workflow, proposed fix, current workaround, acceptance criteria, affected environments, release notes, retest evidence, and rollback criteria: [paste details]. Create a verification checklist that includes setup, test data needs, happy path, edge cases, regression checks, evidence to capture, owner, environment, version, pass/fail criteria, and sign-off question. Do not claim the fix passed unless the evidence says it passed. What to check after: A fix without retest evidence is a hope with a ticket number. Pair this with AI smoke test prompts when the remediation affects launch-critical paths. Prompt 9: Write audit-ready evidence notes You are helping document remediation evidence. Here are sanitized details: finding, decision history, owner actions, approvals, implementation notes, retest results, screenshots described safely, dates, ticket IDs, and remaining risks: [paste details]. Create an audit-ready evidence note with sections for original finding, corrective action taken, evidence reviewed, verification result, approvals, remaining risks, and final sign-off needed. Keep it factual. Do not invent evidence, approvals, timestamps, or compliance conclusions. What to check after: Audit-ready does not mean “sounds official.” It means a future human can trace what happened without interviewing eight ghosts. Prompt 10: Create the final remediation decision log You are creating a remediation decision log. Here is sanitized context: problem, evidence, impact, root-cause status, chosen fix, rejected options, owner decisions, approvals, verification results, stakeholder updates, and open risks: [paste details]. Create a decision log with date, decision, rationale, evidence, owner, approver, alternatives rejected, risks accepted, follow-up action, and final sign-off status. Separate facts from judgment. Do not invent approval, risk acceptance, or closure. What to check after: The decision log is where future-you finds out whether past-you was brave, careful, rushed, or just caffeinated and doomed. If this remediation followed a bigger failure, use AI incident review prompts to keep the learning clean instead of turning the review into blame karaoke. A simple remediation plan template Use this structure when you want one practical document instead of a pile of comments: Problem statement: What happened, in one plain paragraph. Confirmed evidence: What proves the problem exists. Impact: Who or what was affected, using approved summaries. Scope: Systems, workflows, customers, environments, versions, or teams involved. Current workaround: Temporary relief, owner, risk, and expiration condition. Root-cause status: Confirmed cause, hypothesis, or still unknown. Corrective actions: Permanent fixes, owners, dependencies, and due-date questions. Preventive actions: What changes so this does not become a subscription. Verification plan: Tests, evidence, environments, sign-off, and retest schedule. Communication plan: Who needs updates, what can be said, and who approves it. Decision log: What was chosen, rejected, accepted, escalated, and closed. AI can help draft each section. The human has to keep the document honest. Where AI is useful, and where it gets dangerous Use AI for: Turning messy notes into a structured plan. Finding missing owners, dependencies, and evidence. Drafting privacy-safe stakeholder updates. Creating retest and verification checklists. Separating confirmed facts from assumptions. Comparing temporary workaround vs. permanent fix. Writing audit-ready summaries from approved evidence. Do not use AI to: Invent root causes. Decide customer impact without evidence. Approve security, legal, HR, financial, or compliance conclusions. Paste raw private data into unapproved tools. Replace engineering, QA, support, security, or executive judgment. Close remediation without verification. Decide risk acceptance because the generated table looked confident. This is the same pattern as AI rollback plan prompts and AI contingency plan prompts: AI helps you think through options. It does not absorb the consequences. Frequently asked questions Can AI write a remediation plan for me? Yes, if you give it sanitized, accurate context and treat the output as a draft. It can structure the plan, organize owners, identify missing evidence, and draft checklists. It cannot verify the fix, approve the plan, or know whether the root cause is true unless the evidence is actually there. What should I never paste into AI tools? Do not paste customer PII, credentials, access tokens, raw production logs, private tickets, unreleased strategy, legal disputes, security vulnerabilities, financial records, HR issues, regulated data, medical information, private client conversations, or sensitive personal information into unapproved AI tools. Use redacted summaries and approved systems. How is a remediation plan different from an incident review? An incident review explains what happened and what the team learned. A remediation plan turns those findings into corrective actions, owners, dependencies, verification steps, and closure evidence. You usually need both. One without the other is how teams become recurring characters in their own disaster movie. Can AI find the root cause? AI can suggest hypotheses and organize evidence, but it should not declare a root cause from incomplete notes. Root cause needs verified evidence, expert review, and often testing. Ask AI what evidence would confirm or reject each hypothesis instead of asking it to sound certain. How do I know a remediation plan is done? A remediation plan is done when the approved fix is implemented, retested, documented, communicated where needed, and signed off by the right humans. “Ticket closed” is not proof. “We think it is better” is not proof. Save the evidence. What if the fix creates new risk? Then the remediation plan should say that. Add dependency risks, rollback criteria, regression checks, owner questions, and approval needs. If the change affects launch-critical workflows, support coverage, customer commitments, or security posture, route it through the right review before pretending the plan is tidy. Final thought AI remediation plan prompts are not magic. They are scaffolding. They help you turn a pile of findings, arguments, screenshots, half-remembered meetings, workaround notes, and owner shrugs into a plan a real team can inspect. That is useful. It is also not the same as fixing the problem. The job is still human: verify the evidence, choose the right fix, protect sensitive data, communicate honestly, retest the result, and own the call. If you want the broader survival guide for staying useful while the workplace fills with confident autocomplete and prettier dashboards, read Don’t Replace Me by Dmitry Kargaev. It is basically a field manual for remembering that accountability is not something you can outsource to a prompt. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Change Impact Analysis Prompts: Spot Damage AI change impact analysis prompts for mapping affected systems, teams, customers, QA checks, support gaps, and rollout risks. AI promptschange impactrisk AI Root Cause Analysis Prompts: Stop Fixing Symptoms AI root cause analysis prompts for turning messy incidents, bugs, timelines, evidence, and corrective actions into clearer RCA work. AI promptsroot cause analysisRCA AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management --- ## AI Requirements Gathering Prompts: Stop Starting Vague URL: https://dontreplace.me/articles/ai-requirements-gathering-prompts-stop-starting-vague AI promptsrequirements gatheringproject managementAI at workproduct requirements AI Requirements Gathering Prompts: Stop Starting Vague 10 copy-paste AI requirements gathering prompts to turn vague asks into verified briefs. Includes safety rules and a reusable prompt formula. July 11, 2026 Updated July 11, 2026 The project brief says "build a form." That's it. No context, no constraints, no stakeholders named, no definition of done. You paste it into ChatGPT and get back two pages of beautifully formatted AI requirements gathering prompts output that sounds completely authoritative and is about 40% invented. Now you're in a kickoff meeting reading requirements nobody agreed to, and everyone's nodding politely because the document looks so official. This is the trap. AI is very good at making incomplete information sound finished. It fills gaps with plausible-sounding assumptions and presents them with the same confidence it uses for things that are actually true. Tidy does not mean accurate. That's Rule #5 from Don't Replace Me: It's Not Smart. It's Fast. The prompts below are built around a different approach. You bring the real inputs. AI helps you structure, question, and pressure-test them. You own the judgment calls. Why most requirements documents fail before you start Bad requirements aren't usually the result of laziness. They're the result of starting to write before you've finished listening. Someone has a vague ask, you capture the surface version of it, and then everyone spends three sprints building the wrong thing with increasing confidence. AI makes this worse, not better, if you use it wrong. Paste a messy Slack thread into any LLM and it will give you a polished requirements brief with clear sections and bullet points. Half the bullets will be extrapolated from context clues. The other half will be the AI guessing what "sounds right" for a project like yours. It won't flag which ones it made up. The prompts in this guide force you to separate what you actually know from what you're assuming. That gap, named explicitly, is where requirements work actually happens. The reusable formula before any requirements prompt Every prompt here follows the same structure. Get this right once and the rest is copy-paste. Role: Act as a requirements analyst reviewing a project brief. Input: Here is what I have: [paste verified source material only] Task: [specific task from the prompts below] Constraints: Flag every assumption you make. If something is unclear, ask the question rather than fill the gap. Output should distinguish between confirmed facts and open questions. That last constraint is the one that matters. Without it, AI will happily complete the picture for you. Safety note before you start: Do not paste customer PII, employee records, credentials, contracts, legal disputes, financial forecasts, unreleased product details, security vulnerabilities, confidential strategy, or sensitive client data into any AI tool that hasn't been approved by your security or legal team. If the source material contains any of that, remove it before prompting. Use placeholder text or anonymized descriptions instead. 10 AI requirements gathering prompts you can use today Prompt 1: Turn a vague ask into a requirements brief Use when: Someone handed you a one-liner and called it a project. Act as a requirements analyst. Here is the initial request I received: [paste the ask]. Turn this into a draft requirements brief with these sections: - Problem statement (what we know for certain) - Proposed solution (as described, no additions) - Known constraints - Assumptions being made - Open questions that must be answered before this can proceed Label every bullet as CONFIRMED or ASSUMED. Do not invent scope. Prompt 2: Extract requirements from messy notes Use when: You have a recording transcript, a Slack dump, or six pages of meeting notes. I have raw notes from a stakeholder conversation. Extract only what was explicitly stated. Do not infer, extrapolate, or complete the picture. Notes: [paste sanitized notes] Output: - Explicitly stated requirements - Explicitly stated constraints - Explicitly stated preferences (not requirements) - Topics that came up but were not resolved - Things I should go back and clarify If you're working from meeting notes regularly, the AI meeting notes prompts guide has dedicated templates for that step. Prompt 3: Identify missing stakeholder inputs Use when: You have a brief but you're not sure who else should have been consulted. Here is a draft requirements document: [paste document] Based on the scope described, identify: - Which stakeholder groups likely have requirements not represented here - Which teams will be affected by this work (technical, legal, security, privacy, accessibility, finance, operations, customer-facing) - What questions each missing group would probably ask - Which of those gaps could block delivery if not resolved Do not invent requirements. Only flag what's missing. Prompt 4: Separate facts from assumptions Use when: A requirements doc already exists but you're not sure how much of it is real. Review this requirements document and sort every item into one of three categories: CONFIRMED: Stated explicitly by a named source ASSUMED: Inferred or extrapolated without verification UNKNOWN: Requires research or stakeholder input to resolve Document: [paste requirements] For each ASSUMED or UNKNOWN item, write the specific question someone would need to answer to confirm it. This is the prompt most people skip. It's the most useful one. Prompt 5: Map users and use cases Use when: The brief mentions "users" without specifying who they actually are. Based on this project description, help me map the user landscape: Description: [paste brief] For each user type you can identify from the description: - Who they are (role or context, not invented demographics) - What they're trying to accomplish - What constraints or access levels they likely have - What could go wrong for them specifically Flag every item as INFERRED if it's not explicitly stated. Note which user types need direct research. Prompt 6: Find dependencies and constraints Use when: You need to scope what this project bumps into before you start. Review this requirements brief and identify: - Technical dependencies (systems, APIs, platforms this touches) - Team dependencies (who else has to do something for this to work) - Timeline dependencies (what has to happen first) - Compliance or regulatory constraints (data, accessibility, privacy, security) - Constraints that were implied but not explicitly stated Brief: [paste brief] For each dependency, note whether it is CONFIRMED or INFERRED and who owns the resolution. For deeper work on this, the AI workflow audit prompts are useful for surfacing process-level blockers your brief might be missing. Prompt 7: Surface risks and edge cases Use when: The requirements look clean but something feels off. Here are the requirements for a project: [paste requirements] Identify: - Edge cases that are not addressed - User behaviors the requirements assume won't happen but probably will - Technical failure modes not covered - Accessibility or inclusion gaps - Privacy or data handling risks - Scenarios where the stated requirements conflict with each other For each risk, note its likely severity (high/medium/low) and whether it needs legal, security, privacy, or engineering review. Before you get here, AI risk assessment prompts can help you run a broader blast radius check on the overall project. Prompt 8: Prepare stakeholder interview questions Use when: You have a draft brief and you're about to talk to real humans who will fill the gaps. I'm preparing to interview stakeholders for this project: [paste brief] Generate interview questions designed to: - Validate or challenge the assumptions in the brief - Surface constraints I haven't captured - Clarify decision rights (who approves what) - Understand success criteria in their words, not ours - Identify what they're worried about that isn't in the document Group questions by stakeholder type: business owner, end users, technical team, legal/compliance if relevant. Prompt 9: Rewrite requirements for different audiences Use when: You need the same requirements to work for a technical team and a business sponsor. Here are the project requirements: [paste requirements] Rewrite them twice: Version 1, For the technical team: Include implementation context, system references, data handling specifics, edge case handling, and acceptance criteria language. Version 2, For the business sponsor: Focus on outcomes, business impact, user experience, and decision points without technical jargon. Do not add new requirements in either version. Only change the framing and language. If the technical version needs to go further into acceptance criteria, the AI acceptance criteria prompts have dedicated templates for making requirements testable. Prompt 10: Run a pre-kickoff readiness check Use when: You think you're ready to kick off but you want to be sure. Here is our current requirements document for an upcoming project: [paste document] Run a pre-kickoff readiness check. Evaluate whether we have: - A clearly stated problem (not just a solution) - Named stakeholders with confirmed buy-in - Defined success criteria - Known constraints and dependencies - Open questions logged with owners and deadlines - Risk items flagged for specialist review (legal, security, privacy, accessibility) - Decision rights documented - Scope boundaries (what's out of scope) Rate each area as READY, NEEDS WORK, or BLOCKED. For anything that's not READY, say what's missing. This one pairs well with AI project kickoff prompts for the meeting itself. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What AI should never replace in requirements work AI requirements gathering prompts are useful for structure, questions, checklists, and language. They are not useful for the things that actually determine whether requirements are real. Those things are: Stakeholder conversations. AI can generate interview questions. It cannot sit across from a product owner and read the room when they say "yes, sure" while clearly meaning "I have no idea." User research. A language model can map plausible user types. It has never met your users and doesn't know what their actual context is. Political judgment. Which requirements will get cut in budget discussions. Which ones are negotiable but nobody will say so. Which stakeholder's opinion overrides the brief on paper. That's what Dee calls the taste moat in Don't Replace Me: the human value that doesn't show up in a document. Legal, security, privacy, and compliance review. If your requirements touch personal data, financial systems, health information, regulated industries, contracts, or security-sensitive infrastructure, a prompt is not a review. Get the actual experts involved. An AI-generated requirements brief does not constitute sign-off from anyone. Delivery ownership. Requirements signed by nobody change easily. Requirements with a named owner who has committed to them change through a process. AI can't create accountability. You have to. A note on confidence levels One practical habit: when you finalize a requirements brief, add a confidence field to each requirement. Not complicated. Just: HIGH (confirmed by named source), MEDIUM (inferred but plausible), LOW (assumed, needs verification). This forces you to have a conversation about the LOW items before the kickoff, not in a sprint retro six weeks later. It also makes it easier to log decisions when things change, which they will. The AI decision log prompts can help you track those downstream. The goal of requirements gathering is not a polished document. It's shared understanding. AI can help you write that document faster. It can't create the understanding for you. Frequently asked questions Can AI write requirements documents for me? AI can draft a requirements structure from your inputs, but it will fill gaps with plausible-sounding assumptions that may not reflect what stakeholders actually need. Use it to organize and question your material, not to complete it. Every ASSUMED item in an AI-generated brief needs a human to verify it before you build anything. What's the biggest mistake people make with AI requirements gathering prompts? Pasting a vague brief and accepting the output as real requirements. AI will produce something that looks complete even when the source material is missing half the information. The prompts in this guide are designed to surface gaps, not hide them. What information should I never paste into an AI tool for requirements work? Customer PII, employee records, credentials, contracts, legal disputes, financial forecasts, unreleased product details, security vulnerabilities, medical or safety information, and anything your security or legal team hasn't approved for use in external AI tools. Anonymize or remove sensitive details before prompting. How do I handle requirements that need legal or compliance review? Flag them explicitly in the document and do not treat an AI-generated requirements brief as approval. Escalate to legal, security, privacy, accessibility, or finance as appropriate. The prompt for surfacing risks and edge cases (Prompt 7 above) will help you identify which items need that review. Do I need to tell stakeholders I used AI to draft the requirements? That depends on your organization's AI disclosure policies, and you should check them. What you definitely need to do is verify the requirements with stakeholders regardless of how you drafted them. A document drafted by AI that stakeholders haven't reviewed and confirmed is not a requirements document. It's a guess with good formatting. How does requirements gathering connect to project kickoff? Requirements work happens before kickoff, and a lot of kickoff chaos comes from starting with requirements that haven't been properly validated. If your brief has unresolved open questions, unknown owners, or unverified assumptions, those problems don't disappear when the kickoff meeting starts. The AI project management prompts cover the broader project structure these fit into. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Launch Checklist Prompts: Ship Without Gaps 10 copy-paste AI launch checklist prompts for ChatGPT and Claude. Turn messy plans into structured checklists with owners, risks, and rollback steps. ai promptslaunch checklistproject management AI Migration Plan Prompts: 10 Templates That Work 10 copy-paste AI migration plan prompts for system mapping, dependency checks, rollback criteria, cutover plans, and stakeholder comms. With safety reminders. AI promptsmigration planningsystem migration AI Assumption Log Prompts: Stop Planning on Guesswork Use 10 AI assumption log prompts to expose project guesswork, design evidence checks, assign review owners, and stop unverified beliefs becoming commitments. AI promptsproject managementassumption logs --- ## 10 AI Research Prompts That Actually Work URL: https://dontreplace.me/articles/ai-research-prompts-desk-research-without-lying-to-yourself ai research promptsdesk researchchatgpt researchai prompts for workresearch methodology 10 AI Research Prompts That Actually Work 10 copy-paste AI research prompts for desk research. Learn to summarize sources, extract claims, spot gaps, and build decision memos without hallucinations. June 03, 2026 Updated June 03, 2026 AI has a research problem. Not a capability problem. A confidence problem. Ask it a question and it answers immediately, fluently, and with the quiet certainty of someone who has never once said "I don't know." That's useful when you want a first draft. It's a trap when you need the truth. The difference between a good AI research prompt and a bad one isn't about clever wording. It's about whether you're using AI to find answers or to confirm the answers you already wanted. This guide gives you 10 copy-paste AI research prompts for desk research, plus the safety rules that stop you from publishing something embarrassing. The prompts are organized in order: from clarifying your question all the way through to building a decision memo. Use them in sequence or grab the one you need. Before we get into the templates, one rule that runs underneath all of them. The core mistake people make with AI research They treat AI like a search engine that gives better answers. It isn't. A search engine returns links. You decide which ones are credible. AI returns synthesized prose that sounds like the conclusion, skips the sources, and buries any uncertainty in confident-sounding language. That's great for organizing material you already have. It's dangerous when you're trying to discover something true. The mental model that actually works: AI is a fast, tireless intern. Exceptional at summarizing, comparing, formatting, and drafting. Terrible at knowing when it's wrong, citing real sources unprompted, or telling you that your original question was bad. What AI can and can't do is worth reading if you want the full breakdown. The short version: the judgment is yours. The legwork is theirs. Every prompt below is built on that split. You bring the sources, the context, and the question. AI organizes, compares, and drafts. You verify, decide, and sign your name to the conclusion. What not to paste into AI before we go any further This matters and people skip it. Do not paste: confidential company strategy, private customer data, unreleased product plans, proprietary research, personal medical or financial information, regulated data of any kind. That means no client lists, no patient records, no legal case details, no HR notes, no NDA-covered materials. Anonymize before you paste. "Our client, a regional hospital" becomes "a mid-size healthcare organization." Specific revenue numbers, internal projections, personal names from private contexts: leave them out or genericize them. And one more thing. AI will invent citations if you let it. It will give you a convincing author name, a plausible journal, a real-looking year. None of that means the source exists. Verify every citation, date, statistic, and quote against a primary source before it goes anywhere near something you're publishing or presenting. If you're unsure whether your workplace allows AI tool use at all, that's a separate conversation worth having. Whether to tell your boss you use AI has become a real workplace question, and the answer depends on your organization and what you're working on. The 10 AI research prompts, ready to copy Prompt 1: Turn a vague question into a research plan You know the general topic. You don't know what you're actually trying to find out. Use this first. Prompt: I'm researching [topic]. My goal is to [decision I need to make / output I need to produce]. Here's what I already know: [brief summary]. Help me define a focused research question, identify 5-7 subtopics I should investigate, and flag 3 assumptions I might be making that I should test. Don't answer the research questions yet. Just help me structure the inquiry. The last sentence is important. If you don't add it, the AI will just answer the questions. You want the plan first. Prompt 2: Build a source checklist Prompt: I'm researching [topic] for [purpose]. What types of primary and secondary sources should I look for? List them by category: official data, industry reports, academic research, expert commentary, practitioner examples. For each category, give me 2-3 specific places or organizations I might look. Flag which source types are likely to be biased toward a particular conclusion and why. This gives you a structured sourcing plan before you touch a single document. Run it, then go find the actual sources yourself. Don't ask AI to find sources for you without verifying they exist. Prompt 3: Summarize a provided article Prompt: Here is the full text of an article I'm researching: [paste text]. Summarize it in 150 words. Then list: the main claim, the evidence used to support it, the methodology or basis for the evidence (survey, study, expert opinion, case study), and any notable limitations or caveats the author mentions. Don't add any information from outside this article. That last instruction is critical. Without it, AI will happily blend the article with everything else it knows, and you'll lose track of what the source actually said. Prompt 4: Extract claims and evidence Good research separates claims from evidence. Most documents blur the line. Prompt: Here is a document I'm analyzing: [paste text]. Create a table with three columns: Claim, Evidence Provided, and Evidence Type (data/expert opinion/anecdote/analogy/assertion). List every significant claim the document makes. If a claim has no evidence, mark the Evidence Provided column as "None stated." This prompt alone is worth the price of the article. It makes bad arguments visible immediately. Prompt 5: Compare two sources Prompt: Here are two sources on [topic]: [paste Source A] [paste Source B]. Compare them across these dimensions: main argument, evidence quality, areas of agreement, areas of contradiction, and what each source seems to be missing. Present your comparison as a table. Do not synthesize a conclusion. Just show me the comparison. Again: no conclusion yet. You make the conclusion. The AI makes the table. Prompt 6: Find missing viewpoints This one catches the blind spot in your own research. Prompt: I'm researching [topic]. My current sources all come from [describe the type of sources you've used: industry reports, academic papers, news coverage, etc.]. Based on the topic, what perspectives, stakeholder groups, or types of evidence am I likely missing? Be specific. For each gap, explain why that perspective would matter to a complete picture of this issue. Dee covers the judgment problem in Don't Replace Me: the human work is deciding which sources matter, which assumptions are weak, and which conclusions are actually useful. Spotting your own gaps before someone else does is exactly that kind of work. Prompt 7: Prepare interview or customer research questions Prompt: I'm about to interview [describe person: role, context, what they know] about [topic]. My research goal is [what you're trying to understand]. Here's what I already know or suspect: [brief summary]. Write 10 open-ended questions for the interview. Prioritize questions that test my assumptions rather than confirm them. Avoid leading questions. Flag which questions might feel sensitive to ask. This is good for customer discovery, stakeholder interviews, and expert conversations. It's not a replacement for doing the interview. It's preparation. Prompt 8: Build a decision memo from research notes You've done the research. Now you need to turn scattered notes into something a person can act on. Prompt: Here are my research notes on [topic]: [paste notes]. My audience is [who will read this]. The decision we need to make is [specific decision]. Organize my notes into a decision memo with these sections: Background (what we know), Key Findings (what the research shows), Open Questions (what we still don't know), Tradeoffs (options and their implications), and Recommended Next Step. Mark any finding where the evidence is thin or uncertain. Do not add information from outside my notes. The instruction to mark thin evidence is one you should never skip. It's the difference between a memo that helps people decide and one that overconfidently misdirects them. Prompt 9: Fact-check a draft This prompt works on your own draft or someone else's. Prompt: Here is a draft document: [paste draft]. Identify every specific factual claim: statistics, dates, named studies, quotes, organizational claims, historical assertions. For each one, flag whether it's verifiable from context in the document, requires external verification, or appears to be an unsupported assertion. List them in a table. Do not verify the claims yourself. Just identify which ones need checking. You then go verify the flagged items against primary sources. Don't ask the AI to verify them. It will often produce a confident-sounding wrong answer. Prompt 10: Build an evidence table Prompt: I'm writing a [report/article/presentation] arguing that [your thesis]. Here are the sources I've collected: [paste summaries or excerpts]. Build an evidence table that shows: Source, Claim from source, How it supports or challenges my thesis, Evidence strength (strong/moderate/weak, with brief reason). Order the table from strongest to weakest evidence. Then list any claims in my thesis that have no supporting evidence in my sources. That last instruction will show you where you're arguing from hope rather than evidence. Fix those gaps before you present, not after. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Which prompts work best for which research tasks Different jobs call for different prompts. Here's how they map to common work situations. Research taskBest prompts to use Starting a new topic from scratch1, 2 Analyzing a report or article3, 4 Comparing vendor options or competing arguments5 Pressure-testing your own research6, 9 Preparing for a meeting or interview7 Turning research into a deliverable8, 10 Checking someone else's draft or claims4, 9 Use prompts 1 and 2 at the start of any unfamiliar topic. Use prompts 9 and 10 before anything goes to a decision-maker. Prompts 3, 4, and 5 are the workhorses you'll return to constantly. A note on research versus writing People conflate these two things and it causes problems. Research is finding out what's true. Writing is communicating what you found. They're different jobs, and AI is better at one than the other. For writing, AI is genuinely fast and useful. You paste your notes, it drafts a structure, you edit. For research, AI is a trap if you're asking it to generate the underlying facts. It doesn't have access to real-time data (in most configurations), it can't call an expert, and it has no way of knowing that the study it's confidently citing was retracted two years ago. The prompts in this article are designed for the research phase, not the writing phase. They work on material you bring to the conversation. If you want prompts for the writing and communication side of work, the broader guide to AI prompts at work covers that territory. The one place the line blurs is the decision memo (Prompt 8). That's a writing task built on your research. It works because you're pasting your notes and asking AI to organize them, not asking AI to know things. Quality rules that apply to every single prompt Verify before you publish. Every statistic, every date, every named study, every quote. If it came from AI, assume it might be wrong until you've checked it against the original source. AI invents references. It doesn't mean to. It just pattern-matches toward what a citation in this context might look like. A realistic-sounding author, journal, and year doesn't mean the paper exists. Search for it. Find the primary source. If you can't find it, don't cite it. Don't let AI write your conclusion. Use it to organize evidence. The conclusion is yours. You're signing your name to it. Keep the "do not add outside information" instruction in any prompt where you're working from specific documents. Without it, the AI blends your sources with everything else it knows, and you lose track of provenance. The actual skill here isn't prompting It's asking good questions in the first place. Knowing which sources are credible. Spotting when evidence is thin. Understanding the context that makes a finding relevant or irrelevant to your specific situation. Recognizing the difference between "the data says X" and "this data, with these limitations, in this context, suggests X." Those are the skills that make you hard to replace. AI makes a mediocre researcher faster. It doesn't make a fast researcher good. The human skills that matter in research, including judgment, source evaluation, and intellectual honesty, are exactly what AI can't fake consistently. The prompts above are tools. The tool doesn't decide what question to ask. It doesn't decide what evidence is trustworthy. It doesn't take responsibility for the conclusion. You do. Frequently asked questions Can AI find sources and citations for desk research? Not reliably. AI can suggest types of sources to look for and organizations that might publish relevant data, but it frequently invents specific citations, including plausible-sounding author names, journal titles, and publication years that don't exist. Always find and verify primary sources yourself before citing anything in work you'll publish or present. What's the safest way to use AI for research without hallucinations? Paste the source material yourself and instruct the AI to work only from what you've provided. Add "do not add information from outside this document" to any prompt where you're analyzing specific texts. Use AI to organize, compare, and summarize, not to generate facts from memory. Should I use ChatGPT or Claude for research prompts? Both work for the prompts in this article. Claude tends to be more cautious about flagging uncertainty. ChatGPT is slightly more fluent in formatting tables and structured outputs. The more important factor is how you write the prompt. A good prompt works in either tool. See the broader guide to AI prompts for work for more on this. Can I paste company data or client information into AI research prompts? No, unless your organization's policy explicitly permits it and you're using an approved enterprise tool with appropriate data handling guarantees. Strip out or anonymize any confidential strategy, customer data, financial projections, HR information, or regulated data before you paste. When in doubt, leave it out. How do I know if the evidence AI finds in my documents is actually there? Keep the original source open. When AI produces a summary or claim extraction, check it against the source text directly. If the AI attributes a claim to a source you pasted but you can't find that claim in the original, it may have blended your source with outside knowledge. The prompt instruction "do not add outside information" reduces this but doesn't eliminate it entirely. Is researching with AI cheating? That's the wrong question. Using AI to organize notes, summarize articles, or build comparison tables is the same category of work as using a spreadsheet to sort data. The question is whether the judgment, sourcing, and conclusions are yours. If you're letting AI decide what's true without verifying it yourself, that's not cheating. It's just bad research. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Client Communication Prompts (10 Templates) 10 copy-paste AI prompts for client emails: delays, scope creep, tense calls, and bad news. Draft faster, verify everything, own every word. ai client communicationclient email promptsai prompts for work AI Prompts for Work: 12 Templates You Can Use Today 12 copy-paste AI prompt templates for emails, reports, and decisions. Real examples for normal people, with what to check before you send anything. ai prompts for workchatgpt promptsworkplace ai AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation --- ## AI Resume Bullet Prompts That Actually Work URL: https://dontreplace.me/articles/ai-resume-bullet-prompts-rewrite-your-experience resumeAI promptscareerChatGPTjob search AI Resume Bullet Prompts That Actually Work 8 copy-paste AI prompts for resume bullets that get results. Includes the CAR formula, before/after examples, and safety rules you can't skip. May 22, 2026 Updated May 22, 2026 Your resume bullets probably read like a job description. "Responsible for managing social media accounts." "Assisted with quarterly reporting." "Supported the team on various projects." Congratulations, you've described having a job. You haven't made a case for keeping one, or getting a better one. AI resume bullet prompts can fix this, but only if you use them right. Dump your old bullets in, ask for "something better," and you'll get fluffy corporate garbage dressed up in slightly different words. Ask the right way, and you'll get structured, specific, results-focused bullets that actually sound like you. This is a practical guide. You'll get the formula, eight copy-paste prompts, and before/after examples. Read to the end, because the quality warnings matter as much as the templates. Why most AI resume rewrites fail People open ChatGPT, paste a bullet, and type "make this better." The AI obliges. It makes it longer, more confident, and completely generic. The problem is context. The AI doesn't know what you actually did, what the stakes were, what you were working with, or what happened as a result. It fills those gaps with plausible-sounding language, which is exactly what a recruiter's eyes slide off without registering. Good resume bullets follow a pattern. Context or challenge (what was the situation), action (what you specifically did), result (what changed because of it). You'll see this called CAR or STAR in career coaching circles. The labels don't matter. The structure does. AI can build that structure for you. But it needs the raw material to work with. The rule from the book side of this site applies here too: garbage in, garbage out. If you give the AI a vague half-sentence, you'll get a vague full sentence back. The reusable prompt formula for resume bullet prompts Before the copy-paste prompts, understand what makes a good one. Every effective resume prompt has five parts. Role: Tell it who you are. "I'm a marketing coordinator" or "I manage a team of six in a regional logistics operation." Context: What was the situation? What were the constraints? What was broken, growing, or at risk? Action: What did you specifically do? Be concrete. Tools used, decisions made, processes changed. Result: What happened? Numbers if you have them. If not, describe the direction: faster, cheaper, fewer complaints, more revenue. Format constraint: How many bullets, what length, what tone. "Give me three punchy bullets, each under 20 words" produces very different output than "give me one detailed bullet with a subordinate clause." Miss any of these and the AI will guess. It's fast and wrong. You're slow and right. Work together. The master prompt template Copy this and fill in the brackets each time: I'm a [job title] with [X years] of experience in [industry/function]. Rewrite the following into [number] strong resume bullet(s) using the CAR format: Challenge/context, Action, Result. Keep each bullet under 25 words if possible. Use active verbs. Do not invent numbers or claim results I haven't mentioned. Here's my raw experience to work with: [paste your notes or existing bullet here]. Here's the job I'm applying for: [paste job description excerpt or describe target role]. That last line matters more than people think. Tailoring to the job description is how you get past applicant tracking systems and actually into the interview pile. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 8 copy-paste AI resume bullet prompts These are ready to use. Fill in the specifics. Don't skip the brackets, that's where your actual experience lives. Prompt 1: The vague duty rewrite I'm a [job title]. I had a duty that I've described vaguely as "[paste old bullet]." Rewrite this into one specific, results-focused bullet using an active verb. Assume I did this well. Don't add metrics I haven't provided. Ask me if you need more context before rewriting. Prompt 2: Turning notes into bullets Turn the following rough notes into two polished resume bullets. Use the CAR format (context, action, result). Active verbs, no filler phrases, under 20 words each. Here are my notes: [paste your raw notes or memory dump about the project]. Prompt 3: Adding a metric you actually have I have a result I'm proud of: [describe the result in plain English, e. g., "we cut the time it took to onboard new clients from six weeks to two"]. Help me turn this into one sharp resume bullet that leads with an active verb and includes this specific result. Job I'm applying for: [describe]. Prompt 4: The leadership bullet I managed a team of [number] people doing [describe the work]. During [time period], we [describe what the team achieved or improved]. Write two bullets that show my leadership contribution specifically, not just what the team did. Don't make me sound like I'm taking credit for everything. Prompt 5: The process improvement bullet I changed or improved a process at work. Before: [describe the old way]. After: [describe the new way]. Result: [describe what got better]. Write one to two bullets that communicate this improvement clearly and concisely. Prompt 6: The "soft skill with evidence" bullet I want to show [soft skill, e. g., stakeholder management, cross-functional collaboration] on my resume without just listing the skill. Here's a specific situation where I demonstrated it: [describe the situation]. Turn this into one bullet that demonstrates the skill through the action and result, not just names it. Prompt 7: The quantity-to-quality upgrade I have too many bullets on my current resume. Here are five of them: [paste five bullets]. Rank them from strongest to weakest and explain briefly why. Then rewrite the two weakest ones using the same raw experience but stronger framing. Prompt 8: The tailored version for a specific job I have a resume bullet that's decent but generic: "[paste bullet]." Here's the job description I'm applying for: [paste relevant excerpt]. Rewrite the bullet so it speaks directly to what this employer is looking for, without changing the facts of what I actually did. These eight prompts cover most situations you'll run into. If you want more templates across different work scenarios, the broader list at AI prompts for work has you covered. Before and after: what this actually looks like Example 1: Marketing Before: "Responsible for managing the company's social media accounts." After: "Grew LinkedIn following by 40% in six months by shifting to a weekly long-form content cadence; zero ad spend." What changed: specific platform, specific metric, specific action, specific constraint. Example 2: Operations Before: "Assisted with process improvements across the supply chain team." After: "Mapped three manual reporting processes and consolidated them into one shared dashboard, cutting weekly prep time by four hours across six team members." What changed: the AI asked for more detail, you provided it, and suddenly "assisted with" became "mapped and consolidated." Example 3: HR Before: "Supported onboarding for new employees." After: "Redesigned the 90-day onboarding checklist for a 200-person office, reducing time-to-productivity for new hires by an estimated three weeks." Note: "estimated" is doing work here. That's fine. Don't invent exact figures. If you believe it was roughly three weeks faster, say roughly three weeks. Example 4: Finance Before: "Prepared monthly financial reports for management." After: "Produced 12 monthly variance reports for senior leadership, flagging a recurring budget discrepancy in Q3 that led to a $45K cost recovery." If you have the specific story, put it in. If you don't, the AI can't invent one. Don't let it. Example 5: Project management Before: "Helped coordinate projects across multiple departments." After: "Coordinated delivery of a cross-functional product launch across four teams, bringing the project in two weeks ahead of schedule by consolidating weekly standups into a shared tracker." What changed: "helped coordinate" became a specific deliverable with a measurable outcome and a named action. The AI didn't invent this. You described what actually happened and gave it something to work with. The safety rules you can't skip This is the section people skim and regret. Don't paste confidential data. Revenue figures from a private company, client names, internal project codenames, anything under NDA. Strip it before it goes into ChatGPT. Yes, even if you trust the tool. The habit matters. Never let the AI invent metrics. If you didn't measure it, don't claim it. "Improved customer satisfaction by 23%" is a lie if you don't have data to back it up. Interviewers verify. "Reduced customer complaints significantly" is vague but honest. Verify everything before sending. Dates, job titles, names of tools and systems. AI confidently misremembers things it never knew in the first place. You're responsible for what's on your resume. Read it aloud. If the bullet doesn't sound like you, fix it. A recruiter who likes your resume is going to interview a human being. If your bullets read like a McKinsey press release and you sound like a normal person in conversation, that's a problem. If you want the bigger picture on how to build AI into your work without turning into a copy-paste machine, how to use ChatGPT at work covers the practical foundation. What AI can't do for you here It can't tell you what your actual results were. It can't remember the project you're most proud of. It can't decide which version of your experience is most relevant to the job you're applying for. And it can't read the room of a specific company culture well enough to calibrate tone. That's your job. AI is fast at language-shaped work. You're the source of truth about your own career. Rule #5 from Don't Replace Me puts it plainly: it's not smart, it's fast. You bring the substance. It brings the structure. Neither of you should be doing the other's job. This is also why chasing every AI resume tool is a waste of time. There are now dozens of products that claim to optimize your resume with AI. Most of them wrap ChatGPT or Claude, charge a monthly fee, and add a dashboard that makes it feel like something special is happening. Nothing special is happening. Just use the underlying model directly. AI skills you actually need in 2026 makes this point about the whole category, not just resume tools. How to build a session that actually works Don't just prompt once. Resume work with AI is a conversation. Start by dumping your raw notes into the chat. Everything you remember about a job: what you did, what broke, what you fixed, what your boss said at your review, what you're most proud of. Then ask the AI to help you find the two or three most compelling stories in that dump. Once you have those stories, use the templates above to shape them into bullets. Ask for three versions. Pick the one closest to right. Edit it until it sounds like you. Then run the tailored version prompt for each job you apply to. This is the step most people skip because it takes time. It's also the step that doubles your response rate. A note on iteration One thing people get wrong: they accept the first output. Don't. After the AI gives you a bullet, respond with: "Give me two more versions. One shorter and punchier. One that leads with the result instead of the action." Now you have three options to work with. Pick the best parts of each and combine them. This back-and-forth takes maybe five extra minutes per bullet. It consistently produces better output than the first pass. The AI isn't worse at round two. It's actually better because it has more context from the first exchange. Think of it the way you'd work with a good editor. You don't accept their first suggestion wholesale. You read it, take what works, push back on what doesn't, and end up somewhere better than either of you started. The broader framework for future-proofing your career includes this kind of deliberate AI fluency as one of the things that actually compounds over time. A better resume is just the start. Common mistakes that kill otherwise good bullets You've done the work. You've used the prompts. You've iterated. And the bullets are still somehow off. Here's what's usually going wrong. Too much jargon. Internal terminology from your last company means nothing to someone who works somewhere else. "Drove adoption of our Agile sprint cadence" reads fine internally. Externally it sounds like someone threw business words at a wall. Ask the AI to explain the bullet to someone outside your industry. If it can't, rewrite it. Leading with the wrong thing. Bullets that start with "Responsible for" or "Helped to" signal a passive contributor. Lead with the active verb. "Rebuilt," "launched," "cut," "negotiated," "trained." The first word sets the tone. Burying the result at the end. Most people write bullets chronologically: I did X, and then Y happened. Flip it for high-impact roles. Lead with the result, then explain how. "Cut client onboarding time by three weeks by redesigning the intake checklist and automating document collection." The result is the hook. Overloading one bullet. If a bullet needs three semicolons to hold together, it's two bullets. Or a story for the interview, not a line on the page. Keep bullets to one accomplishment, one result. Using the same verb twice. If "managed" appears in three bullets, you look like you only did one thing. Ask the AI to audit your verb list. It's good at this. Frequently asked questions Can AI write my entire resume for me? Technically yes, practically no. AI can draft and structure your bullets, summary, and skills section, but it needs your real experience as input. Give it vague instructions and you'll get a generic document that reads like every other AI-written resume in the pile. The more specific detail you provide, the better the output. Think of it as an editor, not a ghostwriter. Is it cheating to use AI for my resume? No. Hiring managers use AI to screen resumes. Using AI to write yours is fair game. The only form of cheating that matters is lying: invented metrics, exaggerated titles, results you didn't achieve. That's on you, not the tool. What's the best AI tool for resume bullets? ChatGPT or Claude, used directly. Not a $29/month resume tool that wraps one of them and adds a template. Use the free or paid tier of the underlying model, and apply the prompts in this article. You'll get better results and keep your money. How do I handle metrics if I don't have real numbers? Use honest approximations. "Reduced turnaround time by roughly 30%" or "handled approximately 50 accounts" is fine if that's your genuine estimate. "Reduced turnaround time by 32.7%" when you made that up is not. Recruiters and hiring managers check. The specific-but-honest approach is always better than the precise-but-false one. Should I use the same resume bullets for every job? No. Use the tailored prompt (Prompt 8 above) to adjust your bullets for each application. You don't need to rewrite everything from scratch, but the top few bullets should speak directly to what the job posting asks for. This is tedious. It's also the thing that actually works. What if the AI changes facts I know are true? Edit them back. AI will sometimes rephrase things in ways that subtly shift meaning, replace one tool with another that sounds similar, or add confidence to a claim you stated tentatively. Read every bullet against your original notes. You're the editor. The AI is the first draft. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Performance Review Prompts That Work 10 copy-paste AI performance review prompts that turn your messy notes into polished self-review language. Real templates, no fluff. performance reviewAI promptsself-review AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management --- ## AI Retrospective Prompts: Learn From Work Without Blame URL: https://dontreplace.me/articles/ai-retrospective-prompts-learn-from-work-without-blame ai retrospective promptsretrospective templatesblameless retrospectiveproject retrospectiveai at work AI Retrospective Prompts: Learn From Work Without Blame 10 copy-paste AI retrospective prompts for turning messy notes into themes, blameless summaries, action plans, and follow-up checklists. June 21, 2026 Updated June 21, 2026 Most retrospectives are lies. Polished, blameless, action-item-complete lies dressed up as learning. Someone writes "communication could be improved" when what actually happened is that two departments stopped talking to each other for six weeks and shipped the wrong thing. Someone lists three action items that never get a real owner. Everyone nods. The doc goes into a folder. Nothing changes. AI retrospective prompts won't fix that. But they can help you do the structural work faster, so you spend your actual meeting time on the conversation that matters instead of arguing about how to word a bullet point. Here's how to use AI for retrospectives without turning a team conversation into a one-person hallucination dressed in corporate formatting. What AI is actually good at in a retrospective AI is fast and neutral. That's most of it. Give it a wall of chaotic notes and it'll pull out recurring themes without caring who said what. Ask it to reword something and it won't take sides. Tell it to organize a timeline and it'll do it without forgetting things the way humans do when they're tired. That's genuinely useful. Facilitation prep, neutral language suggestions, theme clustering, action item formatting, follow-up summaries. These are the parts where the meeting runs long and someone's handwriting is illegible and the Google Doc is fourteen colors of comments. What AI is not good at: knowing what actually happened. It doesn't know your team dynamics. It can't tell you whether a missed deadline was a process failure or one person's bad month. It doesn't know which risks your leadership will actually act on versus archive politely. If you feed it vague complaints, it produces polished nonsense. Rule #13 from Don't Replace Me says it plainly: garbage in, garbage out. So treat AI like a fast retrospective facilitator. Good at structure. Not authorized to draw conclusions. Before you paste anything into an AI tool Stop. Check what's in your notes. Do not paste any of the following into an AI tool that hasn't been specifically approved for sensitive data at your company: customer names or personal information, employee performance records, anything that could identify a specific person's behavior or mistake, confidential financials, legal disputes, security incidents, HR issues, client contract details, or board materials. Retro notes are often full of this stuff. Someone writes "the issue started when [Name] missed the handoff" or references a client complaint by name or mentions a financial target that wasn't public. Scrub that before it goes anywhere near an AI tool. What's safe: anonymized timelines, themes and categories, questions and experiments, action items with role labels instead of names, general project descriptions. If you're not sure, check with whoever runs your company's AI use policy. This matters more for retrospectives than almost any other work document, because retro notes capture what went wrong. That's exactly the kind of material that can cause real problems if it leaks. The reusable AI retrospective prompt formula Before the 10 templates, here's the structure every good AI retrospective prompt uses. You can build any variation from this: "You are a neutral retrospective facilitator. Here are [notes / raw input / outcomes]. Please [specific task: cluster themes / reword blamefully / create action items]. Format the output as [format]. Do not invent root causes, assign blame, or fill gaps with assumptions. Flag anything unclear with a question instead." That last sentence is the one most people skip. Without it, AI fills gaps confidently. With it, you get a document that tells you where the humans need to have a real conversation. Now the templates. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI retrospective prompts you can use today Prompt 1: Turn messy retro notes into themes "Here are raw notes from our retrospective session. Cluster these into 3-5 themes without changing the meaning, attributing comments to individuals, or adding context that wasn't in the notes. For each theme, list the specific points that belong to it. Flag any notes that seem contradictory or unclear." Use this first, before your meeting or during synthesis. It's the one that saves the most time and carries the least risk of distortion. Prompt 2: Write a blameless project retrospective "Here is a summary of what happened during [project name]: [timeline, decisions, outcomes]. Write a blameless retrospective in this format: What we were trying to achieve, What happened, What worked, What didn't, What we'll do differently. Use 'we' language throughout. Do not name individuals. Do not assign fault. Do not soften material problems into vague observations." That last line matters. AI loves to turn "we shipped two weeks late and lost the client" into "there were some timeline challenges." Don't let it. Prompt 3: Find root causes without pretending certainty "Based on these notes, identify possible contributing factors to [specific problem]. Present each as a hypothesis, not a conclusion. Use language like 'this may have contributed to' or 'one possible factor is.' Do not state root causes as facts. List what additional information would be needed to confirm each hypothesis." This one is underused. The 5 Whys framework is great in theory. In practice, teams often arrive at a confident root cause that's actually just the most comfortable answer. AI making it explicit that these are hypotheses keeps the conversation honest. Prompt 4: Convert complaints into experiments "Here are the problems our team identified: [list]. For each problem, suggest one small, testable experiment we could run in the next 30 days to address it. Each experiment should have a clear action, a measurable signal, and a time limit. Do not suggest structural changes, hiring decisions, tool purchases over $X, or anything that requires approval we haven't mentioned." That constraint at the end is crucial. Without it, AI helpfully suggests "hire a project manager" or "rebuild the process" as if those are things you can just do on Tuesday. Keep it small and actionable. Prompt 5: Summarize what went well "Based on these notes, write a summary of what worked well during [project]. Be specific. Don't generalize. If the notes only mention one specific thing that worked, say one specific thing worked and don't pad the section. The goal is an honest record, not a morale document." People often use this section as a warm-up. AI will stretch it if you let it. Tell it not to. Prompt 6: Summarize what went wrong "Based on these notes, write a clear summary of the problems we encountered during [project]. Be direct. Do not soften language that describes material problems. Do not reframe failures as learning opportunities unless the team explicitly identified them that way. Preserve the severity of each issue as described in the notes." This is where AI is most likely to fail you quietly. It defaults to optimistic framing. Explicitly override that default. Prompt 7: Create an action plan with owners "Based on these agreed actions from our retrospective: [list], create an action plan. For each action, include: what will be done, who is responsible (use role titles, not names), deadline, and how we'll know it's done. If any action is missing an owner or a deadline, flag it rather than inventing one." The "flag it rather than inventing one" instruction is the difference between a useful document and a fake plan. For more on structuring work this way, the AI project management prompts guide covers the full framework. Prompt 8: Write a stakeholder-safe retro summary "Here is our internal retrospective: [summary]. Write a shorter version suitable for sharing with [stakeholder group: leadership / clients / partner teams]. This version should: communicate what happened honestly, explain what we're doing differently, be professionally worded, and not contain internal criticism of individuals, speculative root causes, or unresolved conflicts. Flag anything in the original that you're not sure how to handle appropriately." This exists because internal retrospectives should not be forwarded to clients raw. They're two different documents with two different purposes. AI can bridge that, but a human needs to read it before it goes anywhere. Prompt 9: Compare expected vs. actual outcomes "Here are our original goals for this project: [goals]. Here are the actual outcomes: [outcomes]. Write a structured comparison that shows where we hit, missed, or exceeded each goal. Be factual. Do not explain gaps unless the explanation was explicitly agreed in the notes. Do not minimize misses." Useful for quarterly reviews, post-mortems, and any situation where someone will ask "but did it work?" Related: if your workflow produced unexpected gaps, the AI workflow audit prompts can help you dig into where things broke. Prompt 10: Create a 30-day follow-up checklist "Based on our action plan: [list], create a 30-day follow-up checklist. Include: check-in dates for each action, questions to ask at each check-in to verify progress, and a trigger for escalation if something is off track. Format as a simple checklist." Retrospectives die because no one checks on the actions. This prompt turns good intentions into a calendar item. What AI cannot do in a retrospective It can't tell you if your team's psychological safety is real or performed. It doesn't know if someone listed as an action owner actually has the authority or bandwidth to do the thing. It can't read the room when a theme cluster hides a serious interpersonal conflict. It doesn't know which of your "what went wrong" items is a management problem that won't be fixed by any experiment. Those things require a human who was in the room, knows the people, understands the organizational politics, and has the judgment to decide what actually gets documented versus what gets handled directly. AI also cannot do your legal or HR review. If a retrospective surfaces something that involves a specific person's conduct, a security incident, or a client dispute, that content leaves the retrospective document and goes to the appropriate human who handles those things. Full stop. The AI risk assessment prompts article is worth reading before you try to use AI to analyze anything with legal exposure. AI makes retrospectives faster to structure. The quality of the learning still depends entirely on the quality of the conversation and the honesty of the people in it. If you're using AI to avoid the hard conversation, the hard conversation still needs to happen. You've just got better-formatted proof that you didn't have it. Before you share anything Read it. The whole thing. Not skim it. Read it as if you were someone on the receiving end who wasn't in the meeting. Check that no individual is identifiable from the wording even without their name. Check that no material risk was softened into nothing by AI's tendency toward diplomatic language. Check that every action item has a real owner who actually agreed to own it. Check that the root causes are labeled as hypotheses, not conclusions. Then run it past whoever needs to review it before it goes to leadership, clients, or anyone outside your immediate team. Retrospectives that are wrong in small ways tend to become the official version of history. That's worse than no retrospective at all. For the broader picture on using AI at work without the hype, how to use AI at work is the practical starting point. Frequently asked questions Can AI write a retrospective for me? AI can structure and format retrospective notes quickly, but it can't replace the team conversation, the human judgment, or the psychological safety that makes retrospectives work. Use it to organize what your team said, not to substitute for saying it. The AI meeting notes guide covers how to handle raw notes before they go into any AI prompt. What should I not paste into an AI tool for a retrospective? Never paste customer names or personal data, employee performance details, HR issues, legal disputes, security incidents, confidential financials, client contract information, or anything that identifies an individual's specific mistake or conduct. Anonymize before you paste. If in doubt, don't paste it. How do I stop AI from softening problems in a retrospective summary? Tell it explicitly in the prompt. Add an instruction like: "Do not soften language that describes material problems. Do not reframe failures as learning opportunities unless the team identified them that way." Without that instruction, AI defaults to diplomatic language that can make serious failures sound like minor hiccups. What's a blameless retrospective, and can AI help write one? A blameless retrospective focuses on system and process failures rather than individual fault. It's a method from engineering culture, used to get honest post-mortems without people becoming defensive. AI can help reframe notes into "we" language and avoid naming individuals, but the actual psychological safety that makes people tell the truth has to come from the team and the facilitator. How do I make sure retrospective action items actually get done? Use Prompt 10 in this article to create a 30-day follow-up checklist with named check-in dates and escalation triggers. The bigger structural issue is that AI can format an action plan, but someone still has to own it. If no one owns it, it doesn't happen regardless of how polished the document looks. Can AI identify the root cause of what went wrong in a project? No. AI can suggest hypotheses based on the information you give it, but it doesn't know your organization, your team dynamics, or the context behind decisions. Always prompt it to frame potential root causes as hypotheses, not conclusions, and always have a human validate any root cause claim against what actually happened. For more on this, what AI can and can't do gives the full picture. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Automation Prompts: Automate the Boring Stuff 10 copy-paste AI automation prompts for mapping workflows, writing SOPs, adding approval gates, and QA testing. Practical templates with safety rules included. ai automation promptsworkflow automationSOP templates AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts --- ## AI Risk Assessment Prompts: 10 Templates URL: https://dontreplace.me/articles/ai-risk-assessment-prompts-check-the-blast-radius AI risk assessmentAI promptsrisk managementAI at workAI tools AI Risk Assessment Prompts: 10 Templates 10 copy-paste AI risk assessment prompts for decisions, launches, and process changes. Includes what not to paste and where AI breaks down. June 16, 2026 Updated June 16, 2026 AI risk assessment sounds like something consultants charge $40,000 to deliver in a 47-slide deck. It doesn't have to be. You can use AI to do a fast, structured first pass on almost any decision, launch, or process change, and walk into the real review meeting already knowing where the landmines are. That's the point of this article. Ten copy-paste prompts. A reusable formula. And a clear list of what not to let AI touch. Why AI risk assessment prompts are worth using (and where they break) AI is fast. That's not the same as being right. When you're assessing risk, speed is useful for one thing: getting the obvious stuff on paper quickly so humans can argue about the hard stuff. AI is genuinely good at drafting stakeholder maps, generating failure mode lists, structuring scenario questions, and producing first-draft summaries. It's a fast analyst who has read a lot and forgotten none of it. The problem is it will also confidently invent compliance requirements, make up legal obligations, and treat a made-up statistic with the same tone as a real one. That's not a bug you can see coming. It looks the same as good output. So the rule is: AI drafts the structure. Humans validate every claim. Nothing goes into a decision doc without an accountable owner attached to it. If you're building out broader processes around this, the AI workflow audit prompts guide covers how to map what's actually happening before you automate or assess anything. The reusable AI risk assessment prompt formula Before the templates, here's the formula. Every good risk prompt needs four things: Context. What are you assessing? Be specific about the decision, change, or launch. Constraints. What's already off the table? What are the known guardrails? Audience. Who will use this output? A board? A security team? A product manager? Output format. What do you want back? A list? A table? A narrative summary? A vague prompt like "assess the risks of our new AI chatbot" produces a vague, generic response that could apply to any company on earth. A prompt like "identify failure modes for a customer-facing AI chatbot that handles billing disputes, where our constraints include GDPR compliance, no access to payment systems, and a 48-hour max resolution SLA" gets you something you can actually work with. Garbage in, garbage out. That's not a metaphor. It's the operating principle. Here's a quick comparison of what a vague prompt produces versus a specific one: Prompt typeWhat you getWhat you can do with it Vague ("assess the risks of X")Generic list, applies to any companyNot much. Delete it. Specific (context + constraints + audience)Targeted failure modes, relevant questionsA real starting point for review Specific + output formatStructured table or summary ready to shareBring it straight into a meeting The difference isn't the AI. It's the inputs you give it. What never to paste into an AI tool Before the prompts, a hard stop. Do not paste any of the following into a public or unapproved AI tool: Customer PII (names, emails, IDs, addresses, health data, financial data) Employee records, performance data, or HR files Credentials, API keys, or access tokens Contracts, NDAs, or documents involved in legal disputes Confidential financial data or unreleased earnings Security architecture details or vulnerability reports Unreleased product plans, pricing, or roadmaps Proprietary workflows or trade secrets Use placeholders. "Our third-party payroll system" instead of the vendor name and contract terms. "A subset of enterprise customers" instead of the actual account list. The prompts below are designed to work without sensitive details. If your organization has an approved enterprise AI setup with proper data handling agreements, your rules may differ. Check before you paste. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → AI risk assessment prompts: 10 templates to use right now Prompt 1: Define the decision or launch I need to assess the risks of [describe the decision, change, or launch in plain language]. The scope is [brief scope description]. Key constraints include [list known constraints: budget, timeline, legal, technical]. Help me write a one-paragraph problem statement and a list of 5-8 clarifying questions I should answer before the risk assessment begins. What to do with the output: Answer the clarifying questions yourself, with your team. Don't let AI answer them. Those answers are the actual inputs to everything that follows. Prompt 2: Map the stakeholders and their interests We are [describe the project or decision]. List the internal and external stakeholder groups who could be affected. For each group, describe: (a) how they might be impacted, (b) what their primary concern is likely to be, and (c) whether they need to be consulted, informed, or involved in approval. Present this as a table. Review note: Add any stakeholders AI misses. Frontline employees and end customers are routinely underweighted in AI-generated stakeholder maps. Go check with actual people. Prompt 3: Identify sensitive data involved We are [describe the project]. List the types of data this project might involve, including data collected, processed, stored, or shared. Flag which categories are potentially sensitive, regulated, or subject to privacy laws. Note where a formal privacy review or data protection impact assessment may be required. Do not make compliance determinations, flag questions for a qualified reviewer. Hard rule: AI cannot tell you whether you're compliant. It can tell you what questions to ask. Legal and privacy teams answer those questions. Prompt 4: Find the failure modes We are [describe the project or system]. Generate a list of realistic failure modes, ways this could go wrong operationally, technically, or from a user experience perspective. For each failure mode, describe: (a) what breaks, (b) who is affected, and (c) an early warning sign we might notice before it becomes serious. Focus on plausible failures, not theoretical worst cases. What to check: Run this list past someone who has actually operated a similar system. AI generates plausible-sounding failures but misses the weird, specific ones that come from operational experience. The person who ran the last product launch knows things Claude doesn't. Prompt 5: Rate impact and likelihood Here is a list of risk scenarios for [project name]: [paste your failure modes from Prompt 4]. For each scenario, estimate: (a) likelihood (Low / Medium / High), (b) impact if it occurs (Low / Medium / High), and (c) a one-sentence rationale for each rating. Present as a risk matrix table. Flag any scenarios where you don't have enough information to rate confidently. Important: These ratings are a starting point for human discussion, not finished assessments. Anyone with accountability for the project should review and challenge every rating. Prompt 6: Spot legal and compliance questions We are [describe the project]. Based on this description, what legal, regulatory, or compliance areas might be relevant? List the questions a legal or compliance reviewer should be asked, do not provide answers or conclusions. Flag areas where regulation varies by jurisdiction and where specialized legal review is likely needed. Do not let AI: Write compliance conclusions, confirm regulatory requirements, or make statements about what you "must" or "don't need to" do. Those are legal opinions. Get actual legal opinions from actual lawyers. Prompt 7: Review customer and employee harm potential We are [describe the project]. Describe the realistic ways this could harm customers or employees, including direct harm (data exposure, service failure, discrimination), indirect harm (erosion of trust, confusion, loss of access), and cumulative harm (patterns that seem minor individually but significant at scale). For each harm type, suggest a mitigation question the team should answer. Check this one carefully. AI tends to generate generic harms and miss context-specific ones. People closest to customers and employees will spot what AI misses. Talk to them. Prompt 8: Check security and privacy boundaries We are [describe the system or process, using no sensitive technical details]. What security and privacy boundary questions should we be asking? Generate a checklist of questions covering: data access controls, authentication, third-party integrations, data retention and deletion, audit logging, and incident response. This checklist is for a security reviewer, not a substitute for one. Non-negotiable: Security review by a qualified person still happens. This checklist is prep material for that conversation, not a replacement for it. Prompt 9: Design a small pilot We want to test [describe the project or change] at small scale before broader rollout. Suggest a pilot design including: (a) scope and participant selection criteria, (b) duration, (c) what we would measure to call the pilot a success or failure, (d) what would trigger early termination, and (e) what a rollback looks like if we need to stop. Frame this as a one-page pilot brief. Small pilots before broad rollout. Every time. This is where you find out what the risk assessment missed. Prompt 10: Write an executive risk summary Here is a risk assessment for [project name]: [paste a summary of your findings from the prompts above]. Write a one-page executive summary covering: (a) what we are doing and why, (b) the top 3 risks and their current mitigations, (c) what approvals and reviews are still outstanding, (d) the proposed pilot scope and success criteria, and (e) a recommended go/no-go decision framework. Write for a non-technical executive audience. Before this goes anywhere: Every claim in the output needs a named owner who can stand behind it. "AI identified this risk" is not accountability. "[Name], [role], has reviewed and confirmed this" is. How to run these prompts in sequence The ten prompts above aren't random. They follow a logical order, and running them that way produces a coherent body of work rather than a pile of unconnected outputs. Start with Prompt 1 to get your problem statement right. Then Prompt 2 to know who's affected. Prompt 3 to flag what data is involved. Prompts 4 and 5 to build and rate your failure modes. Prompts 6, 7, and 8 to cover your compliance, harm, and security questions. Prompt 9 to plan the pilot. Prompt 10 to summarize everything for whoever needs to make the call. Between each step, you (or someone on your team) should be doing real work: answering the clarifying questions, reviewing outputs against actual experience, filling in what AI got wrong or left out. The AI drafts. The humans edit, challenge, and validate. If you skip the human steps and just run the prompts end to end, you'll have a polished document that looks like a real risk assessment and has the structural integrity of a house of cards. It'll survive exactly until someone asks a specific operational question. Budget roughly 30-45 minutes of real human time per prompt for review and gap-filling, on top of the time it takes AI to generate the output. A full ten-prompt sequence done properly is a half-day of work, not a 20-minute shortcut. That's still faster than starting from a blank page, which is the point. What AI can't do in a risk assessment This is worth saying plainly, because the output looks authoritative even when it's wrong. AI cannot: tell you what's legally required in your jurisdiction, approve a security configuration, guarantee a compliance outcome, make customer promises or SLAs, produce headcount projections you should act on, or calculate ROI figures you should put in a board deck. It can: draft structures, generate question lists, organize information you give it, suggest failure modes, and summarize findings you've already validated. Dmitry Kargaev covers exactly this distinction in Don't Replace Me: Rule #5 is "It's Not Smart. It's Fast." The usefulness of AI in risk work is speed and structure. The accountability is yours. The difference between a useful risk assessment and compliance theater is whether real humans with real authority reviewed the outputs and attached their names to the conclusions. For the governance side of this, the AI policy prompts article covers how to build the rules your team actually needs around AI use. And if you're figuring out which processes to assess or automate first, the AI automation prompts guide and the AI decision-making prompts guide are both useful starting points. What a complete risk assessment still requires The prompts get you a structured first draft. A complete assessment still needs: A named owner for every identified risk Privacy review where personal data is involved Security review where systems or data access are involved Legal or compliance review where regulation applies Frontline input from people closest to the work and the customers Measurable success criteria that aren't just vibes A rollback plan written before you launch, not after something breaks A small pilot with a real go/no-go decision point AI can help you draft the agenda for every one of those reviews. It cannot replace any of them. One practical way to think about this: the prompts get you to about 60% of a finished risk assessment. That 60% is genuinely useful because it means the human experts you bring in aren't starting cold. They're reviewing and challenging a structured document instead of writing one from scratch. That's where the time saving actually comes from. Not from cutting the humans out, but from making their time more productive when they do show up. Frequently asked questions What are AI risk assessment prompts used for? AI risk assessment prompts help you structure a first-pass risk review for a decision, launch, or process change. They're useful for generating stakeholder maps, failure mode lists, compliance questions, and pilot designs. They don't replace legal, security, privacy, or compliance review by qualified people. Can I use ChatGPT or Claude for risk assessment? Yes, for drafting and structuring. Use an approved tool and don't paste sensitive data, PII, contracts, credentials, or confidential financials. AI can help you organize what you know and generate questions to answer. It can't tell you what's legally compliant or approve a security decision. What should I never paste into an AI tool during a risk assessment? Customer PII, employee records, credentials, contracts involved in legal disputes, confidential financial data, security architecture details, unreleased product plans, or proprietary workflows. Use placeholders and anonymized descriptions. Check your organization's AI use policy before pasting anything work-related. How do I make sure an AI-generated risk assessment is actually useful? Attach named owners to every risk. Validate every claim against a source of truth. Run outputs past people with operational experience. Confirm legal, security, and privacy questions with qualified reviewers. Treat the AI output as a first draft, not a finished document. What's the biggest mistake people make with AI risk prompts? Treating the output as the assessment rather than the starting point. AI-generated risk documents look polished and complete. That's the trap. The gaps are in what it doesn't know: your specific regulatory context, your actual security configuration, your customers' real experiences, your legal obligations. Those require humans. How detailed should my prompts be? Specific enough to be useful, general enough not to paste sensitive data. Describe the project type, the constraints, the audience, and the output format. A vague prompt produces a generic response. The more context you give (without sensitive details), the more relevant the output. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Contingency Plan Prompts: 10 Templates 10 copy-paste AI contingency plan prompts for ChatGPT and Claude. Map failures, assign owners, draft comms, and verify assumptions before a crisis hits. contingency planningAI promptsbackup plans AI Customer Service Prompts: 10 Templates That Work 10 copy-paste AI customer service prompts for ChatGPT and Claude. Draft replies, de-escalate complaints, and summarize tickets without sounding like a robot. AI promptscustomer serviceChatGPT at work AI Dependency Register Prompts: Keep Projects From Surprising You Use AI dependency register prompts to map owners, blockers, assumptions, and downstream impact before one small change becomes a three-week incident. AI dependency mappingproject managementAI prompts --- ## AI Roadmap Prompts: Turn Vague Plans Into Sequenced Work URL: https://dontreplace.me/articles/ai-roadmap-prompts-turn-vague-plans-into-sequenced-work ai roadmap promptsproduct roadmapai at workplanning templatesproject management AI Roadmap Prompts: Turn Vague Plans Into Sequenced Work 10 copy-paste AI roadmap prompts for ChatGPT and Claude. Turn messy ideas into sequenced plans, executive summaries, and honest milestones. July 08, 2026 Updated July 08, 2026 Most roadmaps start as a document that contains some bullet points, a few post-it notes exported to Notion, and the vague memory of a conversation someone had three months ago. You know the kind. "Make onboarding better." "Fix the thing customers complain about." "Something with AI." That's your strategy. AI roadmap prompts won't fix the strategic thinking gap. But they'll do the grunt work of turning that pile of intentions into sequenced, readable, structured work, fast. That's what these templates are for. Before we get into the prompts: this is the part where most "AI productivity" articles go off the rails. A clean roadmap slide is not the same thing as good strategy. If your inputs are vague, the output is confident nonsense in a nicer font. AI makes things look polished. Polished is not correct. You still own the judgment. What AI roadmap prompts are actually good for AI is genuinely useful at the boring-but-important layer of roadmap work: clustering scattered ideas into themes, surfacing dependencies you didn't notice, drafting milestone language, translating a messy list into a now-next-later structure, and making a version for executives without you rewriting the whole thing. It is not good at deciding what matters most. It can't know your company's actual risk tolerance, which stakeholder will torpedo a launch, or whether the engineering team is already stretched thin on another project. It doesn't have that context unless you give it, and even then it will guess where it can't know. The prompts below are designed to give it enough information to be useful, and to keep you in the driver's seat on the things that actually count. One more thing before the templates. Don't paste confidential strategy, unreleased product details, financial forecasts, board materials, customer PII, employee records, security incidents, legal disputes, contracts, or client data into any AI tool that isn't explicitly approved for that purpose. Anonymize. Generalize. Protect what needs protecting. The reusable AI roadmap prompt formula Every prompt below follows the same structure. Learn the structure and you can write your own. You are a [role]. Context: [what we're planning, goals, constraints, current state]. Task: [what I need you to produce]. Format: [structure, length, what to include]. Assumptions to surface: [what the AI should flag rather than guess]. The "assumptions to surface" line is the one most people skip. It's the most important one. It tells AI to show you where it filled in gaps with guesses, so you can replace the guesses with facts before the roadmap gets in front of anyone. 10 AI roadmap prompts you can use today Prompt 1: Turn messy ideas into roadmap themes "You are a product strategist helping to organize raw input into a draft roadmap. I'll give you a list of ideas, requests, and notes. Your job is to cluster them into 3-5 roadmap themes, name each theme, write a one-sentence goal for it, and list which raw inputs fall under it. Flag any inputs that don't fit cleanly and explain why. Do not invent priorities or assign dates. Here are the raw inputs: [paste your notes]." Use this when you're staring at a Notion page full of sticky notes and need to stop looking at it. Prompt 2: Convert customer feedback into roadmap inputs "You are a product manager reviewing customer feedback. Here is a summary of themes from recent feedback sessions: [paste anonymized, non-PII feedback themes]. Identify the top 3-5 problems customers are describing, suggest how they might translate to roadmap work, and flag any assumptions you're making about what customers actually want. Do not invent data. Mark anything you're inferring rather than drawing directly from the input." Remove all customer names, account IDs, or identifying details before pasting. You're summarizing themes, not uploading a CRM export. Prompt 3: Sequence work by dependencies "You are a project planner. Here is a list of work items for our roadmap: [list them]. Identify logical dependencies between these items (what needs to happen before something else can start), flag any items that appear to block multiple other things, and suggest a rough sequencing order. Note anything that might need a team or resource we haven't mentioned. Do not assign specific dates. Show your reasoning." This is the prompt that catches the thing where you planned the launch before you planned the thing the launch depends on. Prompt 4: Make a now-next-later roadmap "You are a product strategist. Using the following items [paste list], draft a now-next-later roadmap. 'Now' = things actively in progress or ready to start. 'Next' = things we'd start once current work completes. 'Later' = things we want to do but don't have clear starting conditions yet. For each item, write a one-sentence description and a confidence level (high/medium/low) based on what I've given you. Flag any item where confidence is low and explain what information is missing." The confidence column is not optional. It's the difference between a roadmap and a wish list. Prompt 5: Pressure-test priority tradeoffs "You are a critical advisor reviewing a product roadmap. Here are our top priorities: [list them]. Here are our constraints: [team size, time horizon, any fixed deadlines]. For each priority, describe what we'd be deprioritizing or delaying by choosing it. Identify any priorities that seem to conflict with each other. Ask me the questions you'd need answered to give better advice. Do not tell me what to prioritize. Surface the tradeoffs so I can make the call." This is what Chapter 7 of Don't Replace Me calls the taste moat: the human judgment about which tradeoff is worth making. AI can show you the tradeoffs. It can't weigh them for you. Prompt 6: Draft milestone descriptions "You are a technical program manager. Here is a roadmap item: [description]. Draft a milestone description that includes: what done looks like, what success criteria we'd use to evaluate it, the main risks, and what assumptions we're making. Keep it to 150 words. Flag anything you don't have enough information to write accurately." Useful when your milestones currently say things like "launch improvements to the checkout flow" and nothing else. Prompt 7: Create an executive-friendly roadmap summary "You are a communications strategist. Here is a detailed roadmap for our product/team: [paste your roadmap content, anonymized as needed]. Write a 200-word executive summary that explains what we're building, why, in what order, and what the key risks are. Avoid jargon. Do not include details that could be misread as commitments we haven't made. Flag anything in my input that sounds like a firm commitment but doesn't have a named owner or date." For more templates in this zone, the AI executive summary prompts collection has you covered. Prompt 8: Map risks and assumptions "You are a risk analyst reviewing a product roadmap. Here is the roadmap: [paste]. For each major work item or phase, identify: the key assumptions we're making, the risks if those assumptions are wrong, and any dependencies on teams or factors outside our control. Rank risks as high/medium/low. Do not invent risks. Base everything on what I've given you, and tell me where you're speculating." This pairs well with the AI risk assessment prompts if your roadmap sits inside a bigger decision with real consequences. Prompt 9: Adapt one roadmap for different audiences "You are a communications strategist. I have one roadmap and three audiences: [engineering team / executive leadership / customers or external stakeholders]. Using this content: [paste roadmap], write a version of the key messages tailored for each audience. Engineering version should include technical sequencing and dependencies. Executive version should focus on business outcomes and risks. Customer version should describe impact without internal details. Flag anything in my roadmap that shouldn't appear in the customer version." This saves the three separate rewrites you were about to do manually. Prompt 10: Check whether the roadmap is honest enough "You are a senior advisor reviewing a roadmap for clarity and honesty before it goes to stakeholders. Here is the roadmap: [paste]. Check for: items that sound like commitments but have no named owner, dates that appear specific but have no clear basis, items where the 'why' is missing, anything that sounds confident but is actually an assumption, and any work that has no clear success criteria. List what you find. Do not rewrite the roadmap. Just tell me what's missing or misleading." Run this before you share anything. It will find things you've been glossing over. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What AI gets wrong about roadmaps AI is fast and AI is confident. That combination is dangerous when you're doing roadmap work, because roadmaps are supposed to reflect real decisions, real constraints, and real commitments, not an optimistic arrangement of nice-sounding themes. The specific failure modes to watch for: AI will invent dates when you leave time ambiguous. It will imply consensus that doesn't exist. It will make guesses about business priorities based on how you've phrased things, and it will present those guesses without flagging them as guesses. It will turn an uncertain tradeoff into a clean recommendation. All 10 prompts above include an instruction to surface assumptions and flag gaps. Don't remove those lines to make the prompt shorter. That's exactly the part that keeps you honest. If your roadmap touches real financial decisions, legal exposure, hiring or headcount, security posture, customer contracts, or anything that affects external commitments, get the right people in the room. AI can help you structure the conversation. It can't replace it. For a broader look at how to fit AI tools into real project work without creating a mess of fabricated timelines and invented consensus, the AI project management prompts collection covers the surrounding territory. And if you're trying to make the prioritization layer more rigorous before the roadmap even starts, AI prioritization prompts is the right place to start. When the roadmap is done, the work has barely started A good roadmap is not the goal. It's a communication device. It tells people what you've decided, in what order, and why. It surfaces what you don't know yet. It creates accountability by naming owners and milestones. AI can help you build that artifact faster. It can structure your thinking, catch your blind spots, and translate the same plan into five different formats for five different audiences. That's real, and it's genuinely useful. But if you hand someone a roadmap that was mostly generated by AI from vague inputs, with dates that were invented and priorities that were guessed, you've just created a more attractive version of a problem you haven't solved. The AI quarterly planning prompts are useful here too if you're doing this on a regular planning cycle. The humans who will keep getting asked to lead roadmap work are the ones who use AI to do it faster while keeping their own judgment sharp on the parts that matter: what to build, what to skip, what the real risk is, and who needs to know what. A clean slide is not strategy. You are. Frequently asked questions What are AI roadmap prompts? AI roadmap prompts are structured instructions you give to tools like ChatGPT or Claude to help convert scattered ideas, customer feedback, or rough priorities into a structured, sequenced roadmap. They work best when you provide clear context, constraints, and explicit instructions to flag assumptions rather than guess. They don't replace product judgment, they speed up the drafting work. Can AI create a product roadmap for me? AI can draft a roadmap structure, suggest themes, surface dependencies, and translate your work for different audiences. It can't decide what matters most, what your company can actually commit to, or what tradeoffs are worth making. Those calls require real context, stakeholder input, and human judgment. Use AI for speed, not for strategy. Is it safe to paste my roadmap into ChatGPT? It depends on what's in it. Never paste confidential financial forecasts, unreleased product details, customer PII, legal disputes, board materials, security incidents, or anything covered by an NDA into a tool that isn't explicitly approved for that data. Anonymize specifics, use general descriptions, and check your company's AI use policy before pasting anything sensitive. What's the difference between a roadmap and a project plan? A roadmap shows what you're building and why, across a medium-to-long time horizon, at a theme or outcome level. A project plan shows how specific work gets done, with tasks, owners, and dates. Roadmaps communicate strategic direction. Project plans execute it. For project-level AI templates, see the AI project management prompts. How do I stop AI from inventing dates in my roadmap? Tell it not to explicitly in the prompt. Add a line like: "Do not assign specific dates. Use relative timeframes (short-term, mid-term, long-term) unless I provide real dates." Also run Prompt 10 (the honesty check) before sharing anything, since it's specifically designed to flag confidence that outpaces the underlying evidence. What should I do when AI generates priorities I disagree with? Use them as a starting point for your own thinking, not as a recommendation to accept. AI is pattern-matching against your inputs and making inferences. If you disagree with its output, the most useful question is: what did it see in my inputs that led to this? That often reveals something you need to clarify, or a constraint you forgot to include. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Launch Checklist Prompts: Ship Without Gaps 10 copy-paste AI launch checklist prompts for ChatGPT and Claude. Turn messy plans into structured checklists with owners, risks, and rollback steps. ai promptslaunch checklistproject management AI Migration Plan Prompts: 10 Templates That Work 10 copy-paste AI migration plan prompts for system mapping, dependency checks, rollback criteria, cutover plans, and stakeholder comms. With safety reminders. AI promptsmigration planningsystem migration AI Requirements Gathering Prompts: Stop Starting Vague 10 copy-paste AI requirements gathering prompts to turn vague asks into verified briefs. Includes safety rules and a reusable prompt formula. AI promptsrequirements gatheringproject management --- ## AI Rollback Plan Prompts: 10 Templates That Work URL: https://dontreplace.me/articles/ai-rollback-plan-prompts-undo-bad-launches AI promptsrollback planninglaunch recoveryincident managementrelease management AI Rollback Plan Prompts: 10 Templates That Work 10 copy-paste AI rollback plan prompts for ChatGPT and Claude. Define scope, map owners, write comms, and verify recovery without starting from scratch. July 24, 2026 Updated July 24, 2026 A launch goes sideways at 2pm on a Friday. Something's broken, customers are noticing, and someone in Slack just typed "do we roll back?" Three people are muted. Nobody has a doc. The on-call engineer is on a train. AI rollback plan prompts won't fix your infrastructure. But they will help you stop staring at a blank page and start building the structure you need before the incident gets worse. That's the actual use case here: not automation, not magic, just a fast way to get your thinking organized when your thinking is the last thing working properly. These 10 prompts cover the full rollback workflow, from defining scope to verifying recovery. Copy them, adapt them, use them. The prompt formula every rollback plan needs Before the templates, here's the structure that makes any AI rollback plan prompt actually work: [Role] + [Context] + [What you need] + [Format] + [Constraints] Every prompt below follows this. The role tells AI how to think. The context is your actual situation, pasted in cleanly (no credentials, no PII, no raw production logs). What you need is specific. Format tells AI whether you want a table, a list, or a paragraph. Constraints prevent hallucinated system facts and fake approvals. One hard rule before you start: do not paste customer personally identifiable information, access tokens, database credentials, private tickets, unreleased strategy, security vulnerabilities, financial records, HR issues, regulated data, or anything your legal team would have opinions about into any AI tool that hasn't been explicitly approved by your organization for that data type. Rollback planning involves sensitive operational context. Keep the prompts clean. Prompt 1: Define rollback scope Bad scope kills rollbacks faster than bad code. If nobody agrees on what "rollback" means, half the team is reverting the API while the other half is rolling back the UI layer to a different point in time. Prompt: "Act as a release manager. I'm preparing a rollback plan for [name of release or feature]. The deployment included: [list what changed, e.g., database schema changes, new service version, updated config, frontend build]. Help me define the scope of a full rollback vs. a partial rollback. Output a two-column table: Column 1 is the component, Column 2 is the rollback action and any known dependencies. Do not invent system details. Flag anything that needs engineering confirmation." Fill in the bracketed parts from your actual release notes, not from memory. If your release notes are vague, fix those first. Prompt 2: Map affected systems and owners You can't roll back what you can't find. And "who owns this" is always the question that eats 20 minutes during an incident. Prompt: "Act as an incident coordinator. For a rollback of [release name], help me map affected systems and assign owner categories. Systems involved: [list services, databases, integrations, third-party dependencies]. Output a table with columns: System, Likely Owner Role (e.g., backend eng, DBA, DevOps), Rollback Action Required (yes/no), Dependencies to check. Remind me to verify each owner before the rollback window opens." The "remind me to verify" line matters. AI will fill in plausible owner roles. Your actual org chart may disagree. Engineering sign-off before the rollback window is not optional. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Prompt 3: List rollback triggers When exactly do you pull the trigger? "When it's bad enough" is not a trigger. It's a vibe. Vibes don't hold up in post-mortems. Prompt: "Act as a site reliability engineer. Help me define clear rollback triggers for [release name]. Context: this release affects [describe affected functionality, e.g., checkout flow, login, data pipeline]. Draft a list of specific, observable triggers that would justify initiating rollback. Include severity tiers: Tier 1 (immediate rollback, no approval needed), Tier 2 (rollback within 30 minutes, manager approval), Tier 3 (monitor and hotfix). Format as a numbered list. Note where thresholds need to be filled in by the team." The severity tiers aren't just organizational tidiness. They're the thing that lets your on-call engineer make a decision at 2am without waking up five people first. Prompt 4: Compare rollback vs. hotfix vs. pause Three options, different blast radii, different recovery timelines. Most teams pick based on whoever talks loudest in the channel. Prompt: "Act as an engineering lead facilitating an incident decision. The situation: [brief description of what broke and how it was discovered]. Help me build a comparison of three options: full rollback, hotfix, and feature pause/flag. For each option, list: estimated time to resolution, risk to data integrity, customer impact during recovery, and prerequisites. Output as a comparison table. Flag assumptions I need to verify with engineers before choosing." This prompt pairs well with AI risk assessment prompts if you need to pressure-test the blast radius of each option before committing. Prompt 5: Draft a pre-launch rollback checklist The rollback plan you build before launch is the one that actually helps. The one you build during an incident is the one you have to trust when you're stressed and tired and wrong about three things. Prompt: "Act as a release manager. Help me build a pre-launch rollback readiness checklist for [release name]. Include checks for: rollback procedure documented and reviewed, rollback window defined, owner for each rollback step confirmed, database migration reversibility verified, feature flags confirmed as off by default, monitoring alerts configured, support team briefed on known risks, rollback approval chain identified. Format as a checkbox list. Add a 'verified by' column for names and timestamps." If you're building launch checklists in general, the AI launch checklist prompts article covers the full pre-ship workflow. Prompt 6: Write customer and support communications Something broke. Customers noticed. Support is getting tickets. Someone needs to say something, and it needs to be accurate without being a legal liability and calm without being condescending. Prompt: "Act as a customer communications manager. I need to draft communications for a rollback affecting [describe what customers experience, e.g., checkout errors, login failures, data not saving]. Draft three versions: (1) a status page update, maximum 60 words, factual and neutral; (2) an email to affected users, empathetic, no blame, includes what happened and what we did; (3) an internal support team script for handling inbound contacts. Do not promise compensation or make legal commitments. Flag any language that legal should review." The "flag for legal" instruction does real work here. AI will write confident-sounding sentences that your legal team will hate. Make reviewing those the human's job, not the AI's call. Prompt 7: Summarize blast radius What broke, how many people it affected, for how long, across what surface area. You need this for the post-mortem, for the stakeholder update, and for your own sanity. Prompt: "Act as an incident manager. Help me summarize the blast radius of a failed release. Information I have: [paste timeline of events, systems affected, error types observed, approximate user count if known]. Output a structured summary with sections: What changed, What failed, Who was affected (systems and users), Duration of impact, Data integrity status, Actions taken. Mark any fields where I need to verify the information before publishing internally." Keep your paste clean. No customer PII, no raw logs containing tokens or credentials. Summarize what you know in plain language and let the prompt structure it. This connects naturally to the AI incident review prompts workflow once the dust settles. Prompt 8: Build a go/no-go decision brief Someone senior needs to approve the rollback. They've been on a call for 45 minutes and have 90 seconds of attention left. The brief needs to be tight. Prompt: "Act as a technical program manager. Help me write a go/no-go decision brief for rolling back [release name]. Include: summary of the problem (2 sentences max), rollback option we're recommending and why, what we need to verify before proceeding, estimated time and risk, who needs to approve. Format as a half-page brief. No jargon. Assume the reader knows the product but not the technical details." Dee covers this kind of judgment under pressure in Don't Replace Me, specifically the idea that taste, under Rule #7, is knowing which tradeoffs matter when the situation is live and messy and everyone is looking at you. Prompt 9: Build a post-rollback verification checklist Rollback complete. Now: did it actually work? "It seems fine" is not verification. Verification is a checklist with names and timestamps. Prompt: "Act as a QA lead. Help me build a post-rollback verification checklist for [release name]. Include checks for: services back to previous version (confirmed in deployment log), error rates returned to baseline (with monitoring link), database integrity verified, affected user flows tested manually, no new alerts triggered in the 30 minutes post-rollback, support ticket volume trending down, stakeholders notified of resolution. Format as a checkbox list with a 'verified by' and 'time' column. Add a final sign-off row requiring engineering and product approval." If you built a QA checklist before the launch and it missed something, the AI QA checklist prompts article is worth revisiting before the next release. Prompt 10: Turn lessons learned into prevention notes The part everyone skips. The rollback happened, things are stable, and now everyone wants to close the incident ticket and not think about it. That's how you run the same incident six months later. Prompt: "Act as an engineering team lead running a blameless retrospective. Based on this rollback: [paste a clean summary of what happened, what broke, what was missing from the pre-launch process], help me draft 3-5 prevention notes. Each note should include: what we missed, why we missed it (process gap, knowledge gap, time pressure, unclear ownership), and a concrete change to prevent recurrence. Format as numbered action items with an owner role and a suggested due date. Keep the tone blameless and specific." The "blameless" instruction matters. AI will naturally write post-mortem language that sounds like assigned blame without realizing it. Review the output before you share it with the team. What AI rollback plan prompts can't do for you Speed is not safety. That's Rule #5 from Don't Replace Me, and it applies here directly. AI rollback plan prompts can produce a nicely formatted rollback plan in 30 seconds. That plan might look completely authoritative while containing invented system facts, assumed ownership, and rollback steps that would break your database. AI is a planning assistant. The engineering diagnosis is yours. The compliance review is yours. The final sign-off is yours. The call on whether to actually execute the rollback is never the AI's. A few things AI cannot do for you: Confirm your actual system state Verify that rollback steps are safe for your specific database schema Guarantee that a procedure has been approved Replace an on-call engineer's knowledge of what changed and why Read your deployment logs, access tokens, or private incident tickets (and it shouldn't) If you want to understand the broader question of what AI is and isn't capable of, what AI can and can't do covers it without the hype. How to use these AI rollback plan prompts safely A few ground rules for the whole set: Fill in every bracketed section from verified release notes, not from memory. Engineering reviews the rollback steps before anyone executes them. Legal reviews customer communications before they go out. Every checklist has a human's name and timestamp in the sign-off row. Treat AI output as a first draft, not a final document. Never paste PII, credentials, tokens, raw logs, legal disputes, security vulnerabilities, HR information, regulated data, financial records, or anything your security team would flag into a tool that hasn't been cleared for it. The audit trail matters. When the post-mortem happens, "AI drafted it and I approved it" needs to be followed by a specific human name on the approval row. Frequently asked questions What are AI rollback plan prompts? AI rollback plan prompts are copy-paste templates you give to tools like ChatGPT or Claude to help structure a rollback plan quickly. They cover scope definition, owner mapping, trigger criteria, communications, and verification checklists. They don't replace engineering review or management sign-off. Can I use ChatGPT to write a rollback plan? Yes, with caveats. ChatGPT can structure a rollback plan, draft communications, and build checklists fast. It can't verify your actual system state, confirm that procedures are safe, or replace the engineering judgment that makes a rollback work. Use it for structure, not for facts about your infrastructure. What should I never paste into AI when doing rollback planning? Never paste customer PII, database credentials, access tokens, raw production logs, security vulnerabilities, legal dispute details, HR records, regulated data (HIPAA, financial records), or unreleased strategy. Use clean summaries of what happened and what systems were affected. What's the difference between a rollback and a hotfix? A rollback reverts the system to a previous known-good state. A hotfix patches the current broken state forward. Rollbacks are faster when the code change is clean and reversible. Hotfixes are better when the rollback would break data integrity or take longer than the fix. The AI risk assessment prompts article can help you compare blast radius before choosing. How do I make sure a rollback plan is actually approved before execution? Build an explicit approval row into every checklist with a name, role, and timestamp field. Identify your approval chain before the rollback window opens, not during the incident. Engineering and product sign-off should both be required before execution starts. Do I need a rollback plan for every release? For any release touching production systems, user-facing flows, database schemas, or third-party integrations: yes. The level of detail scales with the risk. A config change in a low-traffic service needs a simpler plan than a database migration on a core payment system. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Regression Testing Prompts: Check What Broke AI regression testing prompts for mapping verified changes to risky workflows, past defects, test evidence, and human-owned release decisions. AI promptsregression testingquality assurance AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management --- ## AI Rollout Plan Prompts: Turn Pilot Wins Into Adoption URL: https://dontreplace.me/articles/ai-rollout-plan-prompts-turn-pilot-wins-into-adoption AI rollout planAI adoptionAI promptschange managementAI at work AI Rollout Plan Prompts: Turn Pilot Wins Into Adoption 10 copy-paste AI rollout plan prompts for phased adoption, stakeholder comms, training, data boundaries, success metrics, and rollback criteria. July 21, 2026 Updated July 21, 2026 Most AI pilots end the same way. A small team tries a tool, it mostly works, someone makes a slide deck, and then... nothing. Six months later the tool is still "in evaluation." The stakeholder who championed it got pulled onto something else. The rollout never happened. AI rollout plan prompts can help you break that pattern, but only if you understand what they're actually good for. They're good for structure: turning messy pilot notes into a readable summary, drafting comms, building training outlines, mapping phases. They're not good for inventing evidence you don't have, manufacturing stakeholder buy-in that doesn't exist, or making a half-baked pilot look like a proven business case. That gap between a working pilot and real adoption is where most AI initiatives quietly die. These prompts are for closing it. Before you paste anything into an AI tool: do not include customer PII, credentials, access tokens, private HR records, legal disputes, confidential financial data, regulated health or financial information, unreleased vendor pricing, security vulnerabilities, or private client conversations. Use only approved tools for your organization. The prompt formula every AI rollout plan needs Before getting to the 10 templates, here's the underlying structure. Every good rollout prompt has four parts. Role. Tell the AI what kind of planning assistant it's being. "You are helping an operations lead plan a phased AI rollout." Context. Give it the real situation: what tool, which teams, what the pilot showed, what's still uncertain. This is where garbage in, garbage out actually matters. Vague pilot notes produce vague rollout plans. Task. Be specific about the output. A phased timeline is different from a comms draft is different from a risk matrix. Constraints. Flag what it shouldn't do. "Do not invent readiness. Flag anything that requires compliance or procurement review." This one step saves you from a confidently-written rollout plan that papers over real problems. With that in mind, here are 10 copy-paste prompts. Drop them into your approved AI tool, fill in the brackets, and treat the output as a working draft, not a finished document. AI rollout plan prompts: the 10 templates Prompt 1: Summarize pilot evidence Use this before anything else. If you can't summarize what the pilot actually proved, you're not ready to roll out. You are helping an operations or program lead summarize a completed AI pilot. The pilot involved: [tool name and version], used by [team name and size], for [duration], to handle [specific tasks]. What we measured: [list metrics, e.g., time per task, error rate, user satisfaction score, support tickets]. What we observed (not measured): [qualitative notes from pilot users]. What went wrong or stayed unresolved: [friction points, edge cases, unanswered questions]. Summarize this pilot in three sections: (1) what the evidence supports, (2) what remains unproven, (3) what needs resolution before broader rollout. Do not invent confidence. Flag gaps honestly. Prompt 2: Choose rollout phases One of the fastest ways to kill a rollout is trying to go company-wide too soon. This prompt helps you sequence it. Based on the following pilot summary: [paste your Prompt 1 output, sanitized] Suggest a phased rollout plan with 2-4 phases. For each phase, specify: which teams or roles to include, what conditions must be true before starting the phase, and what success looks like before moving to the next phase. Flag any phases that depend on compliance, procurement, or security approval. Do not assume those approvals exist. Prompt 3: Map affected teams and change impact Most rollout plans focus on who's using the tool. This one forces you to think about who's affected even if they're not the primary users. The tool being rolled out is: [tool name], used primarily by [primary teams]. The tool touches these workflows: [list key workflows]. Identify all teams or roles that are indirectly affected: teams that receive outputs from this tool, teams whose workload or process will change, teams responsible for oversight, compliance, or data governance. For each, describe the likely change impact (low/medium/high) and what they need to know or do before go-live. Prompt 4: Assign owners A rollout plan without named owners is a wish list. This prompt creates an accountability map. The following roles are involved in this rollout: [list roles, e.g., project lead, IT, HR, legal, training lead, team managers, support lead]. For each role, draft a one-paragraph owner brief that describes: their specific responsibility in this rollout, the decisions only they can make, and what they need to deliver and by when. Leave [NAME] placeholders where human names need to be filled in. Do not invent ownership. Every deliverable must have one named owner. Prompt 5: Draft stakeholder communication This is where most rollout comms go wrong. They either oversell the tool or bury the actual changes people need to make. These prompts on AI change management have more on getting buy-in right. Draft a stakeholder communication for [audience: e.g., all-staff, department heads, direct users] about the upcoming rollout of [tool name]. The communication should cover: what the tool does (in plain language, no jargon), what will change for this audience specifically, what the timeline looks like, where to get help or ask questions, and who approved this rollout. Tone: clear and direct. Do not oversell. Do not minimize concerns. Do not claim the tool is perfect. Flag that training will be available and feedback will be collected. Keep it under 300 words. Prompt 6: Build a training plan outline Training is the thing most rollouts skip or do badly. One 45-minute all-hands Zoom is not training. These AI training prompts go deeper if you want to build something real. We are rolling out [tool name] to [team/role]. The primary use cases are: [list 2-4 specific tasks the tool will handle]. Draft a training plan outline that includes: learning objectives (what people need to be able to do, not just know), format options (live session, async video, job aid, practice exercises), content for each module, estimated time per module, and who delivers each component. Include a section on what to do when the tool is wrong or uncertain. Training should create confident users, not dependent ones. Prompt 7: Set data and privacy boundaries This is the prompt most people skip. Don't. Dee Kargaev's framework in Don't Replace Me makes the point bluntly: AI replaces tasks, not judgment, and nowhere is judgment more important than data handling. We are rolling out [tool name] in [industry/context]. The tool will process the following types of data: [list data types, e.g., customer emails, internal reports, HR communications]. Draft a plain-language data boundary document for end users that specifies: what data is approved for use with this tool, what data must never be pasted into this tool (PII, credentials, regulated data, confidential client information, financial records, health data, legal documents), what to do if you're unsure whether data is safe to use, and who to contact for data classification questions. Note: This is a guidance draft. It must be reviewed and approved by your legal, compliance, and privacy teams before distribution. Prompt 8: Create support and escalation paths When something goes wrong at 3pm on a Wednesday, people need to know where to go. Generic "contact IT" instructions don't cut it. We are rolling out [tool name] to [number] users across [teams]. Draft a support and escalation document that includes: tier 1 support (self-serve: FAQs, known issues, how-to guides), tier 2 support (who to contact and how for tool-specific issues), escalation path for suspected data issues or privacy concerns, escalation path for tool errors that affect real work outputs, and who owns the decision to pause or pull back the tool if a serious issue arises. Leave [NAME] and [CONTACT] placeholders where human information needs to be filled in. Every path should end at a named human. Prompt 9: Define success metrics and feedback loops Vague metrics produce fake confidence. "People are using it" is not a success metric. These status report prompts can help you report on rollout progress once you have real numbers. The primary goals of this rollout are: [list 2-3 specific goals, e.g., reduce time spent on X task, reduce error rate in Y process, improve throughput on Z workflow]. For each goal, define: a baseline metric (what does it look like now, before the tool), a target metric (what does success look like at 30/60/90 days), how it will be measured (data source, owner, frequency), and a qualitative feedback mechanism (how you'll hear about friction, confusion, or unintended consequences from users). Flag any goals where baseline data doesn't currently exist. You cannot measure improvement without a starting point. Prompt 10: Write rollback and next-step criteria Every rollout plan needs a rollback trigger. If you don't define it in advance, you'll find yourself rationalizing problems instead of addressing them. We are rolling out [tool name]. Define clear criteria for two scenarios: Scenario A (Rollback): What specific conditions would require us to pause or reverse this rollout? Include thresholds for error rates, user harm, data incidents, adoption failure, and operational disruption. Who has authority to trigger a rollback? What does rollback actually involve operationally? Scenario B (Proceed to next phase): What evidence must exist before we expand to the next group? What approvals are required? What unresolved issues from the current phase must be closed? Do not invent optimism. If the conditions for rollback are vague, the rollback will never happen. What these prompts can't do A well-structured rollout doc is not the same as a successful rollout. A few things no prompt will fix for you. Fabricated buy-in. If your stakeholders haven't actually agreed to this, a polished communication plan won't change that. Get real alignment first. Missing compliance review. If your tool touches regulated data, personal information, or financial records, it needs procurement and legal sign-off. AI-generated guidance documents are drafts, not approvals. Adoption without training. You can deploy a tool to 500 people and have 12 of them actually use it. Rollout and adoption are different things. The training plan prompt helps, but someone has to actually run the training. Invented metrics. If you don't have baseline data before you roll out, you can't prove the tool improved anything. Establish your baseline before go-live, not after. For the full picture on what a good AI pilot looks like before any of this, these AI pilot program prompts are a good starting point. And if you're worried you're missing something before launch day, the AI launch checklist prompts will catch the gaps. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → The one thing most rollout plans pretend not to know Every rollout has a moment where the plan meets reality and reality wins. A team that wasn't consulted. A use case the pilot didn't cover. A manager who feels threatened. A data issue nobody caught in testing. The prompts above will give you clean documents. What they can't give you is the judgment to know when something that looks solved on paper is still a problem in practice. That's Rule #7 in the book's framework: taste is a moat. Knowing where the friction is real, whose feedback to weight, and when a process is just pretending to work. That's the human job. Use the prompts. Don't trust them blindly. Frequently asked questions What should I include in an AI rollout plan? A solid AI rollout plan needs: a summary of real pilot evidence, a phased adoption timeline, named owners for every deliverable, stakeholder communications, a training plan, data and privacy boundaries, support and escalation paths, defined success metrics with baselines, and explicit rollback criteria. Missing any of these creates rollout theater, where everything looks ready but nothing actually works at scale. Can I use ChatGPT to write my AI rollout plan? Yes, with limits. ChatGPT and Claude are useful for drafting structure, comms, training outlines, and checklists. They're not useful for inventing evidence, replacing compliance review, or generating metrics you don't actually have. Never paste sensitive data, PII, credentials, or regulated information into unapproved tools. Treat AI output as a draft that requires human review and approval. How do I measure whether an AI rollout is working? Set baseline metrics before you deploy, not after. If you're rolling out a tool to reduce time spent on a task, measure how long that task takes today. Then measure again at 30, 60, and 90 days. Pair quantitative metrics with a qualitative feedback loop so you hear about friction, confusion, and unintended consequences from actual users. What's the difference between an AI pilot and an AI rollout? A pilot is a controlled test with a small group to validate whether a tool works for a specific use case. A rollout is the planned expansion of a proven tool to broader teams. The mistake most organizations make is treating a promising pilot as proof the rollout will go smoothly. They're different problems. These AI pilot program prompts cover the pilot stage. When should I trigger a rollback during an AI rollout? Define rollback criteria before you launch, not after something goes wrong. Common triggers include: a data incident involving the tool, an error rate that exceeds a defined threshold, adoption failure after adequate training, significant user harm or complaints, or discovery that the tool is being used for purposes outside approved scope. Name the person who has authority to trigger the rollback. If that person isn't named in advance, the rollback usually doesn't happen. How do I get stakeholder buy-in for an AI rollout? Real buy-in comes from involving stakeholders in the pilot, not just notifying them at rollout. If you're announcing a decision rather than sharing evidence and asking for input, you're not building buy-in. Show the pilot data, acknowledge what's still unproven, name the owners, and give people a real channel to raise concerns. For more on this, these AI change management prompts are worth a look. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Change Management Prompts: Get Buy-In Fast 10 copy-paste AI prompts for change management: announcements, stakeholder maps, resistance diagnosis, training, adoption metrics, and rollback plans. ai change managementchange management promptsai adoption AI Customer Service Prompts: 10 Templates That Work 10 copy-paste AI customer service prompts for ChatGPT and Claude. Draft replies, de-escalate complaints, and summarize tickets without sounding like a robot. AI promptscustomer serviceChatGPT at work AI Launch Checklist Prompts: Ship Without Gaps 10 copy-paste AI launch checklist prompts for ChatGPT and Claude. Turn messy plans into structured checklists with owners, risks, and rollback steps. ai promptslaunch checklistproject management --- ## AI Root Cause Analysis Prompts: Stop Fixing Symptoms URL: https://dontreplace.me/articles/ai-root-cause-analysis-prompts-stop-fixing-symptoms AI promptsroot cause analysisRCAQAoperations AI Root Cause Analysis Prompts: Stop Fixing Symptoms AI root cause analysis prompts for turning messy incidents, bugs, timelines, evidence, and corrective actions into clearer RCA work. August 04, 2026 Updated August 04, 2026 Root cause analysis is where teams pretend they want the truth, then quietly hope the truth fits inside one polite sentence. It usually does not. The thing that broke was probably not caused by one lazy person, one cursed deploy, one confused customer, or one ticket that “should have been clearer.” It was more likely a chain of small decisions, missing checks, unclear ownership, weak signals, rushed timelines, and one spreadsheet everyone respected for no obvious reason. That is where AI root cause analysis prompts can help. Not because AI can inspect your systems, prove causality, read production logs it has never seen, or decide which team should eat the blame sandwich. It cannot. But it can turn incident notes, defect patterns, customer complaints, support tickets, QA findings, timeline notes, system-change logs, process observations, interview notes, evidence summaries, and follow-up actions into a cleaner RCA draft. If you already know the fix and need an action plan, use AI remediation plan prompts. If you are still figuring out what actually happened, stay here. AI can organize root cause analysis. A human still owns the facts, evidence, judgment, accountability, approvals, and final call. What is root cause analysis? Root cause analysis, usually shortened to RCA, is the process of understanding why a problem happened so the team can reduce the chance it happens again. A useful RCA answers: What happened? Who or what was affected? When did the issue start, worsen, get noticed, and get resolved? What evidence supports the timeline? What symptoms showed up? What likely causes contributed? Which causes are confirmed, and which are still hypotheses? What checks failed to catch the problem earlier? What corrective actions will reduce repeat risk? Who owns follow-up, verification, and sign-off? A bad RCA answers one question: “How do we make this go away before Friday?” That is how teams end up fixing symptoms. They patch the visible problem, rename the ticket, add three vague action items, and move on. Then the same failure returns in a different hat. AI root cause analysis prompts are useful because AI is good at structure. It can group symptoms, draft timelines, compare possible contributing factors, run a careful 5 Whys exercise, flag missing evidence, prepare interview questions, and turn findings into action-plan language. But it follows the same rule as what AI can and can't do: AI can make a messy problem easier to inspect. It cannot make an unverified story true. The reusable AI root cause analysis prompt formula Use this formula whenever you ask AI to help with RCA work: “You are a root cause analysis assistant. I am [your role] investigating [incident, defect, customer problem, process failure, launch issue, audit gap, or operational miss]. Here is privacy-safe context: [confirmed facts, timeline notes, symptoms, affected systems, user impact, support themes, QA findings, deploy/change notes, interview summaries, known constraints, current hypotheses, and open questions]. Create [specific output]. Separate confirmed facts from assumptions. List missing evidence, possible contributing factors, verification questions, owner questions, risk concerns, and recommended next checks. Do not invent facts, logs, root causes, blame, customer impact, approvals, due dates, security conclusions, legal conclusions, or final decisions.” That last sentence is not decorative. AI loves finishing the pattern. In RCA work, finishing the pattern can turn “we need to check whether the deploy caused it” into “the deploy caused it.” Congratulations, you just manufactured confidence with grammar. Before using any prompt, sanitize the input. Do not paste customer PII, credentials, access tokens, raw production logs, private tickets, security vulnerabilities, unreleased strategy, legal disputes, financial records, HR issues, regulated data, medical information, private client conversations, or sensitive personal information into unapproved AI tools. Use summaries, categories, redacted excerpts, internal ticket IDs, approved screenshots, and links to authorized systems instead. The model does not need a customer’s phone number to help you ask better RCA questions. It is nosy enough already. What to bring before asking AI for RCA help AI cannot analyze a ghost story. It can only organize the evidence you provide. If the evidence is vague, political, or missing, the output will be confident fog. Bring this first: InputWhy it mattersHuman check Confirmed problem statementKeeps the RCA from driftingVerify against evidence, not vibes Incident or defect timelineShows sequence and delayMark exact times vs. estimates Affected systems or processesDefines scopeConfirm with system or process owners User or customer impactSets seriousnessUse approved, privacy-safe summaries Symptoms observedSeparates visible pain from causeGroup repeated symptoms carefully Change historySurfaces likely triggersInclude deploys, config changes, handoffs, policy changes Existing hypothesesGives AI something to testLabel each as unconfirmed until proven Missing evidencePrevents fake certaintyAsk owners what data is needed Corrective action ideasConnects analysis to preventionDo not approve fixes by chatbot Verification criteriaProves the fix mattersDefine what evidence will close the loop If you do not have these inputs, ask AI to create a missing-evidence checklist first. Do not ask it to explain a system failure from one screenshot and a Slack thread where everyone uses the word “weird.” For upstream sorting, pair this with AI defect triage prompts and AI QA checklist prompts. Better evidence makes better RCA. Bad evidence makes premium nonsense. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI root cause analysis prompts you can use today Prompt 1: Turn incident notes into an RCA brief Use this when your notes are scattered across tickets, Slack, meetings, and someone’s memory palace. “Act as an RCA structuring assistant. Using the privacy-safe notes below, create a concise RCA brief with these sections: problem statement, confirmed facts, affected users or systems, timeline, symptoms, known changes, likely contributing factors, open questions, missing evidence, immediate containment, proposed corrective actions, verification criteria, and owner questions. Clearly label assumptions. Do not invent root cause, impact, owners, or evidence. Notes: [paste sanitized notes].” This prompt is useful because it turns chaos into a first draft. It is not the final RCA. It is the tray where you sort the broken pieces before pretending you are calm. After the draft, check every “fact” against the source. AI may merge two similar events, smooth over time gaps, or make a hypothesis sound confirmed. That is not malice. It is autocomplete wearing a tiny consultant vest. Prompt 2: Separate symptoms from possible causes Use this when the team keeps treating visible pain as the explanation. “Review the following issue summary. Create a two-column table: observed symptoms and possible underlying causes. For each possible cause, add supporting evidence, contradicting evidence, missing evidence, and recommended next check. Label each cause as confirmed, likely, possible, or unsupported. Do not treat a symptom as a root cause unless the evidence proves it. Summary: [paste sanitized summary].” Symptoms are things people noticed: slow load time, failed login, duplicate email, angry customer, missing report, confusing handoff. Causes are why those things happened: bad dependency assumption, missing validation, broken alerting, unclear ownership, untested migration path, stale documentation, weak release gate. AI can help separate those buckets. Humans still decide what the evidence proves. Prompt 3: Build a privacy-safe timeline Use this when nobody agrees when the problem started. “Create an RCA timeline from these privacy-safe notes. Include time, event, source, confidence level, affected area, and open questions. Separate confirmed timestamps from estimated timestamps. Highlight gaps where we need logs, ticket history, monitoring data, customer confirmation, deploy records, or owner interviews. Do not invent times or sequence. Notes: [paste sanitized timeline notes].” Timelines expose the shape of the failure. They show whether detection lagged, escalation wandered, ownership changed, monitoring missed the signal, or the fix arrived three handoffs after everyone thought someone else had it. This is especially useful after launches, migrations, and cutovers. For those, also read AI cutover plan prompts and AI post-launch monitoring prompts. Prompt 4: Run 5 Whys without blame theater Use this when the team wants a simple cause but the problem deserves a better chain. “Facilitate a 5 Whys analysis for the issue below. For each why, provide: question, tentative answer, evidence that supports it, evidence still needed, and whether the answer points to people, process, tooling, communication, quality control, monitoring, training, or decision-making. Avoid blaming individuals. Stop when evidence is missing and list what to verify next. Issue: [paste sanitized issue].” The point of 5 Whys is not to keep asking “why” until someone looks guilty. The point is to move from surface symptom to contributing system conditions. Bad 5 Whys: “Why did the bug ship? Because Alex missed it.” Better 5 Whys: “Why was the bug not caught? The test case did not cover migrated accounts. Why? The migration scenario was not in acceptance criteria. Why? Requirements assumed only net-new users. Why? Customer-data edge cases were not reviewed before scope lock.” Notice how the second version gives you work to fix. The first gives you office drama with a timestamp. Prompt 5: Group contributing factors Use this when there is no single root cause, which is annoying because reality refuses to respect slide decks. “Group the following evidence into contributing-factor categories: technical, process, communication, ownership, training, monitoring, documentation, timeline pressure, vendor dependency, customer data, and decision-making. For each category, summarize the factor, cite the evidence provided, rate confidence, identify missing evidence, and suggest one prevention-oriented question. Evidence summary: [paste sanitized evidence].” Many problems are not one domino. They are ten tiny dominoes, three missing labels, and one person saying, “I thought that was handled by the other team.” AI can help you see patterns across messy notes. It can cluster repeated handoff failures, unclear status updates, weak acceptance criteria, and missing monitoring signals. That pairs well with AI incident review prompts if the event already caused visible damage. Prompt 6: Find missing evidence before conclusions harden Use this before leadership falls in love with the first plausible story. “Review this draft RCA. List every conclusion that needs stronger evidence. For each one, provide: claim, current support, risk if wrong, evidence needed, likely source system or owner, and suggested follow-up question. Flag any statement that sounds more certain than the evidence supports. Draft: [paste sanitized RCA draft].” This prompt is a nonsense detector. It helps catch lines like: “The outage was caused by the deployment.” Was it, or did the deployment reveal another issue? “Support did not escalate quickly enough.” Based on what escalation rule? “Users were minimally affected.” Who verified impact? “The fix is complete.” Where is retest evidence? A good RCA gets stricter before it gets prettier. Prompt 7: Prepare stakeholder interview questions Use this when the evidence is stuck in people’s heads. “Create stakeholder interview questions for an RCA about [issue]. Generate questions for engineering, QA, support, customer success, operations, product, and leadership. For each question, explain what evidence it is trying to uncover. Keep the tone neutral and blame-free. Do not imply fault. Context: [paste privacy-safe context].” Interviews can either uncover reality or trigger self-defense theater. The wording matters. Ask “What signals did we have before detection?” instead of “Why did you miss it?” Ask “What ownership boundary was unclear?” instead of “Who dropped the ball?” Ask “What would have made this easier to catch?” instead of “Please confess into this spreadsheet.” Humans handle the conversation. AI can help you avoid accidentally writing questions like a tiny courtroom prosecutor. Prompt 8: Draft corrective actions from causes Use this when the RCA has findings but the action items are mush. “Turn these confirmed and likely contributing factors into corrective-action options. For each action, include the related cause, expected risk reduction, owner type, dependency, verification method, possible downside, urgency, and whether the action is immediate, short-term, or long-term. Do not mark any action approved. Flag actions needing security, legal, customer, engineering, QA, or leadership review. Findings: [paste sanitized findings].” Corrective actions should change the system, not decorate the ticket. Weak action: “Be more careful.” Better action: “Add migrated-account scenarios to acceptance criteria and smoke tests before billing changes ship.” Even better: “Add migrated-account billing scenarios to the release checklist, assign QA ownership, require evidence links before launch approval, and verify failures alert in post-launch monitoring.” For the fix-planning part, AI remediation plan prompts goes deeper. Prompt 9: Create verification checks Use this when everyone wants to close the RCA but nobody has proof the fix works. “Create a verification checklist for the corrective actions below. For each action, list what evidence would prove it worked, what evidence would show it failed, who should verify it, what system or record should hold proof, and when follow-up should happen. Include regression checks, customer-impact checks, monitoring checks, and documentation updates where relevant. Actions: [paste corrective actions].” This is where the RCA stops being a document and starts being useful. Verification keeps teams from closing action items because the status changed to green in a tool. Green is a color, not evidence. Annoying but true. If the issue happened during a launch, combine this with AI smoke test prompts and AI rollback plan prompts. Prevention beats heroic cleanup. Prompt 10: Write an executive RCA summary Use this when the analysis is long but leaders need the truth without the forensic novel. “Rewrite this RCA into an executive summary. Include: what happened, confirmed impact, timeline summary, most likely root causes or contributing factors, what we know, what remains unconfirmed, immediate containment, corrective actions, verification plan, owners, and decision needs. Keep it clear, calm, and factual. Do not hide uncertainty. Do not invent blame, certainty, impact, or commitments. RCA draft: [paste sanitized draft].” Executives do not need every log line. They do need an honest summary that does not sand off the risk until it becomes useless. Good executive RCA language sounds like this: “The strongest evidence points to two contributing factors.” “Customer impact is confirmed for these segments and still being checked for these others.” “The permanent fix is approved, but verification is not complete.” “This action reduces repeat risk; it does not eliminate all related failure modes.” That kind of language is less sexy than “root cause identified.” It is also less likely to embarrass everyone later. A simple RCA workflow using AI Here is the sane order: Gather privacy-safe evidence. Ask AI to structure the RCA brief. Separate symptoms from possible causes. Build a verified timeline. Run 5 Whys carefully. Group contributing factors. Ask AI to flag missing evidence and overconfident claims. Interview owners with neutral questions. Draft corrective actions. Create verification checks and an executive summary. Do not start with “write the RCA.” That invites the model to create a polished story too early. Start with structure and questions. Make the evidence earn the conclusion. This also helps with AI operational readiness prompts and AI contingency plan prompts. The best RCA is the one you never need because your checks caught the failure earlier. Sadly, we live on Earth. What AI should not do in root cause analysis AI should not: Declare the root cause without evidence. Assign blame to a person or team. Invent timelines, logs, impact, or approvals. Convert guesses into confirmed facts. Recommend risky production changes without expert review. Write legal, security, HR, financial, medical, or customer-impact conclusions as if approved. Decide risk acceptance. Close corrective actions. Replace the people who understand the system. Use AI as a sharp formatting and questioning assistant. Do not use it as a judgment vending machine. If you want the broader starter framework, read how to use AI at work without the hype. Same rule applies here: give AI bounded jobs, inspect the output, and keep the human accountable. Frequently asked questions Can ChatGPT do root cause analysis? ChatGPT can help structure root cause analysis, summarize sanitized notes, draft timelines, run 5 Whys, group contributing factors, and identify missing evidence. It cannot verify what happened. You still need approved evidence, expert review, and human sign-off. What should I include in an AI root cause analysis prompt? Include a clear problem statement, privacy-safe timeline, symptoms, affected systems, customer or user impact, known changes, evidence summaries, hypotheses, constraints, and open questions. Ask the model to separate confirmed facts from assumptions and to flag missing evidence. Is it safe to paste incident logs into AI tools? Only if your organization has approved that tool and the logs are allowed to be shared there. In most normal situations, use privacy-safe summaries instead. Do not paste credentials, tokens, PII, raw production logs, security vulnerabilities, regulated data, or private customer information into unapproved AI tools. How do I keep AI from inventing root causes? Tell it explicitly not to invent facts, logs, causes, impact, approvals, owners, or final decisions. Require labels such as confirmed, likely, possible, and unsupported. Then verify every claim against the source evidence. What is the difference between RCA and remediation? RCA explains why a problem happened. Remediation defines what will be done to reduce repeat risk and prove the fix worked. They are connected, but they are not the same thing. A beautiful RCA without corrective action is trivia with formatting. Should AI assign owners for RCA actions? AI can suggest owner types, like engineering, QA, support, operations, product, or leadership. A human should assign real owners. Fake ownership is how action items go to the farm upstate. Final thought AI root cause analysis prompts are useful when they make the work clearer, stricter, and less political. They are dangerous when they make weak evidence sound finished. Use AI to organize the mess, ask better questions, surface missing evidence, and draft cleaner summaries. Keep humans in charge of facts, causality, customer impact, risk, approvals, and accountability. That is the whole survival move: let the machine help with the structure while you keep the judgment. Dee Kargaev’s book Don’t Replace Me is basically a field guide for that posture: stay useful while everyone else mistakes a tidy document for knowing what happened. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Change Impact Analysis Prompts: Spot Damage AI change impact analysis prompts for mapping affected systems, teams, customers, QA checks, support gaps, and rollout risks. AI promptschange impactrisk AI Remediation Plan Prompts: Fix Root Causes AI remediation plan prompts for turning findings, defects, incidents, owners, deadlines, fixes, and retest evidence into real corrective action. AI promptsremediationcorrective action AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management --- ## AI Sales Call Coaching Prompts: Improve the Conversation URL: https://dontreplace.me/articles/ai-sales-call-coaching-prompts-improve-the-conversation AI promptssales call coachingsales enablementsales managementconversation review AI Sales Call Coaching Prompts: Improve the Conversation 10 AI sales call coaching prompts to review evidence, improve discovery, practice objections, protect private data, and coach reps fairly. September 02, 2026 Updated September 02, 2026 Sales call coaching often begins with a noble goal and ends with somebody announcing that a rep said “um” fourteen times. Congratulations: the revenue problem has been solved. AI sales call coaching prompts can turn approved transcripts into structured observations, find unanswered questions, prepare practice drills, and help a manager run a more specific debrief. They cannot read a buyer’s mind, determine whether a rep is good at their job from one call, or turn a bad transcript into truth. AI can organize evidence from a sales conversation. Humans still own consent, context, coaching, privacy, employment decisions, and customer trust. These ten templates help managers, founders, enablement teams, and account executives use AI for sales coaching without appointing autocomplete as regional vice president of vibes. What sales call coaching actually is Sales call coaching is a repeated human process: observe a conversation, compare what happened with an approved standard, understand the context, select one useful behavior to improve, practice it, and check progress across future calls. Keep these ideas separate: Observation: something directly supported by the approved transcript or recording. Interpretation: a possible explanation, such as “the buyer sounded hesitant.” It is not a fact. Outcome: what happened after the call. A positive outcome does not make every behavior good. Rubric: the team’s explicit standard for discovery, accuracy, listening, next steps, and conduct. Coaching: support intended to improve a skill through reflection and practice. Evaluation: a formal employment process with different evidence, policy, and review requirements. Buyer intent: what the buyer actually says or later confirms—not an emotion guessed from punctuation. One transcript is one sample. It may omit screen sharing, chat, prior calls, relationship history, product constraints, or the manager’s own instructions. Treating it as a complete psychological scan is how coaching becomes surveillance wearing a quarter-zip. If the underlying opportunity is already messy, start with an evidence-first sales pipeline review. For reusable discovery and follow-up language, use these general AI sales prompts. The evidence-first call coaching formula Add this instruction to any prompt below: “Act as a sales call review assistant. Use only the sanitized, approved transcript, notes, rubric, and context I provide for [call type, buyer stage, product, and coaching goal]. Produce [artifact]. Separate direct observations, quoted evidence, interpretations, unknowns, and suggestions. Cite a timestamp or excerpt for every material observation. Flag transcript uncertainty, missing speakers, missing context, and claims requiring product, legal, security, pricing, or managerial review. Do not infer emotion, personality, protected characteristics, health, competence, intent, honesty, or future performance. Do not score or rank employees, recommend employment action, fabricate buyer needs, or present a single call as a pattern.” The most important instruction is use only the evidence provided. “Unknown” is a legitimate result. If the buyer’s reaction is not in the transcript, the model does not get to write fan fiction about it. Never paste unapproved recordings, buyer names, emails, phone numbers, contracts, private pricing, payment details, credentials, employee performance notes, protected characteristics, confidential strategy, or legally sensitive information into an unapproved AI tool. Follow recording-consent laws and company policy. Use approved systems, minimum necessary data, de-identification, access controls, retention limits, and human review. What to collect before prompting Build a small, sanitized coaching pack. Keep source records in approved systems. InputWhy it mattersHuman check Call purpose and stageDefines reasonable expectationsManager confirms Approved transcript or notesSupplies observable evidenceCall owner checks accuracy Speaker labels and timestampsSupports traceable feedbackRep corrects attribution Team coaching rubricPrevents invented standardsEnablement owns version Product and pricing factsTests accuracy safelyProduct or commercial owner verifies Prior customer contextPrevents one-call conclusionsAccount owner summarizes Rep’s own reflectionMakes coaching collaborativeRep provides voluntarily Known transcript gapsLimits false confidenceReviewer documents Consent and access rulesProtects people and customersLegal/privacy approves process One coaching goalKeeps feedback usableManager and rep agree Follow-up evidenceChecks whether next steps happenedCRM owner verifies Do not upload the whole CRM because you want feedback on one discovery question. Data minimization is cheaper than an incident review. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI sales call coaching prompts Replace brackets with sanitized, approved evidence. Every output is a draft for human review, not an employee scorecard. 1. Clean and qualify the transcript “Review this approved transcript and call metadata: [paste]. Create a transcript quality note showing speaker-attribution uncertainty, missing sections, low-confidence wording, likely transcription errors, unexplained acronyms, absent screen-share context, and statements that need confirmation. Preserve the original wording beside any proposed correction. Do not silently rewrite what a speaker said.” Start here because a transcript can confidently turn “annual contract value” into “animal contract value” and then critique the rep’s farm strategy. Have the call owner confirm material corrections. If the system cannot distinguish speakers, do not attribute a questionable statement to the rep. Mark it unknown. Exclude personal chatter and irrelevant identifiers before further analysis. This step is preparation, not coaching. It creates a source-quality warning label so later outputs do not convert transcription noise into performance feedback. 2. Separate observations from interpretations “Create a two-column review of this sanitized transcript: [paste]. In the first column, list direct observations with timestamps or short excerpts. In the second, list possible interpretations, each labeled as tentative. Add a third section for information the transcript cannot establish. Remove any claim about emotion, personality, intelligence, honesty, motivation, or buyer intent unless directly stated.” Managers naturally interpret conversations. The problem begins when an interpretation such as “the buyer was disengaged” gets stored as objective fact. Maybe the buyer was reading a technical document. Maybe the transcript dropped their questions. Maybe they were disengaged. The evidence should show which answer is supportable. Use this output to ask the rep what they noticed. Coaching gets better when the person in the call can add context instead of receiving a machine-generated verdict. 3. Check discovery coverage “Compare this approved discovery rubric with the sanitized call transcript: [paste rubric and transcript]. For each rubric area, mark addressed, partly addressed, not observed, or not applicable. Cite evidence. List unanswered questions without claiming they should all have been asked on this call. Distinguish buyer-confirmed facts, seller assumptions, and items requiring follow-up.” A discovery checklist is a map, not an interrogation script. The rep does not need to ask every possible question while the buyer ages visibly on Zoom. Review whether the conversation established the problem, current process, impact, stakeholders, constraints, decision path, timing, and a mutual next step at the appropriate depth. Then choose one or two gaps that mattered for this call. For planning better customer questions without manufacturing answers, adapt the evidence discipline in these AI customer interview prompts. 4. Review question quality “Analyze only the questions asked by the seller in this sanitized transcript: [paste]. Group them as open, closed, leading, stacked, clarifying, reflective, or confirmation questions. Cite each example. Explain the likely conversational function without guessing buyer emotion. Suggest up to five cleaner alternatives that preserve the seller’s intent and do not assume facts not in evidence.” A question can be grammatically open and still steer the buyer toward the seller’s preferred answer. “How painful is your terrible manual process?” is less discovery and more hostage negotiation. Look for stacked questions that give the buyer six exits, leading language, premature solution framing, and missed opportunities to clarify vague words such as “slow,” “expensive,” or “soon.” Also preserve good questions. Coaching that only hunts mistakes teaches reps to fear review rather than learn from it. 5. Find unsupported assumptions and risky claims “Review this sanitized transcript for unsupported assumptions and seller claims: [paste]. Create sections for buyer-confirmed facts, seller interpretations, product claims, pricing or contract statements, security/legal/compliance claims, competitor statements, and unresolved questions. Cite each excerpt. Flag items for the appropriate human owner to verify. Do not decide that a claim is false merely because supporting evidence is absent from the transcript.” Sales calls move quickly. A rep may summarize a buyer’s need too broadly or answer a technical question from memory. AI can locate statements for review; it should not become the official product manual. Route product claims to product owners, security claims to security, contract language to legal, and pricing exceptions to commercial leadership. The coaching goal is accurate communication and better escalation—not catching somebody in a gotcha spreadsheet. Use an AI risk assessment prompt when a questionable claim may affect customers, compliance, or a live deal. 6. Examine participation without fake precision “Using this approved transcript, summarize observable participation patterns: [paste]. Estimate speaking turns and clearly label any talk-time number as approximate unless verified by the recording system. Note interruptions, long seller monologues, unanswered buyer questions, follow-up questions, summaries, and explicit checks for understanding. Cite examples. Do not infer listening skill, confidence, dominance, engagement, gender, accent, disability, or personality.” Talk ratio is context, not a universal grade. A product demonstration, contract review, technical workshop, and first discovery call should not have identical ratios. A buyer may ask for a detailed explanation. Multiple buyer participants can also distort the number. The useful question is not “Did the rep hit 43 percent?” It is “Did the conversation make enough room for the buyer, and did the rep respond to what was actually said?” Managers should review the recording and context before turning an estimate into coaching. 7. Prepare objection-handling practice “From this sanitized transcript, extract only objections or concerns the buyer explicitly stated: [paste]. For each, quote the evidence, summarize the seller’s response, list information still needed, and draft three practice responses: a clarifying question, an evidence-based response using only these approved product facts [paste], and an honest escalation when the answer is unknown. Do not invent urgency, case studies, discounts, competitor claims, or product capabilities.” Good objection handling begins by understanding the concern, not loading the verbal cannon. A security question may need a specialist. A timing objection may hide an implementation dependency. A budget concern may simply mean the project is not funded. Role-play the alternatives with the rep. Ask them to choose language that sounds like them. AI-generated dialogue often arrives wearing a blazer and saying “I completely understand your concern” like a customer-service android from 2007. 8. Draft a collaborative coaching debrief “Turn these verified call observations, the rep’s reflection, and our approved rubric into a 30-minute coaching debrief: [paste]. Include: one success to preserve, one behavior to explore, evidence excerpts, open questions for the rep, the relevant rubric standard, one practice activity, one mutually agreed next action, and a check-in date. Keep interpretations tentative. Do not diagnose motives, assign a performance rating, or recommend disciplinary action.” Specific feedback is more useful than “be more consultative.” Anchor the discussion in evidence and invite the rep’s view before deciding what happened. A strong debrief might focus on clarifying business impact after a vague buyer answer. It should not dump twenty-seven machine observations on someone and call that enablement. Use these AI feedback prompts to make the final language direct without making it robotic. Formal evaluation belongs in the organization’s approved human process. Coaching records can become employment records, so limit access, follow retention policy, and involve HR when appropriate. 9. Create one focused improvement drill “Using this human-approved coaching goal and sanitized evidence, create a 15-minute practice drill: [paste]. Include the skill, scenario, buyer facts, constraints, success criteria, three escalating role-play turns, reflection questions, and a simple repeat plan. Do not add private customer details or assess personality. Create plausible practice material, clearly labeled fictional, rather than pretending it describes the real buyer.” Practice should be narrow enough to repeat. “Improve discovery” is not a drill. “When a buyer says implementation is slow, ask one clarifying question, summarize the answer, and confirm impact before proposing a feature” is. A fictional role-play scenario is fine when labeled. Keep it separate from factual call analysis so invented details do not leak into the opportunity record. For a broader learning sequence, adapt these AI training prompts. 10. Draft an accurate customer follow-up “Draft a concise follow-up email from this sanitized transcript and approved notes: [paste]. Include only buyer-confirmed priorities, decisions, open questions, owners, and dates. Mark uncertain details with [VERIFY]. Separate seller commitments from buyer commitments. Do not invent enthusiasm, urgency, budget, authority, product capability, pricing, legal terms, or next steps. End with a simple request to correct anything misunderstood.” A coaching process should improve the customer experience, not merely produce an internal score. The follow-up is where misunderstanding can be corrected while the call is fresh. The account owner must verify every commitment, date, and commercial statement before sending. Do not paste the model’s summary directly into the CRM and let future forecasts inherit its guesses. If the follow-up changes the opportunity, update the source system with verified facts. A humane sales call coaching workflow Use the prompts in a controlled sequence: Confirm recording consent, purpose, access, and retention rules. Minimize and sanitize the call data. Let the rep verify transcript errors and add missing context. Run transcript qualification before any analysis. Compare observations with an approved rubric. Invite the rep’s self-reflection. Select one useful coaching goal together. Practice the behavior in a fictional, clearly labeled scenario. Agree on a next action and evidence for progress. Review a pattern across future calls instead of declaring victory or failure from one sample. Keep AI outputs out of formal evaluation unless the organization has explicitly approved the use, evidence standard, appeal process, privacy controls, bias review, and human accountability. “The tool suggested it” is not due process. Common ways AI call coaching goes wrong It treats transcription as ground truth Accents, audio quality, crosstalk, acronyms, product names, and bad microphones create errors. Require excerpts and timestamps, then verify important moments against the authorized source. It performs fake emotion recognition A transcript cannot reliably establish frustration, confidence, honesty, engagement, or personality. Even audio and video inference can be biased and inappropriate. Ask what was said and what happened next. It turns coaching into secret surveillance People should know what is recorded, why it is reviewed, who can access it, how long it is kept, and how outputs may be used. Follow applicable law, policy, and consent requirements. It rewards one rigid call style Top sellers can sound different. Buyers, cultures, products, and stages differ too. Coach outcomes and evidence-based behaviors, not imitation of one charismatic rep’s verbal tics. It confuses correlation with causation A won deal does not prove every behavior was effective. A lost deal does not prove the rep caused the loss. Compare patterns, context, and multiple sources before drawing conclusions. It creates an automated performance score A neat number hides subjective choices about rubrics, data quality, context, and weighting. Do not rank reps or make employment decisions from generated call analysis. Frequently asked questions Can ChatGPT analyze a sales call transcript? It can structure an approved, sanitized transcript, identify passages that match a rubric, and draft review questions. It cannot verify the transcript, know hidden context, or reliably infer buyer intent. A human must review the source and conclusions. What is the best AI prompt for sales coaching? Use a prompt that defines the call type, approved rubric, coaching goal, permitted evidence, and output format. Require citations for observations and separate facts, interpretations, and unknowns. The evidence-first formula above is a strong starting point. Can AI score sales representatives from calls? It should not be treated as an objective employee scorer. Transcript quality, call mix, customer context, rubric design, and model bias can distort results. Keep coaching collaborative and keep employment decisions inside an accountable human process. Is it legal to upload recorded sales calls to AI? That depends on location, consent, contracts, company policy, data type, tool terms, and how outputs are used. Use only approved systems and ask legal or privacy professionals about your actual situation. This article is operational guidance, not legal advice. How do I protect customer data during AI call review? Use minimum necessary data, remove direct identifiers, exclude credentials and sensitive details, restrict access, set retention limits, and use an organization-approved tool. A sanitized excerpt is usually safer than an entire recording and CRM history. Can AI improve objection handling? Yes, as a practice assistant. It can extract explicitly stated concerns and draft alternative questions using approved facts. Humans must verify product, pricing, legal, security, and competitor claims before using them with a customer. How many calls should a manager review before identifying a pattern? There is no universal number. Use multiple comparable calls over time, consider call type and account context, and look for repeated evidence. One call can support a coaching conversation; it rarely supports a broad conclusion about ability. Should reps see the AI-generated coaching notes? A transparent process is generally healthier than secret scoring. Reps should be able to correct transcript errors and add context. Follow company policy and applicable employment, recording, and privacy rules for access and retention. Use the machine for structure, not judgment AI is useful when it turns a long conversation into traceable observations, unanswered questions, and a focused practice plan. It becomes dangerous when guesses about tone or intent harden into employee records and customer “facts.” Use approved evidence. Cite the source. Protect the people in the conversation. Let the rep respond. Keep coaching specific, limited, and human-owned. That is the broader lesson in what AI can and cannot do: speed and polish are not understanding. If you want the larger field guide for working with that reality, Don’t Replace Me by Dmitry Kargaev is the low-drama version. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Account Planning Prompts: Build a Better Plan 10 AI account planning prompts to organize buyer evidence, map stakeholders, expose risks, test strategy, and plan next actions without inventing intel. AI promptsaccount planningsales strategy AI Sales Enablement Prompts: Build Useful Materials 10 AI sales enablement prompts for grounded messaging, training, coaching, role-play, objection practice, and seller resources without invented claims. AI promptssales enablementsales training AI Sales Forecast Prompts: Skip the Fake Certainty 10 AI sales forecast prompts to test evidence, build scenarios, challenge assumptions, protect buyer data, and produce an accountable forecast. AI promptssales forecastingrevenue operations --- ## AI Sales Enablement Prompts: Build Useful Materials URL: https://dontreplace.me/articles/ai-sales-enablement-prompts-build-useful-materials AI promptssales enablementsales trainingrevenue operationssales coaching AI Sales Enablement Prompts: Build Useful Materials 10 AI sales enablement prompts for grounded messaging, training, coaching, role-play, objection practice, and seller resources without invented claims. September 05, 2026 Updated September 05, 2026 Sales enablement should help sellers understand the product, ask better questions, explain value accurately, and handle uncertainty like adults. It should not create a landfill of battlecards nobody trusts. AI sales enablement prompts can turn approved product information, messaging, call notes, objection logs, and process documentation into useful drafts. They can reorganize material, create practice scenarios, adapt formats, and expose missing evidence. They cannot verify a product claim, know what a buyer meant, approve legal language, or coach nuanced human behavior without a competent manager. AI can accelerate the paperwork around enablement. Humans still own truth, judgment, coaching, privacy, approval, and revenue outcomes. These ten templates help enablement leaders, revenue operations teams, sales managers, founders, and consultants build seller resources without promoting autocomplete to chief revenue officer. What useful sales enablement actually does Enablement is controlled translation. It translates product knowledge, customer evidence, market context, and company process into behavior a seller can use. The useful version answers questions such as: What can we truthfully say today? Which source supports each claim? What should a seller ask before recommending anything? Which objections recur, and what evidence helps address them? Where does general guidance stop and deal-specific judgment begin? What requires product, legal, security, finance, or leadership approval? How will a manager observe whether training changed behavior? The useless version measures output volume: decks created, modules completed, battlecards downloaded, boxes ticked. Completion is not competence. A seller can finish a course while answering customer questions with magnificent confidence and entirely fictional details. Keep these labels separate in every AI-assisted artifact: Approved fact: current, sourced, and cleared for the intended audience. Observed customer language: what a customer actually said, suitably de-identified. Interpretation: a human reading of evidence, not evidence itself. Draft language: wording awaiting review. Hypothesis: an idea to test in discovery or coaching. Unknown: information nobody has verified. Restricted information: material that cannot enter the tool or output. Owner: the human accountable for review and maintenance. If you blur those categories, a draft becomes a claim, a claim becomes a promise, and a promise becomes a miserable call with legal. For adjacent workflows, use AI sales prompts for discovery and follow-up and AI account planning prompts. Enablement should support those conversations, not replace them with scripts delivered in the emotional register of airport signage. The evidence-first sales enablement formula Add this instruction to any prompt below: “Act as a sales enablement drafting assistant. Use only the sanitized, approved sources I provide for [audience, selling motion, product, region, objective, and artifact]. Cite a source label and date for every product, customer, competitor, process, pricing, security, legal, or performance claim. Separate approved facts, observed language, interpretations, draft wording, hypotheses, and unknowns. Mark unsupported or stale items [VERIFY]. Do not invent capabilities, roadmap items, customer stories, results, competitive intelligence, guarantees, prices, policies, or buyer motives. Preserve contradictions and route specialist questions to the named human owner.” That instruction makes the model less entertaining and more useful. Good. Sellers need reliable material, not a gifted improv partner inventing enterprise features in real time. Never paste customer names, personal contact details, credentials, private call recordings, contracts, pricing exceptions, confidential roadmaps, regulated data, or legally sensitive information into an unapproved AI tool. Use approved systems, minimum necessary data, de-identification, access controls, retention limits, and human privacy, legal, security, product, and commercial review. What to collect before prompting Create a small, controlled source pack. Do not dump the entire revenue drive into a chat window and hope governance emerges. InputWhy it mattersHuman owner Current approved messagingDefines what sellers may sayMarketing or product marketing Product capability documentationGrounds feature claimsProduct owner Roadmap policyPrevents accidental promisesProduct leadership Approved proof pointsSupports outcome languageMarketing and legal De-identified objection examplesGrounds practice in real patternsEnablement lead Sales process and stage criteriaConnects training to executionRevenue operations Approved pricing guidancePrevents invented discountsCommercial owner Security and legal escalation routesSends specialist questions correctlySecurity and legal Role and experience levelSets useful difficultySales manager Source labels, dates, and expiry datesMakes claims traceableContent owner Desired behavior and evidenceMakes training measurableManager Restricted-data rulesProtects customers and companyPrivacy or security owner A current date matters. Last quarter’s positioning may now be wrong. A roadmap slide is not a product capability. One memorable customer objection is not automatically a market trend. AI will flatten these distinctions unless your inputs and instructions preserve them. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI sales enablement prompts Replace brackets with sanitized, approved information. Treat every output as a draft that an accountable human must review. 1. Audit enablement source material “Review these approved source documents: [paste or summarize with labels and dates]. Build an evidence register with columns for statement, classification, source, date, owner, intended audience, approval status, expiry date, conflict, and action. Classify each item as approved fact, observed customer language, interpretation, draft, hypothesis, unknown, or restricted. Flag stale, contradictory, unsupported, and duplicated material. Do not resolve conflicts by guessing.” Run this before asking for a new deck. It tells you whether the raw material can support one. Have source owners inspect every flag. If a product page says a feature is available and the release note says beta, the model should preserve the conflict. Product decides what is true. The machine does not settle the matter by choosing the sentence with stronger punctuation. Use the register as a maintenance tool. Add review dates and owners. Archive superseded claims instead of letting six versions circulate through shared folders like cursed treasure maps. 2. Build a grounded messaging guide “Using only this approved evidence register: [paste], draft a messaging guide for [role, segment, use case, and region]. Include audience problem, discovery questions, approved value statement, capability explanation, proof available, limitations, prohibited claims, specialist escalation points, and source citations. Put any proposed wording in a DRAFT column. Mark missing support [VERIFY]. Do not invent benefits, differentiation, results, or customer intent.” A messaging guide is a decision aid, not a script sellers must recite while their souls leave their bodies. It should help them connect verified capabilities to a problem the buyer has actually confirmed. Keep limitations visible. If an integration requires configuration, say so. If a result came from one approved case study, do not transform it into a universal benchmark. Review the guide with product marketing, product, legal, and frontline managers before release. Then test whether sellers can explain the idea in their own accurate language. Memorizing a paragraph proves memory. Explaining the boundaries proves understanding. 3. Create a discovery-question bank “From these approved use cases, customer patterns, and process criteria: [paste], draft a discovery-question bank for [seller role and conversation stage]. Group questions by current state, desired outcome, impact, constraints, stakeholders, decision process, risk, and next step. For each question, explain its purpose and list unsafe assumptions to avoid. Do not imply a problem exists before the buyer confirms it. Do not ask for sensitive information that is unnecessary.” Questions should create understanding, not steer a buyer toward your preferred answer. “How does your team handle this today?” is discovery. “How much money are you losing because your current system is terrible?” is a leading question wearing a tiny consultant hat. Managers should trim the bank. Nobody needs forty-seven questions on one call. Select a few based on the conversation, listen to the answers, and follow up naturally. For the wider workflow, pair this with a sales call coaching process that evaluates listening and evidence rather than keyword performance. 4. Draft objection-handling practice “Using these de-identified objection examples and approved responses: [paste], create a practice set for [role and experience level]. For each scenario, include context, buyer statement, clarifying questions, relevant approved evidence, response boundaries, escalation triggers, weak-response examples, and a scoring rubric. Label invented scenario details as fictional. Do not create guarantees, attack competitors, infer motives, or advise pressure tactics.” Objection practice should teach diagnosis. “Too expensive” may mean no budget, unclear value, a different priority, procurement friction, or a polite exit. The seller needs to ask, not launch into a canned monologue about return on investment. Use fictional or properly de-identified scenarios. Do not paste a private call transcript into an unapproved model because it would make the exercise feel authentic. Customer trust is not training data you found under the sofa. After role-play, the manager should coach one or two observable behaviors. The model can suggest feedback categories. It cannot read the seller’s judgment, trust, timing, or emotional control as well as an attentive human coach. 5. Turn product updates into seller briefs “Convert these approved release notes and product-owner comments: [paste] into a one-page seller brief. Include what changed, availability, eligible users, customer problem addressed, setup or dependency, known limitation, approved demo path, claims sellers may make, claims they must not make, likely questions, and escalation owner. Cite every factual statement. Keep roadmap items separate and clearly labeled. Mark uncertainty [VERIFY].” Release notes are written for accuracy, not always for field use. AI can help translate them, but translation must not inflate “supports” into “automates everything” or “planned” into “available Tuesday.” Ask the product owner to verify the brief against the live product. Ask legal or security to review claims in their domains. Add a version and expiry date. When the capability changes, retire the old brief everywhere sellers can find it. A useful brief also says who should not use the feature yet. Qualification boundaries save more trust than another adjective in the value statement. 6. Build realistic role-play scenarios “Create [number] fictional role-play scenarios from these approved patterns: [paste]. Vary role, business context, maturity, priorities, constraints, and conversation stage without using real customer identities. For each scenario, provide seller objective, buyer brief, verified facts, hidden information the seller must discover, plausible objections, prohibited assumptions, escalation moments, and observable success criteria. Do not stereotype industries, roles, cultures, or protected groups.” Good role-play has enough context to create choices, not so much that the exercise becomes theater with one approved ending. The seller should need to listen, ask, summarize, and decide whether the product fits. Keep the fictional label visible. Otherwise a scenario can migrate into a deck and later be repeated as a real customer story. That sounds absurd until someone says, “I think a healthcare client achieved this,” and nobody remembers the client was generated on a Tuesday afternoon. Rotate scenarios and difficulty. New sellers may practice accurate explanation and escalation. Experienced sellers may practice ambiguity, conflicting stakeholders, or respectfully disqualifying a poor fit. 7. Create a human-led call-review rubric “Using this approved selling framework and role expectations: [paste], draft a call-review rubric. Include observable criteria for preparation, agenda alignment, discovery, listening, evidence use, accurate product language, handling uncertainty, next-step clarity, and privacy. Define examples for developing, competent, and strong performance. Exclude personality judgments, accent, charisma, demographic inference, emotion detection, and unsupported intent scoring. Identify which criteria require human judgment.” A rubric should improve coaching consistency without pretending conversation quality is fully measurable. Talk-time ratios and keyword counts can be clues. They are not verdicts. Managers must review context. A short customer call may legitimately contain little discovery. A technical specialist may speak more because the buyer asked detailed questions. Automatic scores can create perverse behavior, such as asking robotic questions only to satisfy a dashboard. Use the rubric to structure a human conversation: what happened, what evidence supports that view, what the seller noticed, and what behavior to practice next. 8. Tailor training by role and experience “Adapt this approved training material: [paste] for [role, tenure, region, selling motion, and learning objective]. Preserve all claims, limitations, compliance language, and source citations. Propose role-relevant examples, practice activities, a short knowledge check, and an on-the-job application task. Mark adaptations requiring owner approval. Do not simplify away material limitations or create region-specific legal guidance.” Different roles need different depth. A business development representative, account executive, solution consultant, and customer success manager may touch the same product story from different angles. Adaptation should reduce irrelevant material, not remove inconvenient truth. A new seller still needs to know when to escalate. An experienced seller still needs current product boundaries. Measure application after the module. Ask a manager to observe a call, review a draft, or run a role-play. A perfect quiz score may prove that the seller can identify answer C. Revenue work rarely arrives with answer C highlighted. 9. Find gaps in the enablement library “Compare this enablement inventory, approved product map, sales-process requirements, objection themes, and manager feedback: [paste]. Produce a gap analysis. Separate missing content, stale content, discoverability problems, skill gaps, process gaps, and issues requiring product or leadership decisions. For each gap, cite evidence, affected audience, business risk, proposed owner, smallest useful intervention, and validation method. Do not assume a new asset is the answer.” This prevents the standard response to every field problem: create another PDF. Sometimes the material exists but sellers cannot find it. Sometimes managers are not reinforcing it. Sometimes the process conflicts with the training. Sometimes the product has a real limitation that no enablement asset can wordsmith into submission. Prioritize by seller and customer impact. Fix authoritative sources and retrieval before generating derivatives. For repeatable documentation, these AI SOP prompts can help preserve ownership and review steps. 10. Run the approval and release checklist “Review this draft enablement artifact against the supplied source register, audience, privacy rules, brand guidance, accessibility checklist, and approval matrix: [paste]. Check every claim and link, identify unsupported wording, confirm limitations remain visible, list restricted information, verify version and expiry date, and produce a sign-off table for content owner, product, legal, privacy, security, commercial, accessibility, and frontline manager as applicable. Do not mark any approval complete unless I provide evidence.” This is the final gate, not a ceremonial spell. The model can find inconsistencies and prepare the checklist. Named humans approve their domains. Also test the artifact with a small seller group. Can they find the answer? Can they explain it accurately? Do they know what remains unknown and when to escalate? Run an AI QA checklist before broad release. Track the version after launch. Collect specific feedback and observed behavior, then update the authoritative source. Do not let every manager maintain a private remix until the company has twelve competing definitions of the same feature. A practical operating rhythm AI-assisted enablement works best inside a boring, visible loop: Collect: gather approved, dated sources and restricted-data rules. Classify: separate facts, observations, interpretations, drafts, hypotheses, and unknowns. Draft: use AI for structure, transformation, examples, and gap detection. Verify: trace every material claim to a current owner-approved source. Approve: route product, legal, privacy, security, commercial, and accessibility questions correctly. Practice: let sellers apply the material in realistic human-led scenarios. Observe: measure behavior and customer impact, not content downloads alone. Maintain: assign an owner, version, expiry date, and retirement path. Start with one high-friction artifact. An objection practice set, release brief, or discovery bank is easier to govern than an attempt to regenerate the entire enablement library overnight. Speed is useful only when the direction is sane. If your team is new to these tools, read the no-BS guide to using AI at work and the quick explanation of what AI can and cannot do. The central lesson is gloriously unsexy: give the model controlled evidence, demand labels, and keep accountable humans in the loop. Frequently asked questions What are AI sales enablement prompts? They are structured instructions that help an AI system transform approved sales, product, and process information into draft training, messaging, practice, coaching, or reference material. Good prompts define the audience, source boundaries, output, prohibited assumptions, citations, and human reviewers. They do not ask the model to invent strategy from an empty box. Can ChatGPT create sales training materials? It can draft and adapt materials from approved inputs, create fictional practice scenarios, propose knowledge checks, and organize a curriculum. A human owner must verify product claims, examples, limitations, compliance language, accessibility, and instructional usefulness. Training completion should be followed by manager-observed application. Can AI write battlecards and objection responses? It can produce drafts when supplied with current, approved evidence. Humans must verify competitive statements, customer proof, pricing, legal boundaries, and tone. Competitive rumors and seller anecdotes should never be promoted to facts merely because the model formatted them neatly. Should we upload sales calls to an AI tool? Only when the tool and workflow are explicitly approved for that data. Call recordings may contain personal, confidential, contractual, or regulated information. Follow consent, access, retention, security, privacy, and legal requirements. Prefer minimum necessary data and de-identified excerpts for general training. How do we stop AI from inventing product claims? Provide a controlled source set, require citations for every claim, label unsupported content [VERIFY], forbid roadmap and performance invention, and route open questions to named owners. Then perform human review against the actual source. No prompt can guarantee truth when the source material is missing, stale, or contradictory. Can AI replace a sales enablement manager? No. It can accelerate drafting, repackaging, practice creation, and administrative analysis. It cannot own product truth, read organizational context reliably, build trust with managers and sellers, make accountable tradeoffs, or coach complex human behavior by itself. How should we measure AI-assisted enablement? Measure whether sellers can find accurate information, explain capabilities and limitations, ask better questions, handle uncertainty, follow process, and improve specific observed behaviors. Also track correction rates, stale-content incidents, approval time, and customer-impacting errors. Asset count and course completion are weak proxies. What is the safest first project? Choose one bounded artifact with clear sources and owners, such as a product-update brief or a de-identified objection-practice set. Define approval gates, test it with a small group, observe use, and improve the workflow before expanding. Do not begin by pouring the company drive into a consumer chatbot and calling it transformation. Use the machine for speed, not authority AI is good at turning known material into new shapes. That makes it useful for enablement, where the same approved knowledge may need to become a guide, exercise, brief, checklist, and coaching aid. But format conversion is not truth. A persuasive sentence can still be unsupported. A realistic scenario can still teach the wrong behavior. A polished battlecard can still be stale. Keep sources visible. Keep unknowns labeled. Keep specialist approvals real. Keep managers coaching humans rather than worshipping scores. The machine can help build the practice field; it should not referee the game, rewrite the rules, and announce the winner. For a broader field guide to staying useful while AI makes knowledge work faster and weirder, read Don’t Replace Me by Dmitry Kargaev. It is less interested in magic prompts than in the human judgment that makes any prompt worth using. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Account Planning Prompts: Build a Better Plan 10 AI account planning prompts to organize buyer evidence, map stakeholders, expose risks, test strategy, and plan next actions without inventing intel. AI promptsaccount planningsales strategy AI Customer Renewal Prompts: Prepare With Evidence 10 AI customer renewal prompts to verify outcomes, expose account risks, prepare options, and protect trust without inventing value or customer sentiment. AI promptscustomer renewalcustomer success AI Deal Desk Prompts: Review Complex Deals 10 AI deal desk prompts for checking discounts, terms, risks, approvals, and complex sales deals without inventing policy, evidence, or authority. AI promptsdeal deskrevenue operations --- ## AI Sales Forecast Prompts: Skip the Fake Certainty URL: https://dontreplace.me/articles/ai-sales-forecast-prompts-build-without-fake-certainty AI promptssales forecastingrevenue operationssales managementforecast planning AI Sales Forecast Prompts: Skip the Fake Certainty 10 AI sales forecast prompts to test evidence, build scenarios, challenge assumptions, protect buyer data, and produce an accountable forecast. August 31, 2026 Updated August 31, 2026 Sales forecasts have a beloved corporate ritual: take an uncertain pipeline, multiply it by suspiciously tidy percentages, and present the result with two decimal places. Congratulations. The uncertainty is now wearing a tie. AI sales forecast prompts can clean inputs, separate evidence from optimism, compare scenarios, and prepare useful challenge questions. They cannot know what a buyer will do, rescue stale CRM data, or make a forecast reliable by adding a confident paragraph. AI can structure forecast evidence. Humans still own the data, customer context, commercial judgment, privacy, and final number. These ten templates help sales leaders, revenue operations teams, founders, and account executives use AI for sales forecasting without appointing autocomplete chief prophecy officer. What a sales forecast actually is A forecast is an accountable judgment about likely revenue in a defined period. It is built from current evidence, explicit assumptions, historical context, and known uncertainty. It is not the same as pipeline. Keep these terms separate: Pipeline: opportunities that might produce revenue. Forecast: the portion and timing an accountable human expects to realize. Forecast category: a documented judgment such as pipeline, best case, commit, or closed. Scenario: a conditional outcome based on stated assumptions. Stage probability: often a CRM default, not evidence about one specific buyer. Buyer evidence: observable customer actions, decisions, approvals, and direct statements. Seller belief: useful context that must remain labeled as interpretation. A deal can be late-stage and weakly evidenced. A large pipeline can be concentrated in two fragile deals. A coverage ratio can describe exposure without predicting attainment. A close date can be the buyer's documented target—or the final day of the quarter entered because the form demanded something. Start with an evidence-first pipeline review before forecasting. If the underlying records are stale, forecasting simply laminates the mess. The evidence-first sales forecast formula Add this instruction to any prompt below: “Act as a sales forecast analysis assistant. Use only the sanitized evidence I provide for [team, segment, period, and currency]. Our stage rules are [rules], forecast categories are [definitions], and the decision is [decision]. Produce [artifact]. Separate verified buyer evidence, seller activity, seller interpretation, calculated measures, assumptions, scenarios, and unknowns. For every material conclusion, cite the source and last-updated date, show contradictory evidence, state confidence and limitations, name a human owner, and propose a validation action. Do not invent buyer intent, commitments, quotes, dates, probabilities, competitor facts, or sensitive attributes. Flag stale data, pushed dates, changed definitions, concentration, small samples, and conclusions requiring sales leadership, finance, privacy, legal, or customer review.” The important phrase is use only the evidence provided. If a required fact is missing, the model should return “unknown,” not write a tiny screenplay about the buyer's procurement process. Never paste buyer names, emails, call recordings, contracts, pricing exceptions, payment details, credentials, confidential strategy, private employee notes, or legally sensitive information into an unapproved tool. Use approved systems, minimum necessary fields, de-identification, aggregation, access controls, retention limits, consent, and human review. What to collect before prompting Create a sanitized forecast pack and keep source records in approved systems. InputWhy it mattersHuman check Forecast period and currencyDefines what the number meansFinance confirms Category definitionsPrevents “commit” meaning five thingsSales leadership approves Stage and close-date historyExposes aging and date pushingRevenue operations verifies Buyer-confirmed milestonesSeparates progress from activityAccount owner validates Deal amount and commercial statusSupports timing and valueFinance reviews Historical conversion contextAdds a baseline, not a verdictAnalyst validates Concentration by segment and dealReveals portfolio fragilityLeader reviews Assumptions and dependenciesMakes scenarios inspectableOwners confirm Missingness and data freshnessLimits false confidenceCRM owner audits Prior forecast and actualsExplains movement and biasFinance reconciles Do not silently impute missing facts. An explicit hole can trigger action. A plausible invented value can trigger a bad hiring plan. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI sales forecast prompts Replace brackets with sanitized, approved evidence. Treat every output as a draft for human review. 1. Clean and define forecast inputs “Turn this sanitized data dictionary, forecast period, currency rules, category definitions, source inventory, refresh dates, and known limitations into a forecast input brief: [paste]. Show unit of analysis, required fields, missingness, duplicate handling, amount logic, close-date logic, historical snapshot availability, source owner, and privacy classification. Flag inconsistent currencies, overwritten history, reopened deals, splits, renewals mixed with new business, and fields that should be removed or aggregated.” Use this before requesting a number. If one report uses annual contract value and another uses total contract value, the model can produce a beautiful reconciliation of two different realities. Revenue operations and finance should verify joins, exchange rates, owner changes, opportunity splits, exclusions, and historical snapshots. Today's CRM state does not automatically reveal what the team knew at the previous forecast. 2. Separate buyer evidence from seller belief “Classify these sanitized deal notes and forecast claims: [paste]. Label each as verified buyer action, direct buyer statement, seller activity, seller interpretation, calculated measure, assumption, unsupported claim, sensitive inference, or unknown. Preserve source and date. Rewrite unsupported claims as neutral validation questions. Do not infer urgency, authority, emotion, budget, protected characteristics, or purchase probability.” “The buyer scheduled legal review for Tuesday” is evidence. “Legal should be easy” is belief. Both may matter, but merging them turns confidence into camouflage. This classification is especially useful before executive reporting. Every summary pass tends to shave off caveats until “champion expects approval” becomes “approval expected.” 3. Reconcile forecast categories “Compare these sanitized opportunity assignments with our approved forecast category definitions: [paste]. For each deal, show criteria met, criteria contradicted, missing evidence, date sensitivity, external dependencies, and questions for the accountable forecaster. Recommend review priority, but do not assign a final category unless an explicit human-approved rule permits it.” A category is useful only when it has operational meaning. If “commit” means signed paperwork for one manager and enthusiastic vibes for another, aggregation creates theater. The model can expose inconsistent application. A sales leader must make or approve the final call because customer nuance, portfolio risk, and organizational consequences belong to humans. 4. Build base, upside, and downside scenarios “Using these sanitized opportunities, verified milestones, dependencies, category rules, and human-approved assumptions, build base, upside, and downside forecast scenarios: [paste]. For each scenario, list included deals or aggregates, explicit conditions, amount range, timing risk, concentration, evidence, contradictory evidence, and trigger that would move the scenario. Do not assign invented probabilities or collapse ranges into a single precise prediction.” Scenarios are conditional stories, not alternate universes selected by a dropdown. The base case should not simply be “what leadership wants,” and downside should not mean “subtract ten percent because that feels prudent.” Require a trigger for movement. If legal approval slips past a stated date, what changes? If a large deal signs, what remains exposed? Conditions make the forecast manageable. 5. Test close-date assumptions “Review this de-identified close-date history, buyer milestone evidence, procurement steps, implementation dependencies, and seller notes: [paste]. For each material deal, separate buyer-confirmed timing from internal target dates. Flag repeated pushes, missing approval steps, impossible sequencing, expired milestones, and quarter-end clustering. Provide a competing explanation and one customer-safe validation question for each flag.” A close date is not true because it survived three forecast meetings. Repeated pushes may signal weak qualification, an external dependency, a changed customer priority, or lazy CRM maintenance. The output should identify the uncertainty, not prosecute the rep. Use customer-safe questions. Forecast pressure is not permission to bully a buyer into inventing certainty they do not have. 6. Check pipeline coverage without pretending it predicts attainment “Analyze these sanitized aggregate pipeline values and historical context by segment, source, stage, owner group, and period: [paste]. Show unweighted value, approved weighted value, coverage ratio, age, missingness, definition changes, comparable historical ranges, and sensitivity to assumptions. Explain what each measure can and cannot support. Do not claim a coverage ratio guarantees attainment.” Coverage is a planning signal. It can reveal that the team has little room for slippage. It cannot tell you whether the actual deals are healthy. Historical comparison also needs comparable conditions. Product changes, territory redesigns, pricing shifts, seasonality, and different stage rules can break a neat trend. These AI data analysis prompts help structure checks, but humans must validate every calculation. 7. Identify concentration and dependency risk “Review these sanitized forecast aggregates and dependency labels: [paste]. Show concentration by deal, account group, segment, product, geography, partner, approval path, and implementation dependency where permitted. Model the effect of the largest exposures slipping or shrinking using human-approved scenarios. Flag shared dependencies and small groups that should not be reported for privacy reasons.” Ten deals are not diversified if eight depend on the same partner certification. A forecast may also look healthy while one giant opportunity carries the quarter on its back like a stressed intern. Do not publish tiny rep-level groups or use concentration analysis as employee surveillance. The purpose is portfolio risk and action planning, not generating a leaderboard of blame. 8. Prepare a forecast challenge meeting “Using this sanitized forecast pack, draft a 45-minute challenge agenda: [paste]. Prioritize the decisions and assumptions with the largest impact. For each, provide supporting evidence, contradictory evidence, unknowns, owner, neutral questions, required decision, and follow-up artifact. Separate data repair, deal validation, coaching, finance reconciliation, and leadership judgment. Avoid accusatory language and autogenerated interrogation.” A useful meeting resolves uncertainty or assigns validation. It does not read every row aloud while executives ask why the dashboard color changed. Pick the few questions that can materially alter timing, value, or action. Managers still need territory, tenure, product, and customer context. Sparse notes should never become an invisible performance system. 9. Explain changes from the prior forecast “Reconcile this current sanitized forecast with the prior snapshot and actual outcomes: [paste]. Build a change bridge for new deals, wins, losses, slips, amount changes, category changes, scope changes, currency effects, corrections, and definition changes. For every material movement, cite the source and owner. Separate real business movement from data cleanup. Do not invent causes when only a change is observable.” A number moving does not explain why it moved. “Forecast decreased by $200,000” is arithmetic. Cause requires evidence. This prompt can draft an executive narrative after finance and sales operations validate the bridge. Use an executive summary prompt to compress verified findings, not to erase uncertainty. 10. Turn the forecast into an accountable action plan “Convert these human-reviewed forecast findings into an action plan: [paste]. For each item include the verified issue, affected scenario, source, confidence, competing explanation, owner, due date, customer-safe validation step, success measure, dependency, guardrail, escalation condition, and review date. Separate CRM repairs, customer actions, coaching, commercial approvals, finance decisions, and final forecast calls.” “Close more deals” is not an action. “Account owner confirms the buyer's procurement sequence by Thursday; legal reviews two stated exceptions by Friday; forecast owner revisits scenario placement after both sources are recorded” is inspectable. Run an AI risk assessment before changing discounts, incentives, headcount, or rep evaluation based on forecast output. High-impact choices need complete context and named humans. Common forecasting failures Treating CRM probability as truth Default stage percentages may be workflow conventions. Trustworthy probability requires defined outcomes, suitable historical data, validation, calibration, monitoring, and stable processes. A language model inventing 73% is not a statistical upgrade. Hiding uncertainty in one number A single number is convenient but incomplete. Preserve ranges, dependencies, assumptions, and scenario triggers. Precision should reflect evidence, not formatting options. Confusing seller activity with buyer progress More calls and emails can mean engagement—or repeated contact with silence. Ask what the buyer confirmed, approved, scheduled, shared, or completed. Ignoring forecast process changes A new stage definition, territory model, pricing policy, CRM cleanup, or acquisition can break historical comparability. Label structural changes instead of calling every difference a trend. Automating employee judgment Forecast records reflect territory, product fit, inherited pipeline, manager habits, documentation quality, and customer complexity. Do not infer effort, honesty, or competence from partial records. Employment decisions require fair process and accountable people. Leaking customer information Removing a company name may not de-identify a giant deal in a small region. Combine deal size, timing, product, and quoted language carefully. Use approved tools and minimum necessary data. A practical forecast review checklist Before approving an AI-assisted forecast, confirm: The period, currency, amount basis, and category definitions are explicit. Every material claim has a source and update date. Buyer evidence is separated from seller interpretation. Close dates have customer evidence or remain labeled uncertain. Missing and contradictory records stay visible. Scenarios show assumptions, dependencies, and movement triggers. Coverage and CRM probabilities are not presented as guarantees. Concentration and small-sample risks are disclosed. Buyer and employee data was handled under approved rules. Finance, revenue operations, and the accountable leader reviewed the result. Every action has an owner, deadline, guardrail, and validation step. The final number belongs to a named human, not “the AI.” If several checks fail, do not request a more persuasive deck. Repair the evidence and process. Frequently asked questions Can ChatGPT create a complete sales forecast? It can organize sanitized inputs, apply documented rules, compare scenarios, and draft challenge questions. Humans must verify data, understand customers, reconcile finance, make the forecast call, and own the consequences. What data should I use with AI sales forecast prompts? Use the minimum approved evidence needed: forecast definitions, de-identified opportunity history, buyer-confirmed milestones, sanitized commercial status, prior snapshots, aggregate historical context, assumptions, and limitations. Keep source records in approved systems. Can AI predict whether a deal will close? A generic language model cannot reliably infer buyer intent from sales notes. It may apply a separate validated model supplied by humans, but any prediction still needs calibration, monitoring, appropriate data, and accountable review. Should I ask AI for a probability on every opportunity? No. A number generated from prose is not automatically meaningful. Use probabilities only when the method, data, validation, and limits are documented. Otherwise use evidence labels, scenarios, ranges, and validation actions. How often should a forecast be updated? Match cadence to sales-cycle speed, data freshness, and business decisions. Weekly reviews are common, with more frequent validation near deadlines. Constant refreshes do not create new customer evidence. Can AI move deals between forecast categories automatically? It can compare evidence with approved definitions and flag inconsistencies. An accountable human should approve category changes, especially when commercial context or business consequences matter. Is AI suitable for grading sales reps from forecast accuracy? Not as an automated performance system. Results depend on territory, deal mix, inherited pipeline, process changes, customer behavior, and manager practices. Performance decisions require complete context, fair policy, and human accountability. How do I protect confidential forecast data? Use approved tools, remove unnecessary identifiers, aggregate where possible, restrict access, set retention limits, document consent, and involve privacy or legal reviewers when required. Never casually paste contracts, recordings, credentials, or private notes into consumer tools. The useful boundary AI is fast at reorganizing evidence and exposing candidate inconsistencies. It is not the buyer, sales leader, finance partner, or person accountable when an inflated forecast changes hiring and spending. Use these prompts to make assumptions visible, scenarios inspectable, and validation easier. Then do the human work: talk to customers, repair the CRM, challenge the story, protect the data, and own the call. That boundary is the point of what AI can and can't do and the broader no-BS guide to using AI at work. For more practical rules about keeping judgment human, Don't Replace Me by Dmitry Kargaev is the field guide—not a permission slip to outsource your brain. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Sales Pipeline Review Prompts: Find Risk Early 10 AI sales pipeline review prompts to inspect deal evidence, challenge forecasts, protect buyer data, and turn pipeline risk into accountable action. AI promptssales pipelinesales forecasting AI Sales Territory Planning Prompts: Divide Without Drama 10 AI sales territory planning prompts to compare coverage scenarios, expose bad assumptions, protect sensitive data, and make accountable decisions. AI promptssales territory planningrevenue operations AI Customer Renewal Prompts: Prepare With Evidence 10 AI customer renewal prompts to verify outcomes, expose account risks, prepare options, and protect trust without inventing value or customer sentiment. AI promptscustomer renewalcustomer success --- ## AI Sales Negotiation Prompts: Prepare Better URL: https://dontreplace.me/articles/ai-sales-negotiation-prompts-prepare-without-manipulation AI promptssales negotiationB2B salesrevenue operationsdeal strategy AI Sales Negotiation Prompts: Prepare Better 10 AI sales negotiation prompts to clarify evidence, rehearse objections, plan value trades, and protect buyer trust without inventing leverage or authority. September 07, 2026 Updated September 07, 2026 Sales negotiation gets weird when preparation turns into theater. A seller has three stale CRM notes, a discount request, and a deadline nobody can explain. Then an AI tool produces a twelve-page “strategy” featuring invented buyer motives and the confidence of a hostage negotiator from television. AI sales negotiation prompts are useful for organizing verified information, finding gaps, rehearsing objections, and comparing approved options. They cannot read a buyer’s mind, authorize a concession, interpret a contract, manufacture leverage, or decide what preserving the relationship is worth. AI can help you prepare the conversation. It cannot have the conversation—or own the promise made in it. These ten templates help account executives, sales managers, founders, revenue operations teams, and commercial leaders prepare without manipulating buyers or outsourcing judgment to autocomplete. What good negotiation preparation actually does Negotiation is not a collection of tricks for defeating the person across the table. In durable B2B deals, it is a controlled exchange: both sides clarify interests, constraints, risks, and terms until they find an agreement they can actually keep. Useful preparation answers: What has the buyer explicitly requested? Which interests are verified, and which are assumptions? What can we approve, trade, or decline? Which concessions require finance, legal, security, or executive review? What questions would reveal the real constraint? What happens if no agreement is reached? Who has authority to commit each side? What must be documented after the call? Keep these categories separate: Verified fact: supported by a current, dated source. Buyer statement: something the buyer said, not proof of motive or authority. Seller interpretation: a hypothesis that needs testing. Approved boundary: a limit confirmed by the authorized owner. Possible trade: an option, not permission to offer it. Unknown: information that needs a question or specialist review. Decision: a choice made by a named human with authority. A model tends to blur these categories because smooth prose rewards certainty. Your process should reward visible uncertainty instead. If the deal needs formal exception review, start with AI deal desk prompts. For broader relationship context, use AI account planning prompts. Negotiation preparation should build on those records, not create a competing version of reality. The evidence-first sales negotiation prompt formula Add this instruction to every template below: “Act as a sales negotiation preparation assistant. Use only the sanitized, approved sources I provide for [deal, buyer role, negotiation stage, products, commercial boundaries, and decision]. Cite a source label and date for every price, term, deadline, commitment, capability, concession, and policy statement. Separate verified facts, buyer statements, seller interpretations, approved boundaries, possible trades, decisions, and unknowns. Mark unsupported or stale claims [VERIFY]. Do not invent buyer motives, authority, urgency, leverage, competitor claims, approvals, precedent, legal interpretations, product capabilities, or outcomes.” That instruction removes some sparkle. Good. Sparkle is not evidence. Never paste customer names, personal contact details, contracts, credentials, payment information, confidential margins, private CRM notes, security findings, regulated data, or legally sensitive information into an unapproved AI tool. Use approved systems, minimum necessary data, de-identification, access controls, retention limits, and human legal, finance, security, privacy, and commercial review. What to collect before prompting Build a small source pack instead of dumping the company’s entire commercial nervous system into a chatbot. InputWhy it mattersHuman owner Buyer requests in writingGrounds what is actually being negotiatedAccount owner Discovery notes with datesSeparates current needs from old assumptionsAccount owner Approved proposal and scopeDefines the offer under discussionCommercial owner Price book and discount policyEstablishes valid boundariesFinance or revenue operations Approval matrixShows who can authorize each tradeRevenue operations Product capability sourcePrevents roadmap fan fictionProduct owner Delivery constraintsTests whether promises are feasibleDelivery owner Contract and security statusExposes specialist dependenciesLegal and security owners Alternatives and no-deal impactGrounds the walk-away discussionDeal sponsor Negotiation historyPrevents repeated or conflicting offersAccount owner Restricted-data rulesControls what may enter the toolPrivacy or security owner When sources conflict, preserve the conflict. Do not ask the model to choose the version that makes your quarter look nicer. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI sales negotiation prompts Replace brackets with sanitized, approved information. Treat every output as preparation until an authorized human verifies it. 1. Check the negotiation inputs “Review this sanitized source pack against these negotiation-readiness requirements: [paste]. Create a table with required input, supplied evidence, source, date, status, conflict, missing information, owner, and next action. Do not infer missing values. Mark unsupported claims [VERIFY].” Use this before developing tactics. It catches missing approval limits, contradictory pricing, vague buyer requests, and mystery deadlines while they are still fixable. A negotiation plan built without current boundaries is merely a creative-writing exercise with revenue consequences. The account owner should resolve buyer questions; finance, legal, product, and delivery owners should confirm their own constraints. 2. Separate facts from assumptions “Classify every material statement in these notes as verified fact, buyer statement, seller interpretation, approved boundary, possible trade, decision, or unknown. Include the source and date. For each interpretation, write a neutral question that could test it. Do not infer emotion, intent, budget, authority, or urgency.” “The buyer needs 20% off” may mean the buyer requested it, procurement routinely asks, the budget is fixed, or the seller panicked after one email. Those are different realities. This prompt makes hidden stories visible. Humans then test them through respectful questions instead of optimizing a strategy around imaginary psychology. 3. Map buyer interests and unknowns “Using only these verified notes: [paste], create an interest map with stated need, supporting evidence, operational impact, success measure, constraint, unresolved question, and person who can verify it. Keep inferred interests in a separate hypothesis section. Draft open, non-leading questions to validate each unknown.” Positions are requests: “We need monthly terms.” Interests explain why: cash flow, policy, implementation risk, or something else. You can explore interests without pretending you already know them. Do not use AI to profile a buyer from demographics, location, personality guesses, or unrelated personal data. Ask. Listen. Record the answer accurately. Revolutionary technology, apparently. 4. Define boundaries and approvals “Map the proposed negotiation items against this current approval matrix and policy set: [paste]. For each item, show standard position, approved range, prohibited commitment, approval owner, evidence, expiration date, and open question. Label brainstormed options as NOT APPROVED. Do not treat past deals as authority for this one.” This keeps “we could discuss” from becoming “sales promised.” A seller should know which dimensions are flexible before entering the room: price, volume, term, payment timing, scope, implementation, support, reference participation, or sequencing. Complex combinations matter. A discount that is acceptable on a three-year prepaid agreement may be unacceptable with monthly payments and custom implementation. Route exceptions through AI risk assessment prompts and the actual approval system. 5. Build a give-get matrix “Using only these approved options and boundaries: [paste], create a give-get matrix. For each possible give, list business cost, risk, required approval, minimum reciprocal get, evidence the get has value, and wording that presents the exchange clearly. Do not recommend unapproved concessions or manipulative pressure.” A concession without a reciprocal commitment is not negotiation. It is leakage. Possible trades might pair lower price with greater volume, longer term, faster payment, narrower scope, standardized implementation, or a verified reference agreement. Finance must validate costs. Legal must validate terms. Delivery must confirm feasibility. The model is arranging columns, not granting commercial authority. 6. Prepare discovery questions “Draft 12 concise, open questions for the next negotiation conversation using these verified facts and unknowns: [paste]. Group them by business outcome, constraint, decision process, authority, timeline, implementation, risk, and alternatives. Explain what evidence each question seeks. Avoid leading, intrusive, or coercive wording.” Good questions reduce the need for clever tactics. Ask what outcome matters, what prevents agreement, who must review the terms, and what a workable implementation looks like. Do not turn questions into an interrogation script. Choose the few that fit the moment, listen to the answer, and follow the conversation. Human attention remains annoyingly non-automatable. 7. Rehearse objections without inventing answers “Act as a skeptical buyer using only these documented concerns: [paste]. Run a rehearsal with one objection at a time. After each response, evaluate whether I acknowledged the concern, asked a useful question, cited verified evidence, stayed inside approved boundaries, and avoided unsupported claims. If information is missing, require me to say I will verify it.” Rehearsal is one of the safest uses of AI because the model is helping you practice rather than generating a buyer-facing commitment. Never let it invent a product capability, implementation date, competitor weakness, customer result, or legal answer to make the role-play easier. Practice saying, “I don’t know; I’ll confirm with the owner by Tuesday.” That sentence closes fewer imaginary deals and prevents more real disasters. For broader coaching structure, see AI sales call coaching prompts. 8. Compare approved scenarios “Compare these approved negotiation scenarios: [paste]. Create a table covering price, term, payment timing, volume, scope, implementation effort, margin input supplied by finance, contractual risk supplied by legal, buyer value, company value, dependencies, required approvals, and unresolved questions. Do not calculate missing figures or rank scenarios unless decision criteria are provided.” Scenario comparison helps people see the whole trade instead of obsessing over one discount percentage. It should expose uncertainty, not decorate it. Humans decide which trade-offs fit strategy and relationship context. If the decision is consequential, record the rationale with AI decision log prompts after the authorized owner decides. 9. Draft the negotiation recap “Draft a concise post-meeting recap from these sanitized notes: [paste]. Separate agreed items, proposals still under review, declined items, open questions, owners, deadlines, and next meeting. Quote exact numbers and terms only when supported by a source. Mark ambiguity [VERIFY]. Do not convert discussion into agreement.” Send recaps quickly, but verify them first. The dangerous phrase is “as agreed” when nobody agreed. Use plain language such as “you requested,” “we proposed,” and “both parties confirmed” only when each is accurate. A recap is also a trust test. It should help the buyer correct misunderstandings before they harden into contract drama. For difficult customer-facing wording, AI client communication prompts can help without hiding the substance. 10. Run the final human sign-off “Audit this negotiation plan and proposed recap against the supplied evidence, policies, approvals, and restricted-data rules: [paste]. Check every price, term, date, capability, concession, buyer statement, and commitment. List unsupported claims, expired sources, privacy issues, missing approvals, ambiguous language, and unresolved dependencies. Produce a human sign-off checklist by owner. Do not approve or send anything.” Run this before a material call and again before sending a final recap or proposal. AI QA checklist prompts offer a wider review pattern. The final sign-off belongs to people who can answer for the outcome. AI does not carry quota, professional liability, delivery responsibility, or the memory of a buyer who was misled. A practical workflow that does not become prompt theater Sanitize the source pack. Use the minimum approved information. Run the input check. Stop if critical evidence or boundaries are missing. Separate facts and assumptions. Turn assumptions into questions. Confirm human boundaries. Get approvals before presenting options. Build trades and scenarios. Compare total value and risk. Rehearse. Practice honest responses, including when to verify later. Hold the conversation. Listen instead of reading a generated script. Write and verify the recap. Preserve proposals, decisions, and unknowns. Update the system of record. Keep future preparation grounded. This is slower than asking, “How do I win this negotiation?” It is faster than repairing a damaged relationship, unauthorized discount, or impossible delivery promise. If you need the basic operating model first, read how to use AI at work without the nonsense and what AI can and cannot do. Frequently asked questions Can ChatGPT negotiate a sales deal for me? It can help organize evidence, generate questions, rehearse objections, and draft a recap. It should not represent your company, make offers, approve concessions, or communicate autonomously with a buyer. A named human owns the relationship and every commitment. What information should I remove before using AI? Remove names, emails, contracts, private CRM notes, credentials, payment details, confidential pricing and margins, security findings, regulated data, and legally sensitive information unless an approved system and policy explicitly allow them. De-identify the minimum necessary context. Can AI tell me a buyer’s real motive? No. It can label hypotheses and suggest neutral questions. It cannot infer motive reliably from tone, role, demographics, or sparse notes. Treat stated interests as statements and test interpretations through conversation. Should AI recommend my walk-away point? It may organize criteria supplied by authorized humans. Finance and commercial leaders must define the economics, risks, alternatives, and approval boundaries. The model does not know what the company should value or accept. Can I use AI to create negotiation leverage? Use it to identify verified options and unanswered questions, not to fabricate deadlines, competitor activity, scarcity, authority, or customer demand. Fake leverage is deception wearing a spreadsheet. How do I stop AI from inventing concessions? Provide an explicit approved-options list, require source citations, label all brainstormed options NOT APPROVED, prohibit inferred authority, and route every possible concession to its named approver before buyer-facing use. Are AI-generated negotiation recaps safe to send? Only after a human checks every number, term, commitment, owner, and deadline against notes and approved records. The draft should distinguish requests, proposals, agreements, and open questions. Never auto-send it. What is the best first sales negotiation prompt? Start with the input check. Finding missing evidence and unclear authority creates more value than generating tactics. Once the facts and boundaries are sound, the other prompts become useful instead of decorative. The point AI sales negotiation prompts should make your preparation more honest, structured, and inspectable. They should not make your behavior more synthetic. Use AI to expose gaps, organize choices, and rehearse difficult moments. Keep empathy, listening, judgment, authority, and accountability with the humans in the deal. That is the same broader survival principle behind Don’t Replace Me by Dmitry Kargaev: use the machine for speed, then bring the parts it cannot supply. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Customer Renewal Prompts: Prepare With Evidence 10 AI customer renewal prompts to verify outcomes, expose account risks, prepare options, and protect trust without inventing value or customer sentiment. AI promptscustomer renewalcustomer success AI Deal Desk Prompts: Review Complex Deals 10 AI deal desk prompts for checking discounts, terms, risks, approvals, and complex sales deals without inventing policy, evidence, or authority. AI promptsdeal deskrevenue operations AI Sales Enablement Prompts: Build Useful Materials 10 AI sales enablement prompts for grounded messaging, training, coaching, role-play, objection practice, and seller resources without invented claims. AI promptssales enablementsales training --- ## AI Sales Pipeline Review Prompts: Find Risk Early URL: https://dontreplace.me/articles/ai-sales-pipeline-review-prompts-find-risk-before-forecast AI promptssales pipelinesales forecastingrevenue operationssales management AI Sales Pipeline Review Prompts: Find Risk Early 10 AI sales pipeline review prompts to inspect deal evidence, challenge forecasts, protect buyer data, and turn pipeline risk into accountable action. August 30, 2026 Updated August 30, 2026 Sales pipeline reviews have a recurring magic trick: a deal with no buyer reply, no confirmed next step, and a close date pushed three times remains “90% committed.” Apparently the percentage is load-bearing. AI sales pipeline review prompts can organize deal evidence, flag missing fields, test stage criteria, and draft sharper coaching questions. They cannot read buyer intent, verify a rep’s interpretation, or turn stale CRM data into a trustworthy forecast through the healing power of bullet points. AI can make pipeline evidence easier to inspect. Humans still own CRM hygiene, customer context, coaching, commercial judgment, privacy, and the forecast. These ten templates help sales leaders, revenue operations teams, founders, and account executives review pipeline without appointing autocomplete chief revenue oracle. What a pipeline review should actually do A useful review tests whether recorded deal status matches observable buyer progress. It should expose uncertainty early enough for someone to act—not decorate a forecast call after the quarter is already cooked. Separate these concepts: Seller activity: emails sent, calls made, demos delivered. Buyer progress: stakeholders engaged, requirements confirmed, decisions scheduled, approvals completed. Stage: a label governed by documented entry and exit criteria. Forecast category: a human judgment about likely timing and outcome. Close date: an evidence-backed customer milestone, not the last day of the quarter wearing a fake moustache. Probability: a calibrated estimate, when one exists—not a default CRM percentage. Activity can support progress, but it is not progress. Twelve follow-ups with silence do not equal momentum. A proposal sent is not a proposal reviewed. “Verbal yes” is not procurement approval. A champion is not an economic buyer simply because everyone likes the champion. Pipeline review also differs from prediction. The job is to identify what is known, what is claimed, what is missing, what could change the outcome, and who will validate it. For broader deal learning, use AI win-loss analysis prompts. For general commercial workflows, these AI sales prompts cover discovery and follow-up without pretending a template closes the deal for you. The evidence-first pipeline review formula Add this instruction to any prompt below: “Act as a sales pipeline review assistant. Use only the sanitized evidence I provide for [team, segment, and review period]. Our documented stage criteria are [criteria], forecast definitions are [definitions], and the decision is [decision]. Produce [artifact]. Separate verified buyer actions, direct buyer statements, seller activities, seller interpretations, calculated measures, hypotheses, and unknowns. For every risk or recommendation, cite the source and last-updated date, show contradictory evidence, state confidence, name a human owner, and propose the next validation action. Do not invent buyer intent, quotes, dates, probabilities, commitments, competitor facts, or sensitive attributes. Flag stale records, missing data, stage inflation, close-date pushing, unfair rep comparisons, and conclusions requiring sales leadership, finance, privacy, legal, or customer review.” This prevents “proposal sent” from becoming “legal approval expected Friday” somewhere between input and executive summary. If the evidence does not support a claim, the output should say unknown and ask for a validation step. Never paste buyer names, emails, call recordings, contracts, pricing exceptions, payment details, credentials, confidential roadmaps, private rep notes, or legally sensitive information into an unapproved AI tool. Use approved systems, minimum necessary fields, de-identification, aggregation, access controls, retention limits, consent, and human review. What to collect before prompting Build a sanitized evidence pack. Keep source records in approved systems. InputWhy it mattersHuman check Stage definitions and exit criteriaTests whether labels match evidenceRevenue operations confirms Forecast category definitionsPrevents “commit” meaning five thingsSales leadership approves Stage and close-date historyShows aging and repeated pushesCRM owner verifies Buyer-confirmed next stepsDistinguishes progress from activityAccount owner validates Stakeholder mapExposes missing authority or functionsManager reviews Requirements and success criteriaConnects the deal to buyer needBuyer-facing owner confirms Commercial and procurement statusReveals approval dependenciesFinance or legal reviews Risks and competing explanationsPrevents one-story forecastingDeal team challenges Data freshness and missingnessLimits false confidenceRevenue operations audits Decision owner and review dateMakes action accountableLeader confirms Do not quietly fill blanks with assumptions. “Economic buyer unknown” is useful information. It is more useful than a synthetic stakeholder biography assembled from job titles and optimism. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI sales pipeline review prompts Replace brackets with sanitized facts. Treat every output as a draft for human review. 1. Prepare a clean pipeline review brief “Turn this sanitized CRM dictionary, stage logic, forecast definitions, review period, segment scope, source inventory, and known limitations into a pipeline review brief: [paste]. Include the unit of analysis, required fields, source owner, last refresh, missingness, privacy classification, and calculation logic. Flag duplicate opportunities, reopened deals, inconsistent currencies, changed definitions, copied-forward dates, and fields that should be removed or aggregated.” Run this before asking which deals are risky. Duplicate opportunities can inflate pipeline. Currency mistakes can make one region look heroic. Reopened deals can distort age and conversion unless handled explicitly. Have revenue operations verify joins, owner changes, splits, renewal versus new-business logic, and historical snapshots. A tidy brief based on today’s overwritten fields cannot reconstruct what the team knew last month. 2. Audit CRM completeness and freshness “Audit these sanitized pipeline records against our required evidence standard: [paste]. For each opportunity, show missing fields, stale fields, contradictory entries, unsupported assertions, and the responsible owner. Prioritize repairs by forecast impact and review deadline. Do not infer missing values. Distinguish optional documentation from evidence required by policy.” Completeness is not “every box contains text.” A field reading “TBD” may honestly communicate uncertainty; a copied paragraph can create the illusion of knowledge. Check timestamps and sources, not character count. Use the audit to repair the process, not to shame reps in public. Documentation quality may reflect territory, workload, manager habits, CRM design, or inherited records. Automated punishment encourages better-looking fiction. 3. Test stage-exit evidence “Compare each sanitized opportunity with these documented stage-exit criteria: [paste criteria and evidence]. Return met, partially met, not met, contradicted, and unknown for every criterion. Cite the exact evidence and date. Recommend whether a human should validate, advance, hold, or move the stage backward. Do not change records or invent buyer confirmation.” This prompt is useful when stages drift from definitions. A demo may satisfy a seller activity requirement while leaving buyer requirements, authority, timeline, or approval process unresolved. Moving a deal backward is not failure. It is cheaper than discovering during forecast week that “contracting” meant someone attached a draft agreement to an unanswered email. 4. Identify stale and repeatedly pushed deals “Review this de-identified stage history, activity history, buyer-response history, and close-date history: [paste]. Flag opportunities with no recent buyer action, repeated close-date pushes, unusually long stage age, reopened status, or activity without progress. Compare only with relevant segment and deal-type baselines. For each flag, provide evidence, a competing explanation, and one human validation question.” A long sales cycle may be normal for enterprise procurement and suspicious for a small self-serve expansion. Use comparable groups rather than one universal aging threshold. Repeated close-date pushes deserve investigation, not automatic deletion. The buyer may have a documented external dependency. Or the date may be quarter-end fan fiction. The next step is evidence. 5. Check the quality of next steps “Evaluate these sanitized next-step records: [paste]. Classify each as buyer-confirmed action, seller-owned action, mutual action, vague intention, expired action, or missing. Check for owner, date, expected artifact, dependency, and evidence of buyer agreement. Rewrite weak entries as questions the account owner must validate; do not manufacture a commitment.” “Follow up next week” is not a mutual action plan. “Buyer’s security lead will return the completed questionnaire by September 4; account owner will answer exceptions within two business days” is inspectable. A strong next step has an owner, timing, artifact, dependency, and source. It may still fail. The point is to distinguish a real plan from decorative CRM mulch. 6. Separate buyer evidence from seller assumptions “Classify these sanitized notes and deal claims: [paste]. Put each into verified buyer action, direct buyer statement, seller activity, seller interpretation, calculated measure, hypothesis, unsupported assumption, sensitive inference, or unknown. Preserve source and date. Rewrite assumptions as neutral validation questions. Do not infer emotion, authority, budget, urgency, protected characteristics, or purchase probability.” “The CFO attended the pricing call” is observable. “The CFO loves us” is interpretation. “Budget approved” requires evidence about the actual approval, amount, timing, and conditions. This is particularly valuable before executive reviews, where a caveated account note can become a confident slide title after two rounds of summarization. 7. Review pipeline concentration and coverage carefully “Analyze these sanitized aggregate pipeline measures by segment, owner, product, source, stage, and expected period: [paste]. Show total and unweighted values, concentration, age, missingness, and historical comparability. Identify dependence on a few large opportunities and scenarios that would materially change coverage. Do not claim coverage predicts attainment or treat CRM probabilities as calibrated without evidence.” A pipeline can appear healthy while depending on two giant opportunities sharing the same procurement blocker. Total value hides concentration. Weighted value can hide whatever assumptions are embedded in default probabilities. Have finance and revenue operations verify calculations. If groups are small, avoid publishing rep-level comparisons that expose individuals or imply precision the data cannot support. AI data analysis prompts can help structure the checks, but humans must validate every number. 8. Challenge forecast categories “Pressure-test these sanitized forecast assignments using our definitions: [paste]. For each opportunity, list supporting evidence, contradictory evidence, missing evidence, date sensitivity, external dependencies, and alternative outcomes. Recommend questions for the accountable forecaster. Do not assign a new category or probability unless a human-approved rule explicitly permits it.” “Commit” should mean something operational. If it means “the rep feels good,” “the manager wants coverage,” and “the buyer signed” depending on the deal, the category is a mood ring. Use this prompt to surface inconsistencies, then let the accountable leader make the call. Forecast judgment includes customer nuance, portfolio risk, and business consequences that a text model does not own. 9. Prepare coaching questions for a deal review “Using this sanitized deal evidence and the documented stage criteria, draft a coaching agenda: [paste]. Ask neutral questions about buyer problem, measurable outcome, decision process, authority, stakeholders, alternatives, commercial approval, implementation risk, next step, and evidence gaps. Separate questions that clarify facts from questions that test strategy. Avoid accusatory language, scripts that pressure buyers, or employee performance conclusions.” Good coaching improves thinking. It does not use AI to cross-examine a rep with fifty autogenerated gotchas. Pick the few questions most likely to change understanding or action. Managers should account for territory, tenure, product fit, inherited pipeline, and customer complexity. A model reviewing partial notes should never become an invisible performance-rating system. 10. Turn findings into an accountable action plan “Convert these human-reviewed pipeline findings into an action plan: [paste]. For each item include the verified finding, affected opportunity or segment, source, confidence, competing explanation, owner, due date, customer-safe validation step, success measure, dependency, risk, guardrail, escalation condition, and review date. Separate data repairs, customer actions, internal decisions, coaching tasks, and forecast decisions.” “Save the deal” is not an action. “Account owner confirms procurement sequence with the buyer’s named contact by Thursday and records the source; legal reviews the two open terms by Friday” is at least accountable. Run an AI risk assessment before changing pricing, incentives, approval policy, or rep evaluation. High-impact decisions need complete context and named humans, not an automated confidence adjective. Common pipeline-review failures Treating CRM fields as customer truth CRM records are operational artifacts. Fields can be stale, constrained by taxonomy, copied forward, or written from a seller perspective. Preserve the distinction between recorded claim and verified buyer evidence. Confusing activity with progress More emails may indicate engagement—or a seller repeatedly contacting an unresponsive account. Ask what the buyer did, confirmed, shared, approved, or scheduled. Inventing probabilities A language model can produce a number because you requested one. That does not make the number calibrated. Trustworthy probabilities require defined outcomes, historical data, stable processes, validation, monitoring, and appropriate statistical review. Automating rep surveillance Pipeline summaries are incomplete and context-dependent. Do not infer effort, honesty, competence, or intent from sparse CRM records. Coaching and performance decisions require fair process and accountable managers. Leaking customer and employee data De-identification is more than deleting a company name. Deal size, geography, product, date, and quoted language may identify an account or person. Use the minimum necessary information in approved tools. Producing a dashboard with no owner A red flag without a validation step becomes meeting decoration. Every material finding needs an owner, deadline, source, and decision path. A practical review checklist Before relying on an AI-assisted pipeline finding, confirm: Stage and forecast definitions are documented. Every material claim has a source and update date. Buyer progress is separated from seller activity. Close dates have customer evidence or are labeled uncertain. Missing and contradictory records remain visible. Opportunities are compared only with relevant peers. CRM probabilities are not presented as calibrated facts. Buyer and employee data was handled under approved rules. A human reviewed calculations, context, and recommendations. Every action has an owner, deadline, guardrail, and validation step. If several checks fail, do not request a more persuasive summary. Repair the evidence and operating process first. Frequently asked questions Can ChatGPT run a complete sales pipeline review? No. It can organize sanitized records, test documented criteria, highlight missing evidence, and draft questions. Humans must verify CRM data, understand customer context, coach the team, make forecast calls, and own consequences. What data should I use with pipeline review prompts? Use the minimum approved evidence needed: stage history, forecast definitions, de-identified buyer actions, sanitized notes, next steps, commercial status, known risks, and data limitations. Keep original records in approved systems. Can AI predict whether an opportunity will close? It can summarize patterns or apply a validated model supplied by humans. A generic text model cannot reliably infer purchase intent from notes, and a confident probability is not proof of calibration. Treat predictions as hypotheses unless rigorously validated. How often should pipeline be reviewed? Match review frequency to sales-cycle speed, decision needs, and data freshness. Many teams use weekly operating reviews and more frequent updates near key deadlines. Constant review does not help if no new evidence exists. Should stale deals be removed automatically? No. Staleness is a flag for human validation. Some legitimate buying processes pause. Use documented criteria, verify the customer situation, and apply consistent policy rather than letting a model delete records. Can AI assign forecast categories? It can compare evidence with human-approved definitions and identify mismatches. The accountable sales leader should make or approve category decisions, especially when customer nuance and business risk matter. Is AI appropriate for scoring salespeople from pipeline data? Not as an automated performance system. CRM records reflect territory, segment, product fit, manager practices, inherited pipeline, documentation habits, and customer complexity. Employment decisions require complete context, fair process, and human accountability. How do I keep confidential deal data safe? Use approved tools and policies, remove unnecessary identifiers, aggregate where possible, restrict access, set retention limits, document consent, and involve privacy or legal reviewers when required. Never paste credentials, contracts, recordings, or sensitive notes into consumer tools casually. The useful boundary AI is fast at restructuring records and finding candidate gaps. It is not the buyer, the sales manager, the finance partner, or the person accountable when an inflated forecast changes hiring, spending, or someone’s career. Use these prompts to make evidence easier to inspect and uncertainty harder to hide. Then do the human work: talk to customers, verify the CRM, coach with context, challenge assumptions, protect the data, and own the forecast. That boundary is the point of what AI can and can’t do and the broader no-BS guide to using AI at work. For more practical rules about keeping judgment and accountability human, Don’t Replace Me by Dmitry Kargaev is the field guide—not a permission slip to outsource your brain. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Sales Forecast Prompts: Skip the Fake Certainty 10 AI sales forecast prompts to test evidence, build scenarios, challenge assumptions, protect buyer data, and produce an accountable forecast. AI promptssales forecastingrevenue operations AI Sales Territory Planning Prompts: Divide Without Drama 10 AI sales territory planning prompts to compare coverage scenarios, expose bad assumptions, protect sensitive data, and make accountable decisions. AI promptssales territory planningrevenue operations AI Customer Renewal Prompts: Prepare With Evidence 10 AI customer renewal prompts to verify outcomes, expose account risks, prepare options, and protect trust without inventing value or customer sentiment. AI promptscustomer renewalcustomer success --- ## 10 AI Sales Prompts You Can Use Today URL: https://dontreplace.me/articles/ai-sales-prompts-discovery-follow-ups-objections ai sales promptssales ai toolschatgpt for salesai follow-up emailobjection handling ai 10 AI Sales Prompts You Can Use Today 10 copy-paste AI sales prompts for discovery calls, follow-ups, objection prep, CRM updates, and reactivation. Real templates, no fluff. June 01, 2026 Updated June 01, 2026 Sales reps spend more time on admin than selling. Logging CRM notes, drafting follow-ups, rewriting that passive-aggressive prospect email so it sounds less desperate. AI sales prompts can handle most of that grunt work, and they can do it fast. But here's the thing about sales: the words are the easy part. Knowing when to send them, what the prospect actually cares about, when to push and when to shut up. That's the job. AI can draft. You decide. This article gives you 10 copy-paste prompts for the real work of sales: discovery summaries, follow-ups, objection prep, CRM updates, and more. Use them as starting points, not finished output. Before you paste anything: don't include real customer names, confidential pricing, contract terms, security details, or regulated data in any AI tool unless your company's policy explicitly permits it. Anonymize first. Always. The reusable formula behind every good AI sales prompt Bad prompts produce bad drafts. The fix is simple: give the AI a role, context, task, constraints, and a format. This is what Dee calls the "smart intern" framework in Don't Replace Me. Treat AI like someone who's fast, tireless, knows a lot in general, and knows nothing specific about your customer or your deal. Here's the formula: Role: "You are a B2B sales rep." Context: What happened, who's involved, what stage you're at. Task: What you need the AI to produce. Constraints: What to avoid, what tone to use, what's off-limits. Format: Email, bullet list, short paragraph, whatever you need. Every prompt below follows this structure. You'll see it immediately. The most common mistake people make is pasting in a one-liner and expecting usable output. "Write me a follow-up email for a sales call" gets you something generic enough to be useless. Specificity is everything. The more context you give, the closer the first draft lands. You're not saving time by being vague. You're just moving the work from drafting to editing. 10 AI sales prompts you can use today Prompt 1: Summarize a discovery call When to use it: Right after a call, before your memory goes soft. You are a B2B sales rep. I just finished a 30-minute discovery call with a [job title] at a [industry] company with about [company size] employees. Here's what I remember: [paste your rough notes]. Summarize this into: (1) key pain points mentioned, (2) business goals they described, (3) objections or hesitations, (4) questions they asked, (5) agreed next steps. Use bullet points. Keep it under 200 words. What to check: Did you miss anything? Is anything in the summary a stretch of what was actually said? Don't let the AI invent clarity that wasn't in the call. Prompt 2: Extract buying criteria and decision factors When to use it: When you want to build a smarter pitch for the next meeting. You are a B2B sales strategist. Based on these discovery call notes: [paste anonymized notes], list the apparent buying criteria this prospect seems to care about, ranked by how strongly they emphasized each one. Then flag anything that suggests a potential blocker or political complexity. Be direct. What to check: These are inferences from your notes, not facts. Treat the output as a hypothesis to test in the next conversation. Prompt 3: Draft a useful follow-up email When to use it: After any call, meeting, or demo. The most-used prompt in this list. You are a B2B sales rep. Write a follow-up email to [role, not name] after a [type of meeting]. We discussed: [3-5 bullet points of key topics]. The next step we agreed on is [specific action and date]. The email should be under 150 words, warm but not sycophantic, and end with one clear ask. No filler sentences. What to check: Does the "one clear ask" match what you actually agreed? Is the tone right for this specific relationship? If you've only spoken once, "warm but not sycophantic" means something different than if you've been working together for months. For more on writing better work emails, see the AI email prompts guide. Prompt 4: Rewrite a pushy email so it sounds human When to use it: You wrote something under pressure and it reads like a ransom note. You are an editor specializing in sales communication. Here is an email draft I wrote: [paste draft]. Rewrite it so it sounds like a real person talking, not a sales sequence. Remove urgency language, soften anything that sounds like pressure, and cut any sentence that exists purely to fill space. Keep it under 120 words. What to check: Is the core message still there? Did the rewrite remove something important to the deal? Read it out loud. If you wouldn't say it in a call, don't send it in an email. Prompt 5: Prepare for common objections When to use it: Before a pitch meeting or negotiation call. You are a sales coach. My product is [one sentence description]. My prospect is a [role] at a [type of company]. Based on what I know about this buyer type, what are the five most likely objections they'll raise? For each objection, give me: (1) a likely underlying concern, and (2) a response that addresses the concern without dismissing it. Keep each response under 50 words. What to check: Do these responses match what your product actually does? Don't let AI write claims you can't stand behind. If the objection involves pricing, contract flexibility, or SLA commitments, verify with your manager before using any response. Prompt 6: Qualify an opportunity When to use it: When you're not sure if an opportunity is worth pursuing. You are a B2B sales strategist. Here is what I know about this opportunity: [paste your notes, anonymized]. Using the MEDDIC framework (Metrics, Economic Buyer, Decision Criteria, Decision Process, Identify Pain, Champion), flag what I know, what I'm missing, and what questions I should ask to fill the gaps. Format as a table. What to check: Is the MEDDIC read accurate, or are you being optimistic? The AI doesn't know your gut feel on this deal. You do. Use this as a sanity check, not a forecast. Prompt 7: Create a proposal outline When to use it: Before you start writing a formal proposal or presentation. You are a B2B sales rep preparing a proposal for a [type of company] with [describe their situation in 2-3 sentences]. The key outcomes they care about are: [list]. Create a proposal outline with: (1) executive summary framing, (2) section headers, (3) a suggested order that builds toward the ROI argument. No filler sections. What to check: Does the outline reflect what this prospect actually cares about, or is it generic? Personalize before you use it. Numbers, timelines, implementation specifics, pricing, and legal terms must all be verified before they appear in any real proposal. Prompt 8: Write a CRM update When to use it: End of day, before you forget everything. You are a B2B sales rep. Write a CRM activity note for a call I had today. Here's what happened: [paste rough notes]. Format it as: date, contact role, meeting type, key topics discussed, next action, and owner. Factual and concise. Under 100 words. What to check: Is everything accurate? CRM notes have a way of becoming gospel for colleagues who weren't on the call. Don't let a wrong inference from AI create a misunderstanding down the line. For handling meeting notes more broadly, the AI meeting notes prompts article has more templates. Prompt 9: Prepare a handoff to customer success When to use it: When a deal closes and you need to brief the CS or onboarding team. You are a B2B sales rep writing a handoff document for customer success. The customer is a [type of company, anonymized]. Here's what I know about them: [paste notes]. Write a handoff document with: (1) context on why they bought, (2) what success looks like to them, (3) known risks or sensitivities, (4) relationships to be aware of, (5) agreed commitments made during the sales process. Be specific and honest about risks. What to check: Is the commitments section accurate? If you promised something outside standard terms, that needs to be in there, clearly marked. Don't let the AI smooth over the awkward bits. Prompt 10: Plan a no-pressure reactivation message When to use it: A deal went quiet 3-6 months ago and you want to try again without being annoying about it. You are a B2B sales rep. I had a deal with a [role] at a [type of company] that went quiet [X months] ago. The last thing that happened was: [one sentence]. Write a short reactivation message (under 100 words) that: acknowledges the gap without apologizing excessively, offers something genuinely useful (a new insight, relevant change, or concrete question), and opens the door without pressuring them to respond. Human, not automated-sounding. What to check: Is the "genuinely useful" thing actually useful, or did AI make something up? If you're referencing a market development, company news, or product update, make sure it's real before you send it. How to adapt these AI sales prompts to your specific context The prompts above are templates. They work as written, but they work better when you tune them to your actual situation. Here's how to get more out of them. Adjust the role instruction. "You are a B2B sales rep" is a fine default, but if you sell enterprise software with 18-month sales cycles, say that. If you're in a high-volume transactional role, say that too. The more the AI understands your context, the less generic the output. Add tone anchors. If you have a particular way you communicate with a client, tell the AI. "Match the tone of this previous email I sent: [paste]" is more effective than asking for "professional but friendly." Show it, don't describe it. Use your own examples as input. Past emails that landed well, CRM notes from a deal you won, a proposal that got a signature. Paste those as reference material. Tell the AI to match the structure or tone. You're training it on what good looks like for you specifically, not what good looks like in general. Chain prompts for complex tasks. For something like a full proposal, don't try to get it all in one prompt. Run the discovery summary first, then the buying criteria extraction, then the proposal outline, then each section separately. Each step gives the next one better context. The difference between AI saving you 20 minutes and AI actually making your output better usually comes down to how much context you're willing to put in upfront. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What AI still can't do in sales AI can organize your notes, draft your emails, and help you rehearse objections. It can't read the room. It doesn't know that your champion is fighting a political battle internally, or that the prospect went quiet because their budget got cut, not because they lost interest. It can't tell when the right move is to stop emailing and pick up the phone. There's a reason the best salespeople in any organization tend to be the ones with the deepest relationships and the sharpest instincts about timing. A buyer does not trust you because your CRM note is well-formatted. They trust you because you listened, remembered the real problem, and did not overpromise when the room got uncomfortable. AI can help you prepare for that. It cannot do that part for you. Trust, timing, and judgment are still yours to own. That's not a motivational poster point. It's a practical one. If you want to understand which parts of your job are genuinely hard for AI to take, the human skills AI can't replace breaks it down without the usual cheerleading. And on the question of whether to use AI at work openly or quietly, that's worth thinking through. The should you tell your boss you use AI article has the honest answer. A note on what not to paste into AI tools This deserves its own section because people skip it. Do not paste into any AI tool, unless your company's policy explicitly says you can: Real customer or prospect names and contact details Pricing exceptions, discount approvals, or contract terms Security architecture or compliance documentation Revenue numbers, deal sizes, or pipeline data Any data covered by GDPR, HIPAA, or your company's NDAs Anonymize everything. Replace names with "the prospect" or "the buyer." Replace company names with industry descriptors. Replace specific numbers with ranges or placeholders. If you're using a consumer AI tool like ChatGPT or Claude without a business agreement in place, assume your input could be used for training. Act accordingly. The Microsoft Work Trend Index found that 78% of AI users at work bring their own tools rather than using company-sanctioned ones. That number matters because it means most people are operating without any data policy guardrails. If that's you, be careful about what you put in. Frequently asked questions Can AI write my sales emails for me? AI can draft sales emails, but you need to edit them before sending. Check every claim, verify pricing and timelines, and make sure the tone matches your relationship with that specific person. A generic AI draft sent without editing is often worse than no email at all. Is it safe to paste customer information into ChatGPT? Not without anonymizing it first. Don't include real names, company names, deal sizes, contract terms, or regulated data unless your company has a data processing agreement with the AI provider. When in doubt, replace specifics with placeholders before pasting anything. What's the MEDDIC framework and why does it matter for sales prompts? MEDDIC stands for Metrics, Economic Buyer, Decision Criteria, Decision Process, Identify Pain, Champion. It's a qualification framework common in B2B sales. Using it in prompts gives the AI a structured way to help you spot gaps in what you know about an opportunity, rather than just generating generic advice. Can AI help me handle objections in a live sales call? Not live, no. AI is useful for preparing objection responses before a call. During the call itself, you're on your own. Use AI to rehearse and prepare, then put the phone down and listen to the actual person in front of you. How do I use these prompts without sounding like a robot? Read the draft out loud before you send anything. If you wouldn't say it in a conversation, rewrite it. The prompts include instructions to avoid filler and sound human, but AI still has a tendency toward slightly formal phrasing. Your job is to catch that and fix it. Will using AI for sales make me better or just faster? Faster at drafting, organizing notes, and preparing for calls. Whether that makes you better depends on how you use the time you save. If you spend it thinking more carefully about the deal and the relationship, yes. If you just send more volume with less thought, probably not. Do I need a special AI tool for sales, or will ChatGPT work? ChatGPT or Claude will handle everything in this list. You don't need a sales-specific AI tool. Most of the "AI for sales" products on the market are wrappers around the same underlying models, often with a CRM integration bolted on. Start with what you already have access to, and only pay for a specialized tool if you hit a genuine limitation. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance --- ## AI Sales Proposal Prompts: Draft Better Deals URL: https://dontreplace.me/articles/ai-sales-proposal-prompts-draft-better-deals AI promptssales proposalsproposal writingsales enablementcommercial review AI Sales Proposal Prompts: Draft Better Deals 10 AI sales proposal prompts to organize discovery, define scope, present approved pricing, check risky claims, and draft clearer deals. September 03, 2026 Updated September 03, 2026 A sales proposal should explain a deal. It should not become the place where your AI assistant discovers a bold new pricing model, invents three product capabilities, and promises implementation by next Tuesday. AI sales proposal prompts can turn verified discovery notes, approved scope, pricing, and delivery inputs into a clearer first draft. They can also expose missing information before a buyer does. They cannot know what the buyer secretly values, approve a discount, interpret a contract, or guarantee an outcome. AI can organize approved commercial facts. Humans still own discovery, pricing, scope, legal review, delivery promises, negotiation, and the signature. These ten templates help founders, account executives, consultants, and sales teams draft better proposals without appointing autocomplete as chief revenue officer. What a sales proposal is supposed to do A proposal connects an understood problem to an offer your organization can actually deliver. It gives the buyer enough clarity to evaluate the fit, compare options, involve stakeholders, and decide what happens next. Keep these categories separate: Buyer-confirmed fact: something the buyer said, wrote, or approved. Seller interpretation: your current reading of the situation, which may still need confirmation. Desired outcome: a result the buyer wants, not a result you can automatically guarantee. Deliverable: a specific thing your team has approved and can provide. Assumption: a condition the plan depends on. Dependency: an input, decision, access, or action another party must supply. Estimate: a planning range based on stated conditions. Commitment: a promise an authorized human has approved. Legal term: contract language that belongs in the approved agreement, not in a chatbot improvisation. A polished paragraph can blur all of those into one magnificent liability. “Our solution will increase conversion by 37% within 30 days” sounds efficient right up until somebody asks where the number came from. Before drafting, make sure the opportunity itself is supported by evidence. These AI sales prompts for discovery and follow-up help with the earlier conversation. If the notes came from recorded calls, use an evidence-first call coaching process rather than treating a transcript as divine revelation. The evidence-first proposal formula Add this instruction to any prompt below: “Act as a sales proposal drafting assistant. Use only the sanitized, approved discovery notes, scope, pricing, delivery inputs, and terms I provide for [buyer type, offer, decision stage, and proposal goal]. Produce [artifact and format]. Separate buyer-confirmed facts, seller interpretations, approved commitments, assumptions, dependencies, exclusions, unknowns, and items marked [VERIFY]. Do not invent buyer priorities, metrics, ROI, testimonials, capabilities, integrations, dates, staffing, discounts, legal terms, urgency, or approval. Do not turn desired outcomes into guarantees. Cite the supplied source label for every material claim. Flag conflicts and missing inputs instead of resolving them silently.” The phrase use only the supplied facts does the heavy lifting. A model is designed to complete the pattern. If your proposal looks like a typical software deal, it may helpfully add a typical implementation promise that nobody authorized. Never paste buyer names, emails, phone numbers, contracts, pricing exceptions, payment details, credentials, private procurement notes, confidential strategy, regulated data, or legally sensitive information into an unapproved AI tool. Use approved systems, minimum necessary data, de-identification, access controls, retention limits, and human legal, privacy, security, finance, and commercial review. What to collect before prompting Build a compact, sanitized proposal pack. Keep the original records in approved systems. InputWhy it mattersHuman owner Buyer-confirmed problemAnchors the proposal in evidenceAccount owner Desired outcomes and measuresClarifies what success meansBuyer and account owner Approved solution and deliverablesPrevents capability inventionProduct or delivery lead Scope exclusionsStops accidental expansionDelivery and commercial lead Approved pricing and currencyPrevents unauthorized numbersFinance or sales leadership Timeline range and conditionsKeeps dates honestDelivery lead Buyer responsibilitiesMakes dependencies visibleAccount owner Security and data requirementsRoutes specialist reviewSecurity/privacy owner Procurement requirementsAvoids late process surprisesBuyer contact and legal Approved legal termsKeeps drafting inside guardrailsLegal counsel Source labels and datesMakes claims traceableProposal owner Open questionsPreserves uncertaintyNamed owner and due date If half the table says “we think,” the proposal is not ready for decorative gradients. Go back to discovery or mark the gaps explicitly. These requirements gathering prompts can structure the missing questions without inventing the answers. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI sales proposal prompts Replace brackets with sanitized, approved information. Treat every output as a draft requiring accountable human review. 1. Clean and qualify discovery notes “Review these sanitized discovery notes and source labels: [paste]. Create sections for buyer-confirmed facts, seller interpretations, desired outcomes, constraints, stakeholders, decision process, timing, budget information, open questions, and conflicting statements. Cite the source label for each item. Mark unsupported summaries [VERIFY]. Do not infer priorities, authority, urgency, emotion, budget, or intent.” A proposal built on messy notes merely makes the mess look expensive. Start by separating what the buyer actually confirmed from what the team repeated in Slack until it felt true. Ask the account owner to verify the output against the approved source. If two stakeholders described different priorities, preserve both versions. The model should not choose the more proposal-friendly one. That is a conversation for humans. This step also reveals whether the team has enough evidence to draft. Missing budget information does not mean “budget approved.” A vague request for speed does not mean “must launch this quarter.” Unknown remains a valid and professionally useful answer. 2. Build a traceable proposal outline “Using only this verified proposal pack [paste], draft an outline with: buyer context, confirmed problem, desired outcomes, proposed approach, deliverables, exclusions, timeline, buyer responsibilities, options, pricing, assumptions, risks, open questions, and next steps. After each section, list the approved sources it relies on. Leave unsupported sections as [INPUT NEEDED]. Do not write persuasive claims yet.” An outline is cheaper to correct than six pages of polished nonsense. Review the sequence with delivery, finance, and the account owner before asking AI to expand it. Remove sections the buyer does not need. A proposal is not improved by making it resemble a municipal bond prospectus. The goal is decision clarity, not page count. A traceability note can stay internal. It gives reviewers a fast way to check where claims came from and catches the classic problem where the executive summary promises something the scope section quietly does not include. 3. Map the problem to approved outcomes “Create a problem-to-outcome map from these buyer-confirmed statements and approved capabilities: [paste]. For each problem, show the cited evidence, desired outcome, proposed deliverable, how the deliverable plausibly contributes, what remains outside our control, how progress could be measured, and what needs verification. Do not claim causation or guarantee an outcome.” This is where proposal writing often wanders into fake ROI. A service can contribute to faster processing without guaranteeing a specific revenue increase. A new workflow can reduce manual steps without promising that every employee will adopt it. Make the chain visible: problem, intervention, plausible contribution, measurement, and uncertainty. If the company does not control the buyer’s traffic, staffing, approvals, or implementation, say so. For material customer, compliance, or operational consequences, run an AI risk assessment and have the appropriate specialist review it. 4. Draft an honest executive summary “Draft a 150–250 word executive summary using only these verified inputs: [paste]. State the buyer-confirmed situation, the approved proposed approach, key deliverables, expected contribution, major dependency, and next decision. Label unconfirmed details [VERIFY]. Use plain language. Do not add urgency, market statistics, ROI, guarantees, social proof, or adjectives unsupported by the source pack.” The executive summary should help a stakeholder understand the decision without reading every appendix. It should not behave like a movie trailer voiced by a man who says “in a world.” Check every noun, number, and promise. “Integrated platform” may mean one approved connection, not universal compatibility. “Rapid launch” may mean six weeks if access arrives on time. Replace elastic praise with concrete scope. These executive summary prompts can help tighten the language, but commercial accuracy matters more than elegance. 5. Define scope, exclusions, and change control “Turn these approved deliverables and constraints into a scope table: [paste]. Include deliverable, acceptance evidence, owner, buyer input, estimated timing, dependency, exclusion, and change-control trigger. Use only approved details. Mark conflicts or missing acceptance criteria [VERIFY]. Do not add features, rounds, channels, integrations, support levels, or warranties.” Scope needs edges. “Create campaign assets” is not useful if one person imagines three ads and another imagines a global content factory with localization in nineteen languages. Define what is included, how completion will be recognized, what the buyer must provide, and what would require a change. Use AI acceptance criteria prompts to make deliverables testable, then let delivery owners approve the wording. Exclusions are not hostile. They protect both sides from discovering incompatible expectations after kickoff. Write them plainly, without hiding them in eight-point gray text beneath a stock photo of a handshake. 6. Present approved options and pricing “Format these human-approved commercial options: [paste exact options, currency, taxes, payment schedule, validity date, and approval notes]. For each option, list included deliverables, exclusions, dependencies, timeline conditions, and best-fit situation using only supplied facts. Do not calculate discounts, recommend an option, alter prices, add bonuses, imply scarcity, or invent comparative savings.” AI may format pricing. It does not authorize pricing. Copy the numbers from the approved source and compare them character by character after generation. Check currency, decimal separators, tax treatment, payment milestones, expiration dates, minimum terms, usage limits, and whether optional work is clearly optional. If the model performs arithmetic, recalculate it independently using the approved finance process. A “best fit” note can help buyers compare choices, but it should be grounded in stated needs. Do not manufacture a decoy package or claim that “most clients choose” an option without real approved evidence. 7. Draft a conditional implementation timeline “Create a proposed implementation timeline from these approved phases, estimates, dependencies, and capacity assumptions: [paste]. Show phase, work, owner, prerequisite, estimated duration, decision gate, and evidence of completion. Use ranges where supplied. State that dates depend on kickoff, access, approvals, and change control. Do not invent calendar dates, staffing, parallel work, or guaranteed launch timing.” A timeline is a model of dependencies, not a prophecy. “Four weeks” can become eight when access arrives on day twenty-three and the final approver is hiking somewhere without signal. Ask delivery to verify sequencing and capacity. Ask the buyer to verify their responsibilities. Distinguish elapsed time from effort and estimates from contractual commitments. If a date is genuinely fixed, document the assumptions required to hit it and the decision points where scope, timing, or resources must change. Optimism is not a project-management methodology. 8. Document assumptions, dependencies, and risks “Create an assumptions and dependencies register from this proposal pack: [paste]. For each item, include statement, source, owner, validation method, due date, impact if false or late, and proposed response. Separate confirmed facts from working assumptions. Add a risk section using only risks supported by supplied evidence. Do not invent probabilities or present mitigations as guarantees.” This register turns invisible conditions into visible decisions. It is especially useful when a price assumes clean data, prompt feedback, existing licenses, buyer-provided content, or access to a named system. Review security, privacy, accessibility, procurement, legal, and operational dependencies with the relevant owners. A seller should not casually summarize a data-processing obligation from memory because the paragraph sounded reassuring. A good proposal admits what must be true. A bad one buries those conditions and hopes kickoff develops amnesia. 9. Audit the draft for unsupported promises “Audit this proposal against the approved source pack: [paste both]. Produce a table of every material claim, number, date, deliverable, capability, integration, outcome, testimonial, legal statement, pricing term, and customer responsibility. Cite its source or mark [UNSUPPORTED]. Flag contradictions, desired outcomes written as guarantees, estimates written as commitments, vague scope, missing exclusions, stale information, and language requiring legal, finance, security, privacy, product, or delivery review. Do not repair unsupported claims by inventing support.” This is one of the safest uses of AI in proposal work: finding sentences humans should inspect. It is not a substitute for specialist review. Search for dangerous words such as “will,” “guaranteed,” “seamless,” “fully,” “compliant,” “secure,” “all,” “any,” and “unlimited.” None is automatically wrong. Each deserves evidence and an authorized owner. Also check consistency across the executive summary, pricing tables, scope, timeline, and terms. Proposal errors love duplication because a change made in one section can leave an older promise elsewhere. 10. Build the final human review checklist “Create a role-based pre-send checklist for this proposal: [paste draft and review policy]. Include account owner, delivery, product, finance, legal, security/privacy, accessibility, brand, and executive approval only where relevant. For each check, name the exact section, evidence required, decision owner, and status: approved, changes required, not applicable, or blocked. End with a release gate that prohibits sending while any material item is unsupported or unapproved.” The checklist should make accountability obvious. “Reviewed by team” is not an approval trail. Name who verifies the price, who owns the delivery promise, who checks legal language, and who presses send. Have a human inspect the rendered PDF or document, not only the editable source. Check links, pagination, tables, comments, tracked changes, hidden slides, filenames, metadata, and accessibility. Nothing says enterprise confidence like sending FINAL_v8_use_this_REAL.docx with an internal comment asking whether the margin is “made up lol.” A controlled proposal workflow Use the prompts in a sequence that preserves evidence and authority: Confirm the buyer wants a proposal and clarify the decision it must support. Collect minimum necessary, approved source material. Sanitize sensitive information before using an approved AI tool. Qualify discovery notes and preserve disagreements or unknowns. Build and approve the outline before expanding prose. Map problems to deliverables without promising outcomes outside your control. Insert only approved scope, pricing, timing, and terms. Run the unsupported-claims audit. Route each flagged section to its accountable human owner. Review the final rendered artifact and record approval before sending. Save the approved version in the proper system. Treat buyer feedback as new evidence, not permission for the model to renegotiate. If the buyer asks for a change, update the source of truth first. Then regenerate the affected section and repeat review. Do not let facts live exclusively inside a chat history nobody can audit. Common ways AI proposal writing goes wrong It upgrades aspirations into guarantees The buyer wants lower costs; the proposal suddenly promises a 30% reduction. Describe the intended outcome and your contribution without manufacturing certainty. It invents capabilities and integrations Models complete familiar product patterns. Verify every feature, system, limit, version, and configuration with the product or delivery owner. It improvises pricing Never ask an AI tool to “make the offer compelling” and assume the resulting discount is authorized. Supply exact approved terms and verify the output. It hides scope ambiguity behind good prose A beautiful sentence can still fail to define quantity, owner, acceptance, revision limits, dependencies, or exclusions. Prefer a precise table over a luxurious fog bank. It copies private data into an unapproved tool Proposal packs can contain customer strategy, contracts, contact details, pricing, security requirements, and regulated information. Minimize, sanitize, control access, and follow policy. It treats legal review as a tone edit Contract terms, warranties, liability, intellectual property, privacy, compliance, and regulatory claims need qualified human review. AI can flag text for review; it cannot approve it. Frequently asked questions Can ChatGPT write a sales proposal? It can draft and organize a proposal from approved inputs. It cannot verify those inputs, approve commercial terms, promise delivery, or replace discovery and specialist review. Treat the result as a draft. What is the best AI prompt for proposal writing? Use a prompt that supplies verified buyer context, approved scope, exact pricing, constraints, sources, and a required format. Require the model to separate facts, assumptions, commitments, exclusions, and unknowns. The evidence-first formula above is a useful base. How do I stop AI from making up proposal details? Limit it to supplied evidence, require source labels for material claims, tell it to mark missing information [VERIFY], and run a claim-by-claim audit against the approved pack. A human must still check the result. Should I upload a customer contract to ChatGPT? Not to an unapproved tool. Contracts may contain confidential and legally sensitive information. Follow your organization’s policy, tool agreement, access controls, and legal/privacy guidance. Use minimum necessary sanitized excerpts when explicitly permitted. Can AI calculate proposal pricing or ROI? It can format approved numbers and assist with scenarios, but calculations must be independently verified. ROI also depends on assumptions and buyer-controlled factors. Do not present generated estimates as guaranteed results. Can AI personalize a proposal for a buyer? It can tailor emphasis using buyer-confirmed information you are authorized to use. It should not infer sensitive traits, scrape private information, invent familiarity, or pretend to know priorities the buyer never stated. Who should approve an AI-assisted proposal? The account owner should verify buyer facts; delivery and product owners should verify capability, scope, and timing; finance should verify pricing; and legal, security, privacy, or other specialists should review relevant claims. Approval depends on the actual deal and company policy. Should I tell the buyer AI helped draft the proposal? Follow contractual obligations, organizational policy, professional standards, and applicable rules. Regardless of disclosure, your organization remains accountable for every claim and commitment in the document. Use the machine for structure, not the deal AI can save time by organizing notes, drafting sections, formatting approved options, and finding claims that need review. That is useful work. It becomes reckless when the polished draft outruns the evidence and authority behind it. Keep a source of truth. Label assumptions. Preserve unknowns. Verify every number and promise. Let delivery, finance, legal, security, and the account owner own their decisions. Then send a proposal that describes a deal people can actually keep. That is the practical point in what AI can and cannot do: fast language is not commercial judgment. If you want the wider field guide for using these tools without outsourcing your brain, Don’t Replace Me by Dmitry Kargaev is the low-pressure next read. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Account Planning Prompts: Build a Better Plan 10 AI account planning prompts to organize buyer evidence, map stakeholders, expose risks, test strategy, and plan next actions without inventing intel. AI promptsaccount planningsales strategy AI Sales Call Coaching Prompts: Improve the Conversation 10 AI sales call coaching prompts to review evidence, improve discovery, practice objections, protect private data, and coach reps fairly. AI promptssales call coachingsales enablement AI Sales Enablement Prompts: Build Useful Materials 10 AI sales enablement prompts for grounded messaging, training, coaching, role-play, objection practice, and seller resources without invented claims. AI promptssales enablementsales training --- ## AI Sales Territory Planning Prompts: Divide Without Drama URL: https://dontreplace.me/articles/ai-sales-territory-planning-prompts-divide-without-drama AI promptssales territory planningrevenue operationssales managementcapacity planning AI Sales Territory Planning Prompts: Divide Without Drama 10 AI sales territory planning prompts to compare coverage scenarios, expose bad assumptions, protect sensitive data, and make accountable decisions. September 01, 2026 Updated September 01, 2026 Sales territory planning has a special talent for turning a spreadsheet into a constitutional crisis. One rep sees opportunity. Another sees three airports, 400 dead accounts, and a quota assembled by raccoons. AI sales territory planning prompts can clean inputs, compare allocation scenarios, expose conflicts, and prepare better review questions. They cannot discover true account potential from stale CRM fields, decide what is fair, or absorb the consequences of moving a customer relationship. AI can organize territory evidence. Humans still own strategy, fairness, privacy, communication, and the final decision. These ten templates help sales leaders, revenue operations teams, founders, and account executives use AI for territory planning without promoting autocomplete to vice president of maps. What sales territory planning actually is A sales territory is an explicit assignment of accounts, prospects, products, industries, channels, or geography to a person or team for a defined period. Territory planning is the process of deciding those assignments, testing tradeoffs, documenting rules, and managing the transition. Keep these concepts separate: Coverage: who is responsible for which market or account. Capacity: realistic work a team can handle after ramp time, travel, support duties, and complexity. Potential: an evidence-based estimate of opportunity, not last year's revenue with optimism sprinkled on top. Quota: an approved commercial target. It is related to territory design but is not the same thing. Ownership: the operating rule for who leads a relationship, including overlays and partners. Scenario: a conditional territory option based on stated criteria and assumptions. Fairness: a reviewable process that considers opportunity, workload, constraints, and impact—not identical account counts. Fifty tiny dormant accounts are not automatically equivalent to five strategic accounts. A compact city can require less travel than a “small” rural territory. Historical revenue may reflect the prior rep's tenure, product fit, channel support, or one unusual renewal rather than future potential. Start with an evidence-first sales pipeline review and check your sales forecast assumptions. Territory design built on duplicate accounts and fantasy close dates is just cartography-themed fan fiction. The evidence-first territory planning formula Add this instruction to any prompt below: “Act as a sales territory planning assistant. Use only the sanitized, approved evidence I provide for [market, team, period, products, and decision]. Our design criteria are [criteria], ownership rules are [rules], and constraints are [constraints]. Produce [artifact]. Separate verified facts, calculated measures, assumptions, scenarios, tradeoffs, and unknowns. For every material conclusion, cite the source, definition, owner, and last-updated date; show missing or contradictory evidence; identify affected groups; and propose a validation action. Do not invent account potential, buyer intent, employee capability, protected characteristics, location, relationships, probabilities, or fairness claims. Do not rank employees or recommend employment action. Flag duplicate accounts, parent-child conflicts, stale fields, small samples, proxy discrimination, workload gaps, transition risk, and decisions requiring sales leadership, finance, HR, legal, privacy, or customer review.” The key phrase is use only the evidence provided. If a model cannot find a required fact, “unknown” is useful. A plausible guess about account potential can quietly become somebody's impossible quarter. Never paste buyer names, emails, contracts, private pricing, payment details, credentials, employee performance notes, health or family information, protected characteristics, confidential strategy, or legally sensitive data into an unapproved AI tool. Use approved systems, minimum necessary fields, aggregation, de-identification, access controls, retention limits, and human review. What to collect before prompting Build a sanitized planning pack while keeping source records in approved systems. InputWhy it mattersHuman check Planning period and business goalDefines the decisionSales leadership approves Account hierarchy and ownershipPrevents duplicates and collisionsRevenue operations verifies Product, segment, and channel rulesDefines valid coverageCommercial owners confirm Verified account attributesSupports groupingData owner validates Historical revenue and pipelineAdds context, not destinyFinance reconciles Workload driversMakes capacity less fictionalManagers validate Rep location, role, and ramp statusSupports operating feasibilityHR and managers review Existing customer relationshipsProtects continuityAccount owners confirm Strategic accounts and exclusionsPreserves deliberate choicesLeadership approves Known gaps and update datesLimits false confidenceCRM owner audits Transition dates and dependenciesPrevents dropped coverageOperations coordinates Fairness and privacy guardrailsControls high-impact useHR, legal, privacy review Do not silently fill missing fields. Missingness may itself be uneven: one region may have excellent enrichment while another has old records. Treating that difference as market potential rewards documentation, not opportunity. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI sales territory planning prompts Replace brackets with sanitized, approved evidence. Every output is a draft for accountable human review. 1. Clean the territory planning inputs “Turn this sanitized account dictionary, hierarchy export, ownership file, product rules, segment definitions, source inventory, refresh dates, and known limitations into a territory input brief: [paste]. Show the unit of analysis, required fields, duplicate logic, parent-child treatment, missingness, conflicting owners, stale records, source owner, and privacy classification. Flag definitions that changed across teams and fields that should be removed or aggregated.” Run this before asking for territories. If one system defines an account as a legal entity and another defines it as a buying location, a neat account count can compare entirely different objects. Revenue operations should verify joins, merged companies, subsidiaries, house accounts, partners, named accounts, renewals, and customer ownership. AI can flag suspicious records. It cannot know which record reflects the real commercial relationship. 2. Define territory design criteria “Convert these human-approved business goals and constraints into a territory design scorecard: [paste]. For each criterion, define its purpose, permitted evidence, calculation, weight or priority, known limitation, owner, review threshold, and prohibited proxies. Separate must-have constraints from preferences. Identify conflicts such as growth versus continuity, specialization versus simplicity, and equal account count versus equal workload.” A design cannot be evaluated until “balanced” means something inspectable. It might mean comparable opportunity ranges, manageable workload, customer continuity, travel feasibility, specialist access, or all of those with tradeoffs. Do not let the model choose weights from vibes. Leadership must approve priorities, and HR or legal should review criteria that could affect employment opportunities. A neutral-looking variable such as ZIP code, school, commute, or customer type may act as a proxy for sensitive characteristics. 3. Estimate workload without fake precision “Using these sanitized account aggregates and human-approved workload rules, create workload ranges by proposed territory: [paste]. Consider active opportunities, renewal timing, account complexity, travel, language or certification needs, service obligations, channel coordination, and ramp support. Show assumptions, missing inputs, unusually influential accounts, and sensitivity to each rule. Do not infer employee speed, effort, competence, or availability.” Account count is a terrible universal workload unit. One account may have five business units, an active security review, and a procurement process older than the building. Another may be a clean self-service renewal. Use ranges instead of suspicious precision. Managers should validate whether the workload drivers reflect actual work. Keep employee evaluation out of this analysis; territory workload and individual performance are different questions with different evidence and safeguards. 4. Compare geographic and account-based models “Build three territory scenarios from these sanitized market aggregates and approved rules: geographic, account-based, and hybrid: [paste]. For each scenario show coverage logic, opportunity range, workload range, travel or coordination burden, specialist needs, account continuity, channel overlap, operational complexity, privacy risks, transition cost, and unresolved decisions. Do not recommend a winner until humans approve the criteria.” Geography is easy to explain but may split global account families. Named-account models preserve focus but can create leftovers nobody owns. Hybrid designs can fit reality while producing an ownership manual with the emotional warmth of tax law. The useful output is a visible tradeoff table, not “Scenario B is optimal.” Humans must decide which complexity the organization can operate and which customer relationships it cannot casually move. 5. Balance opportunity and capacity “Compare these proposed territory aggregates against our approved opportunity and capacity ranges: [paste]. Identify territories outside thresholds, explain which inputs drive the difference, show missingness, and suggest rule-based adjustments for review. Preserve strategic-account constraints and relationship ownership. Do not move accounts automatically, rank reps, or claim equal projected revenue means fairness.” Opportunity estimates are assumptions, especially in new markets. Historical revenue can understate whitespace or overstate a territory supported by one inherited whale. Capacity also changes with ramping, specialist support, travel, and customer complexity. Ask for several adjustment options: move a cluster, alter overlay support, narrow a product scope, stage a transition, or leave an intentional imbalance with a documented reason. Territory planning is not a bin-packing contest where customers are interchangeable rectangles. 6. Identify named-account and ownership conflicts “Review these sanitized account hierarchies, current owners, partner relationships, overlays, open opportunities, renewal dates, and proposed assignments: [paste]. Flag parent-child splits, duplicate ownership, global-local conflicts, partner collisions, customer-contact confusion, open-deal disruption, and records with no owner. For each conflict, cite the rule involved, affected roles, customer risk, decision owner, and validation step.” Ownership disputes become expensive when customers receive competing outreach or an active opportunity changes hands without context. The model can compare files and rules, but account leaders must confirm the real relationship. Do not include unnecessary customer or employee identifiers in the output. Use stable internal IDs where possible, restrict access, and route decisions through approved operating channels—not a public spreadsheet titled territories_FINAL_v7_reallyfinal. 7. Stress-test territory scenarios “Stress-test these human-designed territory scenarios against the following conditions: major-account loss, merger, hiring delay, rep ramp, product change, channel conflict, regional disruption, unexpected renewal concentration, and data-quality correction: [paste]. Show which territories become overloaded, uncovered, concentrated, or operationally ambiguous. State scenario triggers, contingency owner, and reversible response. Do not invent likelihoods.” A territory plan that works only when every hire starts on time and every account remains exactly where the CRM put it is not a plan. It is a decorative forecast. Stress tests reveal brittle rules. Use risk assessment prompts to structure downstream impacts, then have sales, finance, operations, HR, legal, and customer leaders review consequences in their own domains. 8. Review fairness, bias, and privacy risk “Audit this proposed territory process and aggregate outcomes for fairness, bias, and privacy risks: [paste]. Review data provenance, missingness by group or region, prohibited attributes, potential proxies, opportunity and workload ranges, access to strategic accounts, support allocation, appeal paths, documentation, and monitoring. Do not infer protected characteristics or declare the design legally compliant. Produce questions for HR, legal, privacy, and sales leadership.” Fairness is not proven because a model produced equal totals. Account quality, relationship maturity, travel, product fit, support, ramp timing, and market volatility can make equal-looking books radically different. Never ask AI to infer ethnicity, age, disability, family status, health, religion, or other sensitive traits from names, locations, writing, photos, or behavior. Do not use territory planning as a back door for automated employee scoring. High-impact decisions require approved policy, complete context, an appeal mechanism, and accountable people. 9. Draft a territory-change explanation “Draft separate change briefs for sales leaders, affected sellers, support teams, partners, and customers using these approved decisions: [paste]. Explain what changes, what stays, the evidence and criteria used, known tradeoffs, effective date, transition support, customer-continuity plan, decision owner, question channel, and appeal or correction process. Do not expose private employee data, confidential strategy, or unsupported claims of fairness.” People can disagree with a decision while still understanding it. Hiding behind “the model optimized coverage” destroys accountability and makes every correction look arbitrary. Keep the explanation proportional. A customer needs to know who owns the relationship and how continuity is protected. An affected seller needs the operating rules, transition expectations, and route for correcting bad data. Neither needs a dump of private notes. 10. Convert the approved design into a transition plan “Turn this human-approved territory design into an accountable transition plan: [paste]. For each action include the territory or account group, current owner, approved future owner, customer-impact level, open opportunities, renewal timing, required handoff artifact, systems to update, owner, due date, dependency, communication, access change, validation check, rollback condition, and review date.” “Update CRM” is not a transition plan. A useful plan says who validates account hierarchies, who briefs the new owner, when customer communication happens, which open opportunities retain temporary ownership, and how the team confirms that routing and permissions work. Use a change impact analysis before launch. Territory changes affect forecasts, quotas, commissions, routing, dashboards, support queues, partner rules, access controls, and customer trust. The map is only the visible part. Common territory planning failures Treating historical revenue as future potential History reflects prior coverage, product availability, pricing, customer concentration, and luck. Use it as one input. Label assumptions about whitespace and future demand, and assign validation work. Optimizing account count instead of work Equal counts can hide enormous differences in complexity, travel, renewal load, and active pipeline. Define workload drivers with managers and test ranges. Confusing an estimate with a fact Enrichment scores and market potential fields often contain opaque methods, stale records, or missing regions. Record provenance and do not let a polished summary upgrade an estimate into truth. Automating employee judgment Territory records do not explain effort, skill, integrity, or future performance. Do not use model-generated territory analysis to rank employees or make punitive decisions. Those choices require fair policy, relevant evidence, and accountable review. Ignoring customer continuity Moving an account can break trust, delay an open deal, confuse a partner, or drop context. Include relationship history, active commitments, handoff quality, and customer communication in the decision. Hiding strategy behind “the algorithm” Every material criterion, weight, exception, and tradeoff needs a human owner. If leadership cannot explain why a territory exists, the model should not be the designated scapegoat. Leaking customer or employee information A supposedly anonymous combination of region, deal size, product, and renewal month may identify one customer or rep. Aggregate carefully, suppress small groups, minimize fields, and restrict access. A practical territory review checklist Before approving an AI-assisted territory plan, confirm: The planning period, goal, unit of analysis, and ownership rules are explicit. Account hierarchies, duplicates, current owners, and strategic exclusions were verified. Every material measure has a source, definition, owner, and update date. Opportunity estimates remain labeled as estimates. Workload includes complexity, travel, renewal timing, and support obligations. Missing and contradictory data remains visible. Scenarios show tradeoffs, transition cost, and stress-test results. Protected characteristics and questionable proxies were excluded. Employee ranking and automated employment decisions are outside scope. Customer continuity and open opportunities have explicit handoff rules. HR, legal, privacy, finance, operations, and sales leaders reviewed relevant risks. Affected people have a route to correct bad data or raise an exception. System, routing, access, dashboard, quota, and commission changes are coordinated. The final decision belongs to named humans, not “the AI.” If several checks fail, do not ask the model to make the deck more persuasive. Repair the evidence, rules, and review process. Frequently asked questions Can ChatGPT create sales territories automatically? It can organize sanitized inputs, compare human-defined scenarios, flag conflicts, and draft review artifacts. It should not automatically assign accounts, decide fairness, evaluate employees, or approve territory changes. Account knowledge, strategy, legal risk, and consequences require accountable humans. What data should I use with AI sales territory planning prompts? Use the minimum approved evidence needed: verified account hierarchies, aggregate market attributes, ownership rules, sanitized revenue and pipeline context, workload drivers, strategic constraints, relationship continuity, assumptions, and known gaps. Keep source records and identifiers in approved systems. Can AI estimate account potential? It can apply a documented human-approved method to supplied data. It cannot make weak or stale data trustworthy. Potential estimates need provenance, validation, ranges, monitoring, and review by people who understand the market and customer. How do I make territories fair? Define fairness criteria before comparing designs. Review opportunity, workload, customer continuity, travel, support, ramp timing, market volatility, and access to strategic work. Exclude protected traits and suspect proxies. Provide transparent documentation, human review, and a correction or appeal route. Should every rep get the same number of accounts? Usually not. Account count is only one workload signal. Complexity, active opportunities, renewal timing, products, travel, partners, and customer needs can make equal counts deeply unequal in practice. Can AI choose which rep should own a strategic account? It can summarize approved criteria and relationship evidence. Humans should decide ownership because customer trust, experience, development opportunity, employment fairness, and commercial strategy require context and accountability. How often should territories be redesigned? Use a stable planned cadence with defined exception triggers. Constant changes damage customer continuity and seller focus. Review when strategy, staffing, market structure, products, channels, or data changes materially, then weigh transition cost against expected benefit. How do I protect customer and employee data? Use approved tools, minimize fields, aggregate or de-identify where possible, suppress small groups, restrict access, set retention limits, document purpose, and involve privacy, HR, or legal reviewers when needed. Never casually paste contracts, credentials, private notes, or protected information into consumer AI tools. The useful boundary AI is fast at reorganizing account evidence, comparing rules, and exposing candidate inconsistencies. It is not the customer, seller, sales leader, HR partner, or person accountable when a territory change damages trust or somebody's livelihood. Use these prompts to make criteria visible, scenarios inspectable, and transitions less chaotic. Then do the human work: validate the data, understand the relationships, review fairness, explain the decision, and own the outcome. That boundary is the point of what AI can and can't do and the broader no-BS guide to using AI at work. For more practical rules about keeping judgment human, Don't Replace Me by Dmitry Kargaev is the field guide—not permission to outsource your spine. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Sales Forecast Prompts: Skip the Fake Certainty 10 AI sales forecast prompts to test evidence, build scenarios, challenge assumptions, protect buyer data, and produce an accountable forecast. AI promptssales forecastingrevenue operations AI Sales Pipeline Review Prompts: Find Risk Early 10 AI sales pipeline review prompts to inspect deal evidence, challenge forecasts, protect buyer data, and turn pipeline risk into accountable action. AI promptssales pipelinesales forecasting AI Customer Renewal Prompts: Prepare With Evidence 10 AI customer renewal prompts to verify outcomes, expose account risks, prepare options, and protect trust without inventing value or customer sentiment. AI promptscustomer renewalcustomer success --- ## AI Scope Creep Prompts: Stop Projects Drifting URL: https://dontreplace.me/articles/ai-scope-creep-prompts-stop-projects-drifting scope creepAI promptsproject managementchange controlstakeholder management AI Scope Creep Prompts: Stop Projects Drifting 10 copy-paste AI scope creep prompts to compare requests, write change logs, flag risks, and keep projects on track. Includes safety rules and escalation tips. July 12, 2026 Updated July 12, 2026 Scope creep doesn't announce itself. It shows up as a "quick ask" in a Slack message at 4pm on a Thursday. It arrives dressed as a helpful suggestion from a stakeholder who wasn't in the original scoping call. By the time you notice it, three undocumented features are halfway built, the timeline is fiction, and nobody can remember what was actually agreed. AI scope creep prompts won't save your project on their own. But they'll help you structure the conversation, document the drift, and put the real question back where it belongs: in front of the people with the authority to say yes or no. Here's the complete set. Copy them, adapt them, use them. Why scope creep is so hard to catch in real time The problem isn't that people are malicious. Most scope creep comes from reasonable people making reasonable requests. A client sees the prototype and thinks "while you're in there..." A product manager hears a new idea in a meeting and adds it to the notes. A developer decides to "just improve" a thing that wasn't in scope because it seemed wrong to leave it. Nobody flags it. Nobody checks it against the original agreement. And then suddenly you're three weeks behind and the budget is bleeding and everyone is confused about how you got here. The other problem is that "it's a small change" is almost never true once you trace the dependencies. That's where AI actually helps: it's fast at pulling apart a vague request and mapping what it touches. The judgment call about whether to do it still belongs to you. Before you try any of these, read the no-BS guide to using AI at work if you haven't already. The tools are only as good as the context you give them. The one thing that makes all AI scope prompts work better Every prompt below follows the same logic: garbage in, garbage out. Rule #13 of Don't Replace Me is blunt about this. Feed AI a half-remembered scope document and a vague description of what someone asked for, and you'll get a confident-sounding analysis that's built on sand. Before you run any of these prompts, gather: The original signed scope, SOW, or project brief The specific new request, in writing if possible Current timeline and budget status Named decision-makers for this project Any existing change-control process your team or client uses The more specific your inputs, the more useful the outputs. And never paste customer PII, employee records, contracts with legal terms, financial forecasts, security vulnerabilities, confidential pricing, or unreleased product details into an unapproved AI tool. Anonymise, summarise, or check with your IT or security team first. The reusable formula for any scope creep prompt Most of the prompts below follow a variation of this structure. You can build your own using it: Act as: [role, e.g. "a project manager reviewing scope on a fixed-fee software project"] Context: [original scope summary, current state, what was agreed] New request: [the specific ask that arrived after sign-off] Constraints: [timeline, budget, team capacity, contractual limits] Your job: [what you want AI to produce: comparison, impact analysis, options, draft message, etc.] Do not: [invent commitments, assume approval, speculate on legal or financial liability] That last line matters. AI will happily write "the client has agreed to..." or "this is within the original contract" if you let it. It doesn't know what was agreed. Only you do. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI scope creep prompts you can use today Prompt 1: Compare a new request against original scope Act as a project manager. I'll give you the original agreed scope and a new request that arrived after sign-off. Compare them and list: (1) what's clearly in scope, (2) what's clearly out of scope, (3) what's ambiguous and needs clarification. Do not assume anything is approved. Flag anything that needs a decision from a named stakeholder. Original scope: [paste summary or key deliverables] New request: [paste the actual ask] Decision-maker: [name] Use this every time someone says "can we just add..." Start here before anything else. Prompt 2: Turn a vague addition into an explicit change request I have a verbal or informal request to change the project. Help me write a formal change request document that includes: a plain-language description of what's being asked, what it replaces or adds to the current scope, estimated impact on timeline and budget (leave blank if unknown), dependencies it creates, and a decision field for the named approver to sign off. Do not invent numbers. Mark anything uncertain as [TO BE CONFIRMED]. Informal request: [describe it] Project name: [name] Named approver: [name and role] This forces the "quick ask" to become a real document. Most scope creep dies here because nobody actually wants to fill in the form. Prompt 3: Estimate impact on timeline and budget I need to think through the impact of adding [describe change] to a project currently at [current status: e.g. week 4 of 10, 60% budget spent]. Help me map the likely impact by listing: tasks that would need to be added or reworked, downstream dependencies that could be affected, risks to the current delivery date, and open questions I need to answer before I can give a real estimate. Do not invent numbers. Flag everything uncertain. Current scope: [summary] Change being considered: [description] Team capacity: [e.g. two developers, one designer, part-time PM] This is the conversation starter for your delivery team, not the final answer. Treat it as a structured starting point. Prompt 4: Identify hidden dependencies Act as a senior delivery lead. I'm considering a change to this project. Help me identify non-obvious dependencies this change might create or affect, including: technical dependencies, process dependencies, third-party or vendor dependencies, data or privacy dependencies, and anything that could affect other workstreams or teams. Ask me clarifying questions if my description is too vague to give useful output. Change description: [describe it] Tech stack or process context: [brief description] This is where AI earns its keep. Tracing dependencies is tedious and easy to miss. It's fast and reasonably thorough at this. Your job is to verify the output with the people who actually own those systems. A good starting point for this is clear requirements before you start, which covers the same dependency-mapping logic at the beginning of a project. Prompt 5: Separate must-have from nice-to-have Help me structure a prioritisation conversation with a stakeholder who has requested several additions to the project scope. For each item I list, help me frame a question that forces a clear must-have or nice-to-have decision, and suggest criteria we could use to make that call (e.g. revenue impact, contractual requirement, customer safety, regulatory compliance, delivery risk). Do not make the prioritisation decision yourself. Requested additions: [list them] Project goal: [one sentence] Stakeholder role: [name and relationship to project] Prompt 6: Write a polite pushback message Help me write a professional message to a [client/internal stakeholder/executive] explaining that their request falls outside the current agreed scope. The message should: acknowledge the value of the request, explain clearly why it's out of scope, describe what happens next (change request process), and avoid sounding defensive or obstructive. Keep it under 150 words. Original scope summary: [paste brief summary] What they asked for: [describe] Change-control process: [describe your process] My relationship with this person: [e.g. direct client, senior stakeholder, partner team] For harder conversations, the client communication prompts have more templates for when the stakes are higher. Prompt 7: Prepare options for stakeholders A stakeholder wants to add [describe change] to the project. I can't just say no, but I also can't absorb it without impact. Help me draft three options to present: Option A (full change, formal change request with cost/time impact), Option B (reduced version of the change that fits within constraints), Option C (defer to next phase with no impact now). For each option, list the tradeoffs clearly. Mark anything I need to verify before presenting. Change requested: [describe] Current constraints: [timeline, budget, capacity] Phase 2 or next milestone: [if applicable] Prompt 8: Update a decision log entry Help me write a decision log entry for a scope change that was discussed today. Include: date, what was requested, who requested it, what was decided, who approved the decision, any conditions or caveats, and what the next action is. Keep it factual. Do not infer approval that wasn't explicitly given. What happened: [describe the conversation or meeting] Decision made: [what was agreed] Approved by: [name and role] Conditions: [if any] The decision log prompts cover this in more depth if your team keeps relitigating past decisions. Prompt 9: Flag risks and tradeoffs before saying yes Before I agree to add this change, help me think through the risks. List: delivery risks (could this delay the current milestone?), quality risks (does it introduce technical debt or reduce test coverage?), budget risks (where does the money come from?), stakeholder risks (who else needs to know?), and contractual or compliance risks I should flag to legal, finance, or security before proceeding. Do not give legal or financial advice. Just help me identify what questions to ask. Proposed change: [describe] Project context: [brief summary] Any known constraints: [list them] If the risks land in territory involving money, legal commitments, customer trust, or security, stop and escalate. AI flagging a risk is not the same as a qualified person reviewing it. Prompt 10: Run a weekly scope check Help me run a quick weekly scope health check on this project. Based on what I tell you, flag: any open items that haven't been formally scoped, any verbal agreements that haven't been documented, any changes that were informally approved but not through the change-control process, and any open questions that are currently blocking clarity on what's in or out. Then give me a prioritised list of what to resolve this week. Project name: [name] Current week: [week number or date] What's happened this week: [brief summary of conversations, requests, decisions] Any known open items: [list] Run this on Friday afternoon. It takes 10 minutes and saves the kind of mess that takes three weeks to untangle. What AI can't do here, and why that matters AI is fast at structure. It's good at drafting, comparing, listing, and prompting you to ask questions you might skip when you're busy. But it doesn't know what was actually agreed. It doesn't know the politics between your client and your delivery lead. It doesn't know that "nice to have" from a certain director means "do this or there will be consequences." Rule #7 in Don't Replace Me calls taste the moat. The human value in scope management isn't writing the change request form. It's knowing which changes are strategic, which are expensive to refuse, which will create downstream problems nobody's mapped yet, and which are genuinely fine to absorb. That judgment doesn't come from a prompt. It comes from knowing the project, the people, and the stakes. Use these prompts to do the structural work faster. Then apply the judgment that only you have. When to escalate instead of prompt Some scope situations aren't a documentation problem. They're a decision problem that needs human authority. Escalate to legal, finance, security, privacy, delivery leadership, client leadership, or executives when: A change affects contractual commitments or payment terms A change creates regulatory or compliance exposure A change significantly affects a client's end customers A change is being pushed through without formal approval A "verbal agreement" is being treated as binding scope No prompt handles those situations. A conversation with the right person does. For the risk side of this, the risk assessment prompts are worth reading alongside these. Frequently asked questions What are AI scope creep prompts? AI scope creep prompts are structured instructions you give to ChatGPT or Claude to help document, analyse, and communicate project scope changes. They're most useful for comparing new requests against original agreements, drafting change requests, estimating impact, and preparing stakeholder conversations. They don't replace human judgment or contract review. Can AI tell me if a request is in scope or out of scope? AI can compare a new request to the scope description you provide and flag what looks different. It can't tell you what was legally agreed, what a stakeholder actually said in a meeting, or what your contract says. Treat its output as a structured starting point, then verify against your actual documentation and with the people who made the original decisions. Is it safe to paste my project scope into ChatGPT? It depends on what's in it. Never paste customer PII, employee records, confidential pricing, financial forecasts, security vulnerabilities, unreleased product details, or legal contract terms into an unapproved AI tool. Summarise or anonymise where possible. If you're unsure, check with your IT or security team before using any AI tool for project documentation. How do I stop scope creep before it starts? Clear scope documentation at kickoff is the first line of defence. The project kickoff prompts and acceptance criteria prompts help you define what done looks like before you start, which makes out-of-scope requests much easier to identify and document later. What should I do when a stakeholder won't accept a scope pushback? Document the request formally, present options with clear tradeoffs, and escalate to the appropriate decision-maker if the informal pressure continues. Prompt 7 above helps you structure the options conversation. If the pushback involves contractual commitments or budget, escalate to legal, finance, or account leadership rather than trying to resolve it at the project level. Do I need a change-control process, or is this overkill for small projects? Even a simple project benefits from a lightweight version. The change-control process doesn't need to be bureaucratic. It can be as simple as: request in writing, one person reviews against original scope, decision documented. The prompts above work for that level too. What you're trying to avoid is informal approvals that nobody remembers and scope that grows by assumption. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Assumption Log Prompts: Stop Planning on Guesswork Use 10 AI assumption log prompts to expose project guesswork, design evidence checks, assign review owners, and stop unverified beliefs becoming commitments. AI promptsproject managementassumption logs AI Contingency Plan Prompts: 10 Templates 10 copy-paste AI contingency plan prompts for ChatGPT and Claude. Map failures, assign owners, draft comms, and verify assumptions before a crisis hits. contingency planningAI promptsbackup plans AI Dependency Mapping Prompts: Find What Breaks AI dependency mapping prompts for finding upstream work, downstream risks, hidden handoffs, owners, schedule conflicts, and launch blockers. AI promptsdependency mappingproject management --- ## AI Security Testing Prompts: Find Risk, Not Fiction URL: https://dontreplace.me/articles/ai-security-testing-prompts-find-risk-without-faking-proof AI promptssecurity testingquality assuranceapplication securitysoftware testing AI Security Testing Prompts: Find Risk, Not Fiction 10 AI security testing prompts for scope, threat models, authentication, authorization, APIs, secrets, dependencies, evidence, and human-reviewed findings. August 16, 2026 Updated August 16, 2026 Security testing gets dangerous when a confident checklist is mistaken for evidence. A model can produce forty plausible attack scenarios before lunch. It cannot tell you which controls exist on the deployed build, whether an exploit actually works, or whether anyone authorized the test. AI security testing prompts are useful for planning bounded, authorized work. They can organize scope, map assets and trust boundaries, suggest checks, expose unanswered questions, and turn sanitized observations into clearer findings. They cannot prove a system is secure, safely probe production by vibes, determine compliance, or approve a release. AI can draft the test map. Humans still authorize the work, operate real tools, collect evidence, judge impact, fix defects, and own the decision. Start with an AI test plan when the overall strategy is unclear. Use AI risk assessment prompts when the team needs to prioritize business impact. Security testing adds one hard rule: written authorization comes before curiosity. Why security testing needs authorization and evidence A security test can affect availability, data integrity, accounts, monitoring, vendors, and customers. Even a harmless-looking request can trigger defenses, corrupt test data, create noisy alerts, or cross a legal boundary. The first artifact is not a payload. It is an approved scope with named systems, environments, identities, windows, exclusions, contacts, and stop conditions. AI does not remove that obligation. Unless a model is explicitly connected to approved tools in an approved environment, it has not tested anything. Generated scanner output, imagined HTTP responses, and hypothetical vulnerabilities are fiction—not findings. Useful security coverage can include: assets, data flows, trust boundaries, roles, and abuse cases; authentication, recovery, enrollment, and multi-factor behavior; authorization across users, tenants, roles, objects, and administrative actions; input handling, file processing, rendering, and downstream interpreters; sessions, tokens, cookies, logout, expiry, rotation, and revocation; API schemas, object access, rate limits, replay, idempotency, and error behavior; secrets, logs, analytics, exports, backups, caches, and client-side exposure; dependencies, build artifacts, deployment settings, headers, and cloud configuration; monitoring, alerting, escalation, evidence retention, and remediation verification. The exact work depends on the product and threat model. A public marketing site, healthcare portal, internal payroll tool, and multi-tenant API do not share one magical checklist. Use AI API testing prompts for broader contract coverage, but keep security assertions tied to executed evidence. The reusable security testing prompt formula Use this base prompt, then add one specialized template below: “Act as a security test-planning assistant. I am testing [named system and flow] in [approved environment/build] under [written authorization reference]. Use only the supplied architecture, requirements, roles, data classifications, threat model, controls, and sanitized evidence. Produce [artifact] with scope, preconditions, safe actions, expected control traced to a source, evidence fields, cleanup, stop conditions, and open questions. Separate hypotheses, planned checks, observed facts, and decisions. Do not invent execution, vulnerabilities, exploitability, severity, compliance, or release approval.” That separation matters. A hypothesis says what might go wrong. A test says how an authorized person will check. An observation records what happened. A finding combines verified evidence with human impact analysis. Mixing those four stages creates security theater with nicer formatting. Never paste credentials, API keys, private source code, production data, customer PII, vulnerability details, logs, network diagrams, regulated information, or confidential client material into an unapproved AI tool. Use synthetic accounts, redacted evidence, approved sandboxes, placeholder secrets, and your organization’s security, privacy, legal, and vendor rules. What to collect before prompting “Hack our app” is not a brief. It is a future incident report. Give the model a narrow packet of verified, non-sensitive context. InputWhy it mattersHuman check Written authorization and ownerEstablishes permissionSecurity lead confirms approval Systems, builds, and environmentsBounds the targetOperators confirm exact assets Exclusions and stop conditionsPrevents unsafe testingIncident owner confirms triggers Roles and synthetic identitiesEnables boundary checksIdentity owner validates accounts Assets and data classificationsFocuses impact analysisData owner confirms sensitivity Architecture and trust boundariesReveals crossing pointsEngineers verify current design Expected controls and sourcesMakes results traceableControl owner confirms requirements Evidence and retention rulesProtects sensitive outputSecurity and privacy approve handling Keep unknowns visible. If nobody knows whether support staff may export customer records, the output should ask for a decision. It should not invent a policy and then congratulate the system for meeting it. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI security testing prompts Replace bracketed text with verified, sanitized information. These prompts produce planning artifacts. They do not authorize or execute tests. 1. Define scope and rules of engagement “Using this authorization, system inventory, environment list, test window, approved identities, exclusions, contacts, and incident procedure: [paste], draft rules of engagement. Include in-scope assets, prohibited actions, safe test limits, evidence handling, cleanup, rate limits, stop conditions, escalation contacts, and unresolved conflicts. Do not expand scope.” Have every accountable owner review the draft before testing. Name domains, APIs, cloud accounts, mobile builds, third-party services, and shared infrastructure precisely. “Staging” is not precise if three teams use different environments with that label. Define what happens when a tester sees unexpected production data, service degradation, active compromise, or a vulnerability outside scope. The correct next action may be stop and escalate—not keep digging because the finding looks spicy. 2. Map assets, data flows, and trust boundaries “Using this sanitized architecture, actor list, components, integrations, data classes, and deployment model: [paste], create a threat-oriented system map. List assets, entry points, trust boundaries, privileged operations, data stores, external dependencies, assumptions, and missing facts. Suggest abuse cases as hypotheses only and cite the supplied source for each element.” The map should show where identity, privilege, and sensitive data cross boundaries. Include background jobs, webhooks, exports, support tools, analytics, object storage, and admin consoles. The forgotten side door is still a door. Engineers must verify the map against the deployed system. Models happily connect plausible components that do not exist and omit boring infrastructure that does. 3. Plan authentication and account-recovery checks “For these approved authentication and recovery flows, identity providers, session rules, MFA policies, synthetic users, and expected controls: [paste], draft safe checks for enrollment, login, failure handling, recovery, verification, account enumeration, lockout, MFA changes, trusted devices, and reauthentication. Include expected evidence, cleanup, rate limits, and stop conditions.” Test normal and edge states: new accounts, disabled accounts, expired links, reused links, changed email addresses, lost authenticators, and concurrent sessions. Avoid denial-of-service behavior against real users or shared identity providers. Do not ask AI to invent working credentials or bypasses. Humans execute approved checks with synthetic identities and record exact results. 4. Test authorization and role boundaries “Using this role matrix, tenant model, object types, administrative actions, API operations, and synthetic accounts: [paste], create an authorization test matrix. Cover same-role object access, cross-role actions, cross-tenant access, direct object references, bulk operations, exports, hidden UI actions, background jobs, and privilege changes. Trace each expected denial or allowance to a source.” Authorization is not the same as hiding a button. Verify enforcement at the server or service boundary. Check read, create, update, delete, export, approve, impersonate, and administrative paths separately. Record the acting identity, target object, tenant, request, response, and resulting state. A 403 is useful evidence only if the protected action did not happen elsewhere. 5. Review input handling without reckless payloads “Given these approved inputs, parsers, file types, renderers, downstream services, validation rules, and sandbox limits: [paste], draft a safe input-handling test plan. Cover boundaries, encoding, canonicalization, malformed structures, file metadata, output encoding, error handling, and downstream interpretation. Use placeholders rather than operational exploit strings and flag specialist testing needs.” The planning model should identify input surfaces and transformations, not supply an unauthorized attack kit. Real specialists can choose appropriate tools and payloads inside the approved lab. Test the complete path. Input may be accepted safely at one layer and interpreted dangerously by a later renderer, queue worker, export, or integration. Evidence needs to identify where behavior changed. 6. Check sessions, tokens, and state transitions “Using these session requirements, token types, cookie settings, expiry rules, revocation behavior, sensitive actions, and supported clients: [paste], draft checks for creation, rotation, idle and absolute expiry, logout, password change, privilege change, replay, concurrent use, storage, transport, and reauthentication. Include safe setup and observable evidence.” A token’s format is less important than its lifecycle. Confirm what happens after logout, account disablement, password reset, role change, and suspected compromise. Check browser, mobile, API, and background-client behavior where supported. Never paste real tokens into an AI chat. Use placeholders in planning and approved evidence stores during execution. 7. Plan API abuse-case and resilience checks “For this API schema, actor model, object ownership, workflows, quotas, idempotency rules, webhook behavior, and approved sandbox: [paste], create security-focused test scenarios for object access, mass assignment, excessive data exposure, replay, pagination, filtering, rate limits, concurrency, webhook verification, and error leakage. Separate contract defects from security hypotheses.” API security lives in workflows, not only endpoints. A request can be valid in isolation and dangerous when repeated, reordered, raced, or performed by the wrong actor. Coordinate load-like checks. Rate-limit testing can affect shared systems and monitoring, so define ceilings and stop conditions before execution. Use AI exploratory testing prompts to vary sequences without abandoning evidence discipline. 8. Review secrets and sensitive-data exposure “Using this approved data-classification policy, secret-management design, logging standards, client bundles, exports, analytics, backups, and error formats: [paste], draft checks for exposed credentials, excessive responses, logs, URLs, browser storage, source maps, build artifacts, caches, notifications, exports, and support tooling. Specify redacted evidence fields and escalation rules.” Search approved artifacts and environments; do not copy discoveries into the prompt. If a real secret appears, follow incident handling immediately. Do not test whether it works unless the response plan explicitly authorizes that action. Sensitive exposure can be indirect: identifiers in URLs, tokens in referrers, personal data in analytics, verbose exceptions, or exports accessible longer than expected. Humans determine impact using actual classification and access context. 9. Examine dependencies and configuration “Using this sanitized dependency inventory, lockfiles, deployment model, security baselines, container or cloud configuration, and approved scan summaries: [paste], create a review plan. Cover provenance, unsupported versions, known advisories, reachable usage, build integrity, least privilege, network exposure, storage permissions, headers, debug modes, and exception handling. Do not fabricate versions or findings.” A dependency advisory is not automatically an exploitable product vulnerability. Humans must verify the installed version, affected feature, reachability, mitigations, and environment. Likewise, a clean scanner report does not prove safe configuration. Keep remediation practical. Upgrading a library can introduce behavior changes; tightening permissions can break jobs. Security owns risk guidance while engineering and operations verify the fix. 10. Turn sanitized observations into findings “Using only these sanitized observations from completed authorized testing: [paste], draft findings with affected asset and build, actor, preconditions, exact steps, observed result, expected-control source, evidence reference, user or business impact questions, reproducibility, cleanup, and remediation owner. Separate facts from interpretation. Do not invent severity, exploitability, compliance, or fixes.” A finding should be reproducible without leaking sensitive evidence into every ticket. Store raw captures in the approved system and reference them. Describe impact in terms of verified access, action, data, or availability—not a dramatic vulnerability name chosen by autocomplete. Humans confirm severity, ownership, disclosure, remediation, and retest. For cleaner defect mechanics, adapt these AI bug report prompts. A practical workflow that avoids security theater Get written authorization. Confirm owner, systems, environment, identities, window, and exclusions. Sanitize the context. Remove secrets, production data, private code, and identifying evidence before prompting. Map the system. Verify assets, actors, data, dependencies, and trust boundaries with engineers. Generate hypotheses and checks. Require traceability, safe limits, cleanup, and unanswered questions. Review the plan. Security, engineering, operations, privacy, legal, and product owners correct it. Execute with approved tools. Humans record exact commands, versions, times, identities, and observations. Escalate surprises. Stop when scope, safety, or incident conditions demand it. Write evidence-based findings. Separate observation from impact judgment and compliance interpretation. Fix and retest. Verify the original path and nearby variants without assuming one patch closed the class. This is the same bounded-work principle in the no-BS guide to using AI at work: delegate structure, verify output, and keep accountability attached to people. Common ways this goes wrong Treating generated scenarios as discovered vulnerabilities A scenario is a hypothesis. Until an authorized tester executes it and records evidence on the named build, no vulnerability has been found. Letting the model expand scope AI may suggest related domains, vendors, or infrastructure. Suggestions do not create permission. Put them in a future-scope list and obtain approval. Pasting secrets or findings into public tools Security evidence is often the sensitive material an attacker would want. Follow approved vendors, retention, access controls, redaction, and incident procedures. Asking for severity before impact is known Severity depends on verified access, affected data and users, prerequisites, reachability, controls, and business context. A familiar vulnerability label is not impact analysis. Confusing scanner output with safety Scanners are valuable and incomplete. False positives need validation; false negatives exist; workflow and authorization defects may never appear in a dependency or dynamic scan. Publishing AI-written fixes without review Generated changes can break authentication, create new bypasses, leak details through errors, or disable legitimate workflows. Review, test, deploy safely, and retest the original finding. If that boundary still feels fuzzy, read what AI can and cannot do. Fast pattern generation is useful. It is not authorization, execution, evidence, or accountability. Frequently asked questions Can AI perform security testing automatically? It can help operate approved automation when explicitly integrated, but a chat response has tested nothing. Even connected tools require scoped authorization, safe configuration, monitoring, evidence review, and expert interpretation. Are AI-generated security tests safe to run? Not automatically. Review every action against written scope, environment limits, data rules, rate limits, and stop conditions. Use approved sandboxes and synthetic identities. Never run generated commands against systems you do not own or have explicit permission to test. Can these prompts prove that an application is secure? No. Security is not a permanent pass state, and testing samples behavior under defined conditions. These prompts can improve planning and coverage. They cannot prove the absence of vulnerabilities or replace ongoing engineering, monitoring, review, and incident readiness. Should we paste source code into an AI tool for review? Only when your organization explicitly approves the tool, data handling, repository, and contract for that code. Prefer minimal sanitized excerpts when possible. Exclude secrets, customer data, private keys, vulnerability details, and unrelated proprietary code. Can AI assign vulnerability severity? It can organize inputs for an accountable reviewer. Humans must validate exploitability, affected assets, privileges, data, user impact, business context, compensating controls, and scoring policy before assigning severity. Does a clean automated scan mean the release is safe? No. Automated tools cover selected classes and configurations. They can miss business-logic flaws, authorization gaps, unsafe workflows, operational weaknesses, and novel interactions. Use layered testing and human review. Can AI replace a penetration tester or security engineer? No. It can accelerate inventory, scenario variation, traceability, and report formatting. Specialists contribute authorization discipline, tool operation, adversarial judgment, evidence validation, impact analysis, remediation guidance, and accountability. What information is safe to include in a prompt? Use only content approved by organizational and vendor policy. Prefer synthetic identities, generalized architecture, placeholder secrets, redacted findings, and non-production examples. Keep credentials, tokens, customer PII, private code, live vulnerabilities, logs, regulated data, and confidential diagrams out of unapproved systems. The point is better testing, not automated confidence AI security testing prompts are useful because security work contains a lot of structure: scope tables, actor matrices, state combinations, control mapping, evidence fields, and repeatable report formats. Let the model accelerate that structure. Keep the consequential work human: granting permission, operating tools, protecting evidence, recognizing dangerous behavior, judging impact, coordinating remediation, and deciding whether risk is acceptable. That is the practical thesis behind Don’t Replace Me by Dmitry Kargaev: use AI for speed without confusing speed for judgment. The goal is not a longer security checklist. It is fewer unexamined assumptions—and findings that are real enough to fix. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance AI API Testing Prompts: Catch Broken Integrations AI API testing prompts for contracts, edge cases, authorization, retries, error responses, and human-verified integration evidence. AI promptsAPI testingquality assurance AI Compatibility Testing Prompts: Catch Breakage 10 AI compatibility testing prompts for browser, device, OS, responsive, locale, network, defect triage, and evidence-based release decisions. AI promptscompatibility testingcross-browser testing --- ## AI Skills for Non-Technical People (2026) URL: https://dontreplace.me/articles/ai-skills-you-actually-need-in-2026-no-coding ai skillsnon-technicalprompt engineeringAI at workfuture-proof career AI Skills for Non-Technical People (2026) The AI skills non-technical people actually need in 2026. No coding, no bootcamp, no $997 course. Four real skills and a minimum viable stack by profession. May 10, 2026 Updated 12 May 2026 There's a $997 course out there promising to teach you "Advanced AI Prompt Engineering for the Modern Professional." The curriculum includes things like "chain-of-thought scaffolding" and "semantic temperature calibration." The guy selling it has 80,000 LinkedIn followers and hasn't held a real job since 2022. The AI skills for non-technical people look nothing like that. Most of them you already have. Some of them you've been practicing for decades without knowing there was a name for it. Let's be specific about what's real. What "learning AI" actually means for non-technical people Here's the thing nobody making money from AI courses wants you to know: there's no discipline called "AI" that you need to study. There are tools. Tools you use for specific tasks in your specific job. That's it. You didn't "learn email" in 1998. You learned to use Outlook for the things email was good at. This is the same thing. Dmitry Kargaev puts it plainly in Don't Replace Me: "AI is just your keyboard now." The keyboard doesn't require a certification. It requires you to sit down and type something. The anxiety around "learning AI" is mostly manufactured by people who sell learning AI. The actual skill gap between where you are and where you need to be is much smaller than the $997 price tag suggests. There are four real skills. None of them require a coding bootcamp. AI skills you actually need: skill 1, clear communication This is the big one. And it's the one the prompt engineering crowd has dressed up in so much jargon that normal people assume it's technical. Prompting is talking clearly. That's it. If you can write a clear email to a colleague explaining what you need and why, you can prompt an AI. The only difference is the AI won't ask clarifying questions back (unless you tell it to), so you need to front-load a bit more context. "Write a marketing email" is a bad prompt not because it violates some prompt engineering principle but because it's vague. You'd never send that as a brief to a copywriter. "Write a 200-word marketing email for our mid-market HR software product, targeting HR directors at companies with 200-500 employees, for a campaign about reducing time-to-hire" is a good prompt because it's clear. You already know how to be clear when you need to be. The skill is just applying that to a new tool. If your outputs are bad, 90% of the time the problem is in your input. The practical guide to using AI at work covers this in detail, but the short version is: give it context, give it a format, give it the constraints. AI skills you actually need: skill 2, critical evaluation AI makes things up. Not occasionally. Regularly. Confidently. With the exact tone of someone who definitely knows what they're talking about. The WEF's 2025 Future of Jobs report lists "analytical thinking" as the most in-demand skill for the next five years. That's not a coincidence. It's a direct response to a world full of plausible-sounding AI output that needs a human to check it before it goes anywhere. This is not a new skill. You've been evaluating information for your entire career. You cross-check figures before a board presentation. You read a contract before signing it. You don't just forward every email you receive. The specific application to AI: check numbers, check citations, check anything that sounds suspiciously convenient for the argument the AI is making. AI is very good at finding evidence for whatever you ask it to prove. That's a feature when you want to brainstorm. It's a problem when you assume the output is neutral research. Build the habit of treating AI output the way you'd treat a first draft from a smart but overconfident intern. Useful starting point. Not something you submit without reading. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → AI skills you actually need: skill 3, workflow redesign This one takes a little more thinking, but it's the skill that actually separates people who get a 20% productivity bump from people who get a 2x productivity bump. Most people who start using AI bolt it onto their existing workflow. They do all the same steps they always did, then occasionally ask ChatGPT to clean up the final paragraph. That's like when factories first got electricity and just replaced the single steam engine with a single electric motor. Same layout. Same process. Just different power source. It took a generation to realize you could put small motors everywhere and redesign the whole factory from scratch. That's when the real productivity gains showed up. Same thing is happening right now. The people winning aren't the ones using AI the most. They're the ones who looked at their workflows and asked: which of these steps only existed because a previous step was slow or expensive? If you had a free, instant research assistant, would you still structure your process the same way? Concretely: which parts of your work are repetitive and pattern-based? Those are AI tasks. Which parts require your specific judgment, relationships, or context? Those stay human. How to future-proof your career against AI goes deeper on this framework if you want the strategic version. AI skills you actually need: skill 4, knowing when to stop This one gets skipped in every "AI productivity" article, and it's probably the most important. AI produces output fast. Fast is seductive. It creates the illusion of progress even when what it's producing is mediocre. The people who use AI badly are often the people who use it too much, not too little. There are categories of work where AI makes things worse: highly sensitive client communications where the human voice matters, creative work where distinctiveness is the whole point, analysis where the person doing the analysis needs to actually understand the subject, and decisions where accountability can't be handed off to a tool. Knowing when to put the tool down is an AI skill. It requires taste. It requires a clear-eyed assessment of what the output is actually worth. What AI can and can't do gives you a sharper framework for drawing that line. The minimum viable AI stack for non-technical professionals Forget the 50-tool listicles. Here's what actually matters, by job type: ProfessionTool 1Tool 2What you use them for MarketerChatGPT / ClaudeCanva AIDrafts, briefs, copy variations / visuals Finance professionalChatGPTExcel CopilotSummarizing reports, explaining data / formulas and analysis HR / People opsChatGPTLinkedIn AI featuresJob descriptions, policy drafts / candidate research Lawyer / LegalClaudeYour firm's AI toolsDocument review, clause research, first draft summaries Project managerChatGPTNotion AIMeeting summaries, status updates, communication drafts Writer / EditorClaudeGrammarly AIResearch, structural feedback, editing passes DesignerChatGPTMidjourney / Adobe FireflyBriefs, concept exploration / image generation Consultant / AnalystClaudePerplexityResearch synthesis, report drafts / real-time sourced research Two tools. Maybe three. That's the stack. Rule #14 in the book puts it directly: pick two or three, master those. Every hour you spend evaluating a new tool is an hour you're not spending getting good at the ones you already have. Real prompt templates for emails, reports, and presentations gives you the practical starting point if you've already picked your stack and want to stop staring at a blank prompt box. What you don't need (and who's selling it) Prompt engineering certification: dead. The models have gotten good enough at understanding natural language that the elaborate prompt syntax people were selling in 2023 is largely irrelevant. If you need a certification to talk to a language model, the problem isn't you. A complete understanding of how LLMs work: you don't need to understand the internal combustion engine to drive a car. You don't need to understand transformer architecture to use ChatGPT. The McKinsey 2024 State of AI report found that the biggest barrier to AI adoption in organizations isn't technical skill gaps. It's change management and workflow integration. The technical stuff gets figured out. The human stuff is harder. Seventeen different AI tools: According to a 2024 survey by Slack, 77% of workers want AI help at work but most haven't settled on a consistent tool they use regularly. The answer isn't more tools. It's actually using the two you have. A course that costs more than $50: If you're paying more than that, you're paying for a community and a coach, not for information. The information is free. OpenAI's own documentation is free. There are free YouTube tutorials for every specific use case you can imagine. The $997 price tag is not a signal of quality. It's a signal of a good marketing funnel. How to actually build these AI skills in the next two weeks Stop reading articles about AI. Start using it for something you actually have to do this week. Take the most annoying, repetitive task on your plate right now. The one that takes two hours and makes you feel like a human photocopier. Draft a prompt. Run it. See what you get. Edit the output. Notice what was wrong with your prompt. Fix it. Run it again. That process, repeated across different tasks over two weeks, will teach you more than any course. You'll develop an intuition for what AI is good at and where it falls flat. You'll build the habit of checking outputs critically. You'll start to see which parts of your workflow are actually candidates for redesign. The Goldman Sachs AI research that got everyone's attention in 2023 estimated 26% of tasks in most knowledge worker roles could be automated. That's not your whole job. That's roughly one day a week of time that could be freed up. You don't need to become technical to claim it. You need to be willing to try something specific. The barrier is lower than you think. The tools are better than they were. The grifters are louder than ever, but you don't have to listen to them. Frequently asked questions Do I need to learn coding to use AI tools at work? No. The main AI tools non-technical workers use, including ChatGPT, Claude, Perplexity, and Copilot, require no coding at all. The real skill is clear communication: giving the tool enough context to produce useful output. If you can write a clear email, you can prompt an AI. What are the most important AI skills for non-technical people? Critical evaluation of AI output is probably the most valuable single skill. AI produces confident, plausible-sounding content that is sometimes wrong. Being able to read AI output the same way you'd review a junior colleague's first draft, checking for errors, bias, and gaps, is what separates people who use AI well from people who get burned by it. Is prompt engineering a real skill worth learning in 2026? The version of prompt engineering that involves elaborate syntax and paid certifications is largely outdated. Modern LLMs are much better at understanding natural language than they were in 2022. What matters now is clear communication: providing context, format, and constraints. You don't need a course to learn that. How many AI tools do I actually need? Two or three, used consistently, will outperform 20 tools used occasionally. Pick the tools that map to your most time-consuming tasks and learn them well before adding anything else. For most knowledge workers, that's ChatGPT or Claude plus one domain-specific tool (Excel Copilot, Canva AI, Notion AI, etc.). How long does it take to get useful at AI tools at work? Most people see real productivity gains within two to four weeks of consistent daily use. The learning curve is not steep. The barrier is usually starting, not the tools themselves. Pick one annoying task, automate part of it, and iterate from there. Should I be worried I'm falling behind if I haven't learned AI yet? Worried, no. Aware, yes. The WEF 2025 Future of Jobs report projects that 39% of core work skills will change by 2030. But "change" doesn't mean replacement. It means tools get incorporated into existing workflows. The people who get left behind aren't the ones who start late. They're the ones who decide not to start at all. See what AI can and can't do for a clearer picture of what actually needs your attention. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading How to Future-Proof Your Career Against AI The combination moat framework: build the career AI can't replace using domain expertise, AI fluency, and relationships. No coding required. future-proof careerAI career strategycombination moat The $997 AI Masterclass Is a Scam The AI masterclass scam is a $5B industry built on fear. Here's how to spot the grifters and actually learn AI for free. AI griftersAI coursesprompt engineering AI Anxiety at Work: How to Stop Spiraling 65% of employees feel AI anxiety at work. Here's how to convert that fear into something useful without buying a $997 course or losing more sleep. ai anxietyai job fearwill ai replace my job --- ## AI Smoke Test Prompts: Catch Launch Bugs Before Users Do URL: https://dontreplace.me/articles/ai-smoke-test-prompts-catch-launch-bugs-before-users-do ai smoke testlaunch testingQA promptsrelease readinessbug reporting AI Smoke Test Prompts: Catch Launch Bugs Before Users Do 10 copy-paste AI smoke test prompts for launch checklists, critical paths, bug summaries, and retest plans. With safety rules for what not to paste. July 30, 2026 Updated July 30, 2026 The worst launch bugs are never mysterious. They are the login button nobody clicked after deploy. The pricing page that loads fine on desktop and looks like lasagna on mobile. The confirmation email that stopped sending because one environment variable was wrong. The dashboard that works for admins and breaks for normal users, which is a cute little detail if your customers are not admins. AI smoke test prompts help with that boring, obvious, career-saving layer of launch testing. Not because AI can test your product for you. It can't. But it can turn scattered release notes, known risks, acceptance criteria, and user flows into a checklist a tired human can actually use before users become your QA department. If you need deeper QA coverage, start with AI QA checklist prompts. If you are testing whether users can accept the thing, use AI user acceptance testing prompts. This article is narrower. Smoke tests are the quick critical-path checks before and immediately after something goes live. AI can structure the smoke test. A human still has to click, verify, judge, and sign off. What AI smoke test prompts are actually good for AI is useful when your launch notes are a junk drawer. Release scope in one doc. Acceptance criteria in tickets. Risks in Slack. Customer-impact notes in someone's head. A launch checklist from last quarter that still says to test Internet Explorer, because apparently your process has ghosts. AI can take that mess and produce a first-pass smoke test plan. It can group checks by user journey, flag missing owners, draft tester instructions, summarize defects, and create retest steps. That is genuinely helpful. What it cannot do is verify production health, inspect private systems it cannot access, know which customer flow pays the bills, or tell whether a bug is acceptable risk. It will happily produce a tidy checklist with a missing checkout flow if your notes never mention checkout. That is not intelligence. That is autocomplete wearing a clipboard. This is the same split explained in what AI can and can't do: use AI for structure, speed, and second-pass gap finding. Do not use it as a replacement for judgment. The reusable AI smoke test prompt formula Use this structure for almost every smoke test prompt: "You are a smoke test planning assistant. I am [your role] preparing to test [release/change/project]. The launch affects [users/systems/pages/workflows]. Here is the sanitized scope, acceptance criteria, known risks, and notes: [paste details]. Create [specific output]. Flag missing information. Do not invent results, owners, URLs, approvals, or test evidence." That last sentence matters. AI should flag gaps, not cosplay as someone who verified them. Before you paste anything, sanitize it. No customer PII. No credentials. No access tokens. No raw production logs. No private tickets. No unreleased strategy. No legal issues. No security vulnerabilities. No financial records. No HR details. No regulated data. No medical information. No private client conversations. Use approved tools and approved data. If you are not sure whether something is safe to paste, it is not safe to paste. Congratulations, you found one of the few easy rules in launch work. What to bring before you ask AI for a smoke test The prompt works better when you bring real launch material instead of asking AI to hallucinate a testing strategy from vibes. Here is the minimum useful input set: InputWhy it mattersHuman check Release scopeKeeps the checklist tied to what changedConfirm the scope matches the actual deploy Critical user journeysPrevents testing only the easy pathsRank by customer and business impact Acceptance criteriaGives expected results something concrete to referenceVerify criteria are current, not stale ticket archaeology Known risksPoints testing at the scary partsAsk engineering, support, and product what worries them EnvironmentsAvoids mixing staging, preview, and production evidenceLabel every result with where it was tested Test data rulesPrevents unsafe copy-paste nonsenseUse approved fake data or approved masked data OwnersMakes follow-up possibleReplace team names with actual humans Rollback triggersConnects testing to launch decisionsConfirm who can call stop That table is not ceremony. It is the difference between a smoke test and a scavenger hunt. If you only give AI the release title, you will get a generic checklist. Login. Submit form. Check email. Very impressive, if your product is a tutorial app from 2011. Give it the real sanitized context instead. Then ask it to point at gaps. A missing owner, unclear expected result, unsupported browser, untested permission level, or absent rollback trigger is not a formatting problem. It is launch risk wearing a tiny hat. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI smoke test prompts you can use today Prompt 1: Turn release notes into smoke test coverage You are a smoke test planning assistant. I am preparing smoke test coverage for this release. Here are the sanitized release notes, acceptance criteria, and known risks: [paste details]. Create a smoke test checklist grouped by user journey. For each check, include: purpose, steps, expected result, owner placeholder, environment, and priority. Flag any release item that has no obvious smoke test. What to check after: Make sure every launch-critical item has a test. AI can map notes to checks. It cannot know the feature your revenue depends on if nobody wrote it down. Prompt 2: Identify the critical user journeys You are a smoke test planning assistant. Here is the product, workflow, or site change we are launching: [sanitized description]. The affected users are: [user groups]. Please identify the critical user journeys we should smoke test before and after launch. Sort them by business impact and user impact. Flag any assumptions you are making. What to check after: The ranking is a draft. A broken admin export might matter less than a broken signup flow. Or it might matter more because one enterprise customer lives inside that export. Human context wins. Prompt 3: Create a pre-launch smoke test checklist You are a smoke test planning assistant. Build a pre-launch smoke test checklist for [release/change]. Scope: [sanitized scope]. Known risks: [risk list]. Environments available: [staging/pre-prod/etc.]. Include checks for core flows, permissions, forms, notifications, integrations, analytics, mobile/responsive behavior, and error states. Mark anything that must be verified by a human. What to check after: Pre-launch checks are only as good as the environment. If staging is not production-like, write that down. A staging pass is useful evidence, not a blood oath. Prompt 4: Create a post-launch smoke test checklist You are a smoke test planning assistant. Create a post-launch smoke test checklist for the first 60 minutes after [release/change] goes live. Include checks for live URLs, login/access, key workflows, payment or submission paths if relevant, notifications, integrations, monitoring dashboards, support channels, and rollback signals. Use placeholders where real links or owners are missing. What to check after: Replace every placeholder before launch. A checklist with placeholder dashboard links is a decorative object, not an operating tool. For the broader launch list, see AI launch checklist prompts. Prompt 5: Test expected vs. actual behavior You are a smoke test planning assistant. Here are the smoke test steps we ran and the observed results: [paste sanitized steps and results]. Compare expected vs. actual behavior. Group findings into: pass, fail, unclear, and needs retest. Do not mark ambiguous results as passed. Flag missing screenshots, dates, owners, or environment details. What to check after: This is where AI is useful as a documentation nag. It can spot vague results like "looks good" or "seems fine." It cannot confirm what happened. If there is no evidence, there is no evidence. Prompt 6: Check browser, device, and environment coverage You are a smoke test planning assistant. Here is our planned smoke test coverage: [paste sanitized checklist]. Review it for browser, device, role, permission, and environment gaps. Identify which checks should run on desktop, mobile, tablet, admin role, standard user role, and any other relevant user role. Do not add unsupported requirements. Flag what needs product or engineering confirmation. What to check after: Do not let AI create fake requirements. If the product officially supports Safari, test Safari. If it does not, clarify that with whoever owns the decision. Do not discover browser policy through vibes. Prompt 7: Draft tester instructions You are a smoke test planning assistant. Draft clear tester instructions for this smoke test: [paste sanitized scope and checklist]. Include: objective, environment, test data rules, steps, expected result format, how to report defects, severity definitions, and deadline. Keep it practical and short enough that a busy teammate will read it. What to check after: Remove anything that implies testers should use real customer data, shared passwords, or unsafe screenshots. The National Institute of Standards and Technology has boring but useful guidance on protecting sensitive data in testing contexts, and boring is exactly what you want here. Prompt 8: Summarize defects for engineers You are a smoke test planning assistant. Here are sanitized smoke test findings and bug notes: [paste findings]. Convert them into engineer-ready defect summaries. For each defect include: title, affected area, steps to reproduce, expected behavior, actual behavior, severity, evidence available, environment, and open questions. Do not invent reproduction steps or evidence. What to check after: A good defect summary saves time. A fake one burns it. If AI creates a step you did not perform, delete it. If severity is unclear, mark it unclear and ask the right human. This pairs nicely with AI bug report prompts when the findings are too messy for a ticket. Prompt 9: Build a retest plan You are a smoke test planning assistant. We found and fixed the following launch issues: [paste sanitized issue list]. Create a retest plan that maps each fix to the checks that must be rerun, the owner, the environment, expected result, and evidence needed. Flag any related flows that might need regression checks. What to check after: Retesting only the exact failing step is how bugs grow side quests. Check adjacent flows. If the login fix touched authentication, test login, logout, password reset, role permissions, and anything else that smells connected. Prompt 10: Convert smoke test gaps into action items You are a smoke test planning assistant. Here are the gaps and open questions from our smoke test planning review: [paste sanitized gap list]. Convert them into action items with owner placeholder, due date, blocker status, severity, and recommended next step. Sort blockers first. Flag anything that requires release owner approval before launch. What to check after: Placeholder owners are not owners. Due dates like "before launch" are not due dates. Turn the list into real assignments or admit you are launching with unresolved risk. How smoke testing fits into launch readiness Smoke testing is not the whole launch process. It is the last practical sanity check before the room starts clapping too early. You still need requirements, acceptance criteria, QA coverage, operational readiness, monitoring, rollback criteria, support coverage, and a real go/no-go decision. The smoke test sits near the end of that chain. It asks a simple question: does the important stuff still work well enough to proceed? If you are still organizing the switch itself, use AI cutover plan prompts. If you are checking whether the team is ready to operate the thing, use AI operational readiness prompts. If the question is whether to launch at all, use AI go/no-go decision prompts. The point is not to create more documents. The point is to make the risk visible while there is still time to do something besides typing "circling back" in Slack with the energy of a hostage note. Dmitry Kargaev talks about this in Don't Replace Me: the moat is not producing artifacts faster. The moat is taste. Knowing which artifact matters, which result is fake comfort, and which missing detail is about to bite you. What AI must not decide in smoke testing AI should not decide whether a launch is safe. It should not approve a defect as acceptable. It should not downgrade a severity because the summary sounds calm. It should not decide that missing evidence is fine. It should not infer that a flow passed because no one mentioned it failing. That is how you get a beautiful report and an ugly morning. Use AI to find gaps, rewrite messy notes, structure test evidence, and make the next action obvious. Then make the human decisions out loud. Who owns the risk? What evidence exists? What did not get tested? What is the rollback trigger? Who can call stop? If nobody can answer those questions, your smoke test did its job. It found the smoke. One more practical rule: write down what you did not test. Teams hate this because it makes the risk visible. Good. Visible risk is useful. Hidden risk is how you end up apologizing to customers with the phrase "edge case" when the edge case was actually every user on mobile. A smoke test report should say three things clearly: What passed, with evidence and environment. What failed or looked weird, with owners and next steps. What was not tested, with the reason and the person accepting that risk. That last bullet is where grown-up launch work happens. AI can help draft it. A human has to own it. Frequently asked questions Can AI run smoke tests for me? AI can help create smoke test checklists, tester instructions, defect summaries, and retest plans. It cannot verify your live system unless it has approved access and a safe automation setup. For normal teams, treat AI as a planning and documentation assistant. Humans still run and verify the checks. What should I include in an AI smoke test prompt? Include sanitized release scope, affected users, critical journeys, acceptance criteria, known risks, environments, browser or device requirements, test data rules, owners, monitoring links, rollback criteria, and the exact output you want. Tell the AI to flag gaps and not invent results. What should I never paste into AI while smoke testing? Do not paste customer PII, credentials, access tokens, raw production logs, private tickets, unreleased strategy, security vulnerabilities, financial records, HR details, regulated data, medical information, private client conversations, or anything your company has not approved for AI tools. How many smoke tests do I need before launch? Enough to cover the critical paths users and the business depend on. That usually means login, permissions, key workflows, forms, notifications, integrations, payment or submission paths if relevant, monitoring, support handoffs, and rollback signals. The number matters less than whether the important flows are covered. What is the difference between smoke testing and UAT? User acceptance testing checks whether the work meets user needs and acceptance criteria before release. Smoke testing is a faster sanity check around launch to confirm the most important flows still work. UAT asks, "Did we build the right thing?" Smoke testing asks, "Did we break the obvious thing?" Can AI summarize smoke test defects for engineers? Yes, and this is one of the safer uses. Paste sanitized findings and ask AI to structure each defect with steps to reproduce, expected behavior, actual behavior, severity, environment, evidence, and open questions. Then delete anything it invented and verify the final ticket before sending it. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Bug Report Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn vague bug complaints into structured tickets. Templates for triage, handoffs, regression checklists, and more. ai bug report promptsbug reportingQA AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management --- ## AI SOP Prompts: Turn Tribal Knowledge Into Repeatable Processes URL: https://dontreplace.me/articles/ai-sop-prompts-turn-tribal-knowledge-into-repeatable-processes AI SOP promptsdocumentationtribal knowledgeprocess documentationoperations AI SOP Prompts: Turn Tribal Knowledge Into Repeatable Processes 10 copy-paste AI SOP prompts to turn messy notes and tribal knowledge into structured SOPs. Includes safety rules and human review guidance. June 24, 2026 Updated June 24, 2026 Your most experienced person just gave two weeks notice. And everything they know about how the actual work gets done is sitting in their head, their inbox, and a Slack thread from 2022 that nobody can find. This is the problem AI SOP prompts were made for. Not replacing the expertise. Just getting it out of one person's skull and into a document before it walks out the door. Standard operating procedures are where most operations documentation goes to die. They take forever to write, they go stale instantly, and the people who need them most hate writing them. AI doesn't fix all of that. But it can cut the draft time down dramatically, ask the questions you forgot to ask, and turn a rambling brain dump into something that looks like a process. Here's how to do it without making a bigger mess than you started with. What AI SOP prompts can actually do (and what they can't) AI is fast. That's the whole thing. It can take your messy notes from an onboarding call, a Loom transcript, or a half-finished Google Doc and turn them into a structured first draft in under a minute. It's genuinely useful for scaffolding: headings, checklists, role-specific versions, quality criteria, exception documentation. What it cannot do is know your business. It doesn't know that the vendor portal times out after 10 minutes and you have to re-authenticate. It doesn't know that the finance team actually approves this, not the ops manager, even though the org chart says otherwise. It doesn't know which steps the outgoing person skipped because they had 10 years of context and didn't need them. Rule #5 in Don't Replace Me says it plainly: it's not smart, it's fast. Speed is not operational truth. The SOP draft AI hands you will look clean and organized. That doesn't mean it's accurate. Your job is to supply the reality. AI's job is to structure it. Before you paste anything: what not to put in an AI tool Before the prompts, a hard stop on data safety. Do not paste any of the following into a general AI tool like ChatGPT or Claude unless your organization has explicitly approved that tool for sensitive data: Customer names, emails, or personal information Employee records or HR-related content Login credentials or access permissions of any kind Legal disputes, contracts, or anything in active litigation Security incidents or vulnerability details Unreleased product details or confidential strategy Private financial information or board materials Sensitive client information or customer commitments If you're not sure whether your company has approved a particular AI tool for business use, ask before you paste. Check your AI workplace policy first. This is not paranoia. It's just the actual rules of using tools responsibly. Anonymize your notes before using them. Replace names with roles. Replace system names with generics if needed. The prompt still works. The reusable AI SOP prompt formula Every good SOP prompt follows the same structure. Memorize this and you can build your own: Role (who's writing this) + Context (what process, what system, what team) + Source material (your notes, transcript, or existing doc) + Output format (what you want back) + Constraints (what AI must not invent) That last part is the one people skip. Always tell the AI what it's not allowed to make up. Otherwise it will fill gaps with plausible-sounding nonsense, and plausible-sounding nonsense in an SOP is genuinely dangerous. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI SOP prompts you can copy and use today Prompt 1: Turn messy notes into a first SOP draft You are an operations writer. I'm going to give you rough notes about a workflow our team runs. Turn them into a structured SOP draft with: a one-paragraph process overview, numbered steps, the role responsible for each step, and a "common mistakes" section at the end. Do not invent steps, owners, system names, or approval paths. If something is unclear in my notes, flag it with [NEEDS CLARIFICATION] instead of guessing. Here are my notes: [PASTE YOUR NOTES HERE] Prompt 2: Extract steps from a transcript I have a transcript from an interview with a team member explaining how they run [PROCESS NAME]. Extract every distinct action step they mention and list them in order. Number them. Where the speaker jumps around, reorganize chronologically. Mark any step that seems implied but wasn't stated explicitly with [ASSUMED - VERIFY]. Do not add steps that weren't in the transcript. Transcript: [PASTE ANONYMIZED TRANSCRIPT HERE] Prompt 3: Find the missing context I have a draft SOP for [PROCESS NAME]. Read it and give me a list of questions I need to answer before this document is safe to publish. Focus on: missing owners, unclear triggers ("when does this process start?"), undefined terms, absent approval steps, unstated assumptions, and steps that could go wrong without more detail. Format your output as a numbered list of specific questions. SOP draft: [PASTE DRAFT HERE] Prompt 4: Create a checklist version Take this SOP and produce two things: 1. A one-page quick-reference checklist for someone doing this process for the first time. 2. A shorter checklist for someone who knows the process and just needs a reminder of the sequence. Use checkboxes. Keep language action-oriented. Do not add steps that aren't in the source SOP. SOP: [PASTE SOP HERE] Prompt 5: Document exceptions I'm going to describe the normal version of our [PROCESS NAME] SOP and also give you notes on common exceptions we handle. Create an "Exceptions and Edge Cases" section that lists each exception, what triggers it, how it differs from the standard process, and who needs to be involved. Flag any exception where you're missing information about the correct resolution with [ESCALATION PATH UNCLEAR]. Normal process notes: [PASTE NOTES] Exception notes: [PASTE EXCEPTION NOTES] Prompt 6: Define "done" and quality checks Based on this SOP draft, write a "Definition of Done" section and a "Quality Checks" section. Definition of Done: what does a completed, correct instance of this process look like? What can the person check to confirm they're finished? Quality Checks: what should a reviewer look for when auditing this work? What are the most common errors? Do not invent quality criteria. If the source material doesn't contain enough information to define a quality check, say so. SOP draft: [PASTE HERE] Prompt 7: Create a role-specific version I have a full team SOP for [PROCESS NAME]. Create a role-specific version for [ROLE NAME] only. Include only the steps this role is responsible for, with clear handoff points to other roles marked as [HANDOFF TO: ROLE NAME]. Add a brief "What you're responsible for" intro paragraph and a "What happens before/after your part" section. Full SOP: [PASTE HERE] Prompt 8: Turn an SOP into onboarding material Take this SOP and turn it into an onboarding guide for someone joining the team and running this process for the first time in their first 30 days. Include: a plain-English "why this process exists" intro, the steps in beginner-friendly language, tips for common mistakes, and a "who to ask if you're stuck" section (I'll fill in the names, just leave [NAME/ROLE] placeholders). Do not soften or skip steps to make it seem simpler. Accuracy matters more than friendliness. SOP: [PASTE HERE] These onboarding-ready outputs pair well with the AI onboarding prompts and AI training prompts we've covered separately. They're different enough to be worth treating as distinct workflows. Prompt 9: Review an existing SOP for risk You are an operations auditor. Review this SOP for risks. Look for: steps with no named owner, approval paths that are vague or missing, compliance or legal language that wasn't reviewed by a qualified person, steps that assume access or permissions without specifying how to get them, and outdated system references. Output a risk log. For each risk, note the specific section, the type of risk, and what action is needed to resolve it. Do not provide legal, compliance, or security guidance. Flag those areas for human expert review. SOP: [PASTE HERE] Prompt 10: Make a lightweight update log entry I made the following changes to our [PROCESS NAME] SOP: [DESCRIBE WHAT CHANGED AND WHY] Write a concise update log entry for this change. Include: date placeholder ([DATE]), what changed, why it changed, and who approved the change ([APPROVER NAME] placeholder). Keep it under 100 words. Use past tense. This is a record, not an announcement. What to do after AI drafts your SOP The draft is not the SOP. This is the mistake people make. They get a clean, well-formatted document back from Claude or ChatGPT and think the work is done. It isn't. Before any AI-drafted SOP becomes official documentation, someone who actually does the work needs to read it step by step and verify it's real. Not skim it. Read it like a new hire would follow it on their first day. Walk through the actual system. Check that the access paths, tool names, and approval flows are accurate. Then you need an owner. Not "the team." One person whose name is on the document and who is responsible for keeping it current. Set a review cycle. Quarterly for high-frequency processes, annually for stable ones. And store the SOP in your actual source-of-truth system, whether that's Notion, Confluence, SharePoint, or wherever your team actually looks things up. A great doc that lives in someone's Google Drive is just elaborate folklore. An AI-generated SOP that nobody validated is just polished chaos with better formatting. For anything touching compliance, security, legal commitments, or customer-facing guarantees, human expert review is not optional. The prompts above will flag these areas. Do not skip those flags. If you're looking for how this connects to broader documentation work, the AI documentation prompts article covers the wider set. And if you're trying to fix broken workflows before documenting them, start with the AI workflow audit prompts first. Documenting a broken process just makes the broken process official. The thing AI will never give you Taste. The ability to look at a step in a process and know it's the one that will quietly burn the building down if someone skips it. The judgment to know which exceptions are genuinely rare and which ones happen every Tuesday. The institutional memory of why the process was built the way it was in the first place. That's the stuff that lives in the head of your person who's leaving. AI can help you get it out of their head faster. It cannot replace it once they're gone. The prompts above are tools for extraction and structure. The validation, the context, the accountability, and the taste: that's still yours. Frequently asked questions Can I use ChatGPT to write SOPs for my company? Yes, with clear limits. ChatGPT is useful for drafting structure, checklists, role-specific versions, and exception documentation. It cannot verify that the steps it generates are accurate for your actual systems and workflows. Every AI-drafted SOP needs human validation before it goes live. Sensitive data including credentials, PII, confidential strategy, and legal content should never be pasted into a general AI tool. How do I make sure an AI-written SOP is accurate? Have someone who actually does the process walk through the draft step by step, in the real system, not just by reading it. AI drafts look polished. That's the danger. Accuracy comes from verification against reality, not from formatting. Use Prompt 3 above to surface missing context before you do that review. What's the difference between an SOP and a process doc? They're often used interchangeably, but an SOP (Standard Operating Procedure) typically implies a higher bar: a specific output standard, a named owner, approval authority, a version date, and a review cycle. A process doc can be informal. An SOP is the official version. AI documentation prompts can help with both. How do I document tribal knowledge before someone leaves? Record an interview or a Loom walkthrough of the process. Get the person to narrate what they're doing and why. Then use Prompt 2 (extract steps from a transcript) to pull the structure. Follow up with Prompt 3 to find what's missing. This won't capture everything, but it captures far more than asking someone to write it down themselves. Should I use AI to update existing SOPs? Yes, carefully. Prompt 10 above gives you a lightweight update log format. Prompt 9 will audit an existing SOP for stale or risky content. The risk with AI-assisted updates is the same as with new drafts: it can generate plausible-sounding updates that don't reflect how the process actually works now. Always validate against current reality before publishing an update. What should I never let AI decide in an SOP? Don't let AI invent: process owners, approval authority, access permissions, compliance requirements, legal obligations, customer commitments, system behavior, security controls, or anything with regulatory implications. AI can suggest placeholders for all of these. The actual values need to come from the people with the authority to define them. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Change Impact Analysis Prompts: Spot Damage AI change impact analysis prompts for mapping affected systems, teams, customers, QA checks, support gaps, and rollout risks. AI promptschange impactrisk AI Dependency Mapping Prompts: Find What Breaks AI dependency mapping prompts for finding upstream work, downstream risks, hidden handoffs, owners, schedule conflicts, and launch blockers. AI promptsdependency mappingproject management AI Dependency Register Prompts: Keep Projects From Surprising You Use AI dependency register prompts to map owners, blockers, assumptions, and downstream impact before one small change becomes a three-week incident. AI dependency mappingproject managementAI prompts --- ## AI Spreadsheet Prompts: 10 Copy-Paste Templates URL: https://dontreplace.me/articles/ai-spreadsheet-prompts-fix-excel-and-google-sheets ai spreadsheet promptsexcel AIgoogle sheets AIChatGPT formulasspreadsheet help AI Spreadsheet Prompts: 10 Copy-Paste Templates 10 copy-paste AI spreadsheet prompts for ChatGPT and Claude. Fix formulas, clean data, debug errors, and build pivot tables without the guesswork. May 29, 2026 Updated May 29, 2026 Most spreadsheet problems aren't really spreadsheet problems. They're "I stared at this VLOOKUP error for 45 minutes and now I hate Excel personally" problems. AI spreadsheet prompts fix that. Not by being magic, but by being faster than a Google search and more specific than Stack Overflow. This article is a copy-paste kit. Ten prompts you can use today in ChatGPT or Claude, a formula for writing your own, and a few reminders about what not to hand over to a chatbot. The prompt formula that makes AI spreadsheet help actually useful AI is not psychic. If you type "my VLOOKUP is broken" you'll get a generic explanation of how VLOOKUPs work, which you didn't ask for. What you need is context. The formula is simple: role + context + task + constraints + format. Concretely: "You are a spreadsheet expert. I have a Google Sheets file with sales data in columns A through F. Column A has order dates, column D has salesperson names, column F has revenue. Write a SUMIF formula that totals revenue for one specific salesperson. Use column D as the criteria range and F as the sum range. Show me the formula and explain each argument." That's it. Five pieces of information. You get a formula you can actually paste. This is Rule #13 from Don't Replace Me: garbage in, garbage out. Treat AI like a smart intern who's fast, tireless, and occasionally overconfident. Give them the context they need to do the job right. 10 copy-paste AI spreadsheet prompts Prompt 1: Explain a formula you didn't write Inherited a spreadsheet from a colleague who left the company? Classic. Paste this: You are a spreadsheet expert. Explain this formula in plain English, step by step. Tell me what it does, what each argument means, and what it would return if any referenced cells were empty or contained errors. Formula: [paste formula here] Replace the bracket with your actual formula. Ask it to use plain language, not jargon. If the explanation doesn't make sense, reply with "simplify that" and try again. Prompt 2: Write a lookup formula from scratch You are a spreadsheet expert. I have an Excel file. Sheet1 has a list of employee IDs in column A and their department names in column B. Sheet2 has a column of employee IDs in column A, and I need column B to show the matching department name from Sheet1. Write a formula for Sheet2 cell B2 that will work when copied down. Use XLOOKUP if available, otherwise VLOOKUP. Explain what each part does. Swap in your actual column letters and sheet names. The explanation matters because you need to understand what you're pasting. Prompt 3: Clean messy exported data Data exports from CRMs, accounting tools, and booking systems are almost always filthy. Phone numbers with inconsistent formatting, names in all caps, extra spaces nobody asked for. You are a data cleaning expert in Google Sheets. I have a column of phone numbers in column C, rows 2 through 500. The numbers are inconsistently formatted: some have parentheses, some have dashes, some have country codes, some have spaces. I want all of them formatted as plain 10-digit numbers with no punctuation or spaces, like 5551234567. Write a formula I can put in a helper column that will clean each number. If a cell is empty, return empty. Prompt 4: Categorize rows using logic You are a spreadsheet expert. I have an Excel file with a column of invoice amounts in column D (numbers only, no currency symbols). I want a formula in column E that labels each row as "Small" if the amount is under 500, "Medium" if it's 500 to 4999, and "Large" if it's 5000 or more. Write a nested IF formula or IFS formula for cell E2 that I can copy down. Simple, but writing nested IFs from memory at 4pm on a Friday is how typos happen. Just ask. Prompt 5: Find duplicate values You are a spreadsheet expert. I have a Google Sheets file with a list of email addresses in column B, starting at row 2. I want to flag any email address that appears more than once. Write a formula for column C that returns "Duplicate" if the email in that row appears more than once in column B, and returns blank if it's unique. Show me the formula and explain how COUNTIF is being used. Good for client lists, inventory codes, order numbers. Any situation where duplicates are a problem. Prompt 6: Summarize a table in plain English This one is different. You're not asking for a formula. You're asking AI to read your data description and tell you what's worth knowing. You are a business analyst. I have a spreadsheet summary: [paste a small, anonymized table or describe the structure and totals here]. In plain English, tell me what patterns or observations stand out. Don't make up numbers. Only comment on what I've shared. List your observations as bullet points. Important: Only paste data that has no personally identifiable information, no client names, no financial data that's confidential. Anonymize before you paste. The rule is: would you be comfortable if this appeared on a shared screen in a meeting? If no, clean it first. Prompt 7: Build a pivot table plan Before you click into the pivot table builder, figure out what you actually want. You are a spreadsheet expert. I have a sales data table in Excel with these columns: Order Date, Region, Product Category, Salesperson, Units Sold, Revenue. I want to analyze monthly revenue by region and identify which salesperson has the highest revenue per region. Describe the pivot table I should build: what goes in rows, columns, values, and filters. Then tell me if there's a simpler formula-based approach that would give me the same answer. You can then build it yourself with a clear plan, or paste the pivot setup instructions and follow them. Prompt 8: Debug a #N/A or #VALUE! error Every spreadsheet user has stared at a red error cell with the energy of someone who wants to close the laptop and go outside. You are a spreadsheet expert. My VLOOKUP formula is returning #N/A but I'm sure the lookup value exists in the source table. Here is my formula: [paste formula]. Here is a description of my data: [column letters, data types, whether there are any leading/trailing spaces, whether the ID is numeric or text]. Walk me through the most common reasons this error appears and what to check first. Then suggest a modified formula that handles the error gracefully using IFERROR. Nine times out of ten it's a trailing space or a number stored as text. But now you'll know which one. Prompt 9: Turn a business question into spreadsheet steps Sometimes the problem isn't the formula. It's that you don't know what formula to use, or whether a formula is even the right tool. You are a data analyst. I have a question I want to answer using a spreadsheet: [describe your business question in plain English, e.g., "which product sold the most units in Q1, and how does that compare to Q4 of last year?"] My spreadsheet has these columns: [list them]. Tell me step by step how to answer this question using formulas, a pivot table, or both. Don't assume I know advanced functions. Explain each step. This is the prompt most people don't think to use. It turns vague requirements into an actual plan. For more templates like this, the general AI prompts guide has 12 you can adapt for any work task. Prompt 10: Build a QA checklist before you share the numbers You've built the spreadsheet. Now you're about to send it to your manager, your client, or the whole team. This prompt helps you not embarrass yourself. You are a spreadsheet auditor. I am about to share a financial summary spreadsheet. The spreadsheet contains [describe briefly: what it tracks, time period, audience]. Create a QA checklist of the most important things I should verify before sending: formula accuracy, broken references, hardcoded numbers that should be formulas, inconsistent date formats, totals that don't match their source data, and anything else a reviewer might catch. Format as a checklist I can print or copy. Run through this before every significant send. It takes three minutes and saves you from the "actually there's an error in row 47" reply. How to write your own AI spreadsheet prompts The ten prompts above cover the most common situations. But your spreadsheet is not a common spreadsheet. It has weird inherited logic, columns named things like "Final_FINAL_v3," and at least one formula referencing a tab that no longer exists. So here's how to write your own prompt from scratch, using the role-context-task-constraints-format structure. Role. Start with "You are a spreadsheet expert" or "You are a data analyst." This tells the model to stop being a generalist and start being specific. It matters more than it sounds. Context. Describe your data. Not the whole file, just the relevant columns. Column letters, what's in them, whether the data is text or numbers or dates. "Column A is order dates formatted as MM/DD/YYYY, column B is product SKU as a text string, column C is units sold as a whole number." That's enough. Task. Say exactly what you want. Not "help me with my formula" but "write a formula that counts how many rows have a date in column A that falls within the last 30 days." Constraints. This is where most people skip ahead and regret it. Does it need to work in Excel 2019? Then XLOOKUP is off the table. Are you in Google Sheets? Then certain Excel-only functions won't work. Do you need the formula to handle blank cells gracefully? Say so. Every constraint you include is a failure mode you're preventing. Format. Tell it how to give you the answer. "Show me the formula, then explain each argument in plain English, then show me an example of what it would return if column A had the date 03/15/2025." Specific format requests get specific answers. Write prompts this way and you'll get something usable on the first try about 80% of the time. The other 20% is iteration: copy the formula into your sheet, see what breaks, describe the breakage, and ask again. Two rounds is almost always enough. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What not to do with these prompts AI is useful here, but it's not infallible and it's not a vault. Do not paste confidential data. That means: client names, employee records, payroll figures, HR data, health information, legal documents, financial forecasts marked confidential, or anything your company's data policy says stays internal. Anonymize first. Use fake names and rounded numbers for structure, then apply the real data yourself once you have the formula. Do not send AI-generated analysis you can't explain. If your manager asks "how did you get this number?" and your answer is "the chatbot calculated it," that's a problem. For more on what AI actually can and can't do under the hood, this plain-English explainer is worth five minutes. Verify every formula before it touches live data. AI gets cell references wrong. It sometimes uses functions that don't exist in older versions of Excel. It occasionally hallucinates a syntax that looks plausible but isn't. Test in a dummy row first. Check your dates. Date formatting is where AI suggestions go sideways most often, especially when mixing Excel and Google Sheets conventions. Always verify. Don't skip the explanation. Every prompt above asks AI to explain what it wrote. That's not just for learning. It's how you catch errors. If the explanation doesn't match your data structure, the formula is wrong. When AI spreadsheet prompts save you time and when they don't The honest answer on timing: AI is fastest on well-defined, language-shaped problems. Writing a formula you know the logic for but can't get the syntax right. Cleaning data with a consistent structure. Debugging an error where you can describe exactly what you're seeing. These are the sweet spots. Where it slows you down is when the problem isn't actually a formula problem. If your data is genuinely inconsistent, if the business logic hasn't been agreed on yet, or if the spreadsheet is doing something it was never designed to do, AI will cheerfully write you formulas that produce wrong answers very efficiently. That's worse than the original problem. A useful question to ask before prompting: can I describe this problem clearly in one paragraph? If yes, AI can probably help. If you spend five minutes trying to write the prompt and keep restarting, the problem is probably not the formula. It's the underlying data or the underlying question. The deeper skill isn't memorizing prompts. It's knowing when the problem is a formula problem versus a judgment call versus a "this data is actually broken" problem that no formula will fix. AI can help with the first. You handle the other two. If you want to get better at this generally, the starter guide to using AI at work covers the mindset shift that makes every prompt better, not just spreadsheet ones. And for the non-technical people who want a broader view of which AI skills actually matter in 2026, without having to learn to code, this breakdown is a realistic look at what's worth your time. Frequently asked questions Can I use ChatGPT or Claude to write Excel formulas? Yes, both work well for writing and explaining formulas. Give them specific context: what your columns contain, what data types, and what result you want. The more detail you provide, the less back-and-forth you'll need. Always test the formula before applying it to real data. Is it safe to paste my spreadsheet data into ChatGPT? It depends on what's in it. Never paste confidential employee records, client data, payroll figures, financial forecasts marked private, or anything your company's data policy restricts. Anonymize sensitive details before using them as examples. When in doubt, describe the structure without the real values. Why does the formula AI gives me return an error? Usually one of three things: the cell reference doesn't match your actual data layout, a number is stored as text (or vice versa), or the function isn't available in your version of Excel or Sheets. Include your column letters and data types in the prompt, ask for an IFERROR wrapper, and test on one row before copying down. Do I need any special AI tools for spreadsheet help? No. A free ChatGPT account or Claude's free tier handles all of these prompts. You don't need a spreadsheet-specific AI plugin or a $97 course. The quality of the output depends almost entirely on how well you describe the problem, not which tool you use. What's the best way to prompt AI for spreadsheet help? Use the role-context-task-constraints-format structure. Tell it you want a spreadsheet expert, describe your data layout, specify the exact task, add any constraints (like compatibility with Google Sheets or Excel 2019), and say what format you want the answer in. That one habit eliminates about 80% of useless responses. Can AI replace knowing how to use spreadsheets? No. It's a faster way to get syntax and explanations, but you still need to understand your data, know when a number looks wrong, and be able to explain your work to someone else. If you want to understand where AI genuinely helps versus where human judgment is irreplaceable, the relevant chapter in Don't Replace Me covers exactly that distinction. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance --- ## AI Stakeholder Update Prompts: 10 Templates URL: https://dontreplace.me/articles/ai-stakeholder-update-prompts-10-templates ai stakeholder update promptsproject communicationstatus updatesai at workprompt templates AI Stakeholder Update Prompts: 10 Templates 10 copy-paste AI stakeholder update prompts for weekly summaries, exec briefs, client updates, and blocker escalations. Real templates, honest results. July 07, 2026 Updated July 07, 2026 Stakeholder updates are where careers quietly go sideways. Not because you lied. Because you let a smooth summary replace an honest one. AI stakeholder update prompts can help you organize messy notes, translate project chaos into plain English, and write the same update five ways for five different audiences. What they can't do is tell you which detail matters, whether a timeline slip is a blip or a signal, or when "we're working through some challenges" is hiding something a senior leader needs to hear today. That distinction is the whole job. The prompts below will handle the formatting. You handle the judgment. Why stakeholder updates are so easy to get wrong with AI There's a specific failure mode here. You dump your notes into ChatGPT, ask for a polished update, and get back something that reads like everything is fine. Confident. Clean. Professional. It's not lying exactly, it's just that AI is very good at making uncertainty sound like competence. The tool doesn't know what it doesn't know. It can't tell whether your timeline shifted because of a technical blocker you're actually on top of, or a dependency you forgot to mention to anyone. It just picks the interpretation that produces the smoothest paragraph. Rule #5 in Don't Replace Me puts it plainly: AI is fast, not smart. It can organize your notes beautifully. That's different from understanding your project. Use these prompts to do the fast part faster. Then read the output yourself before it goes anywhere. Before you use any of these: what not to paste A quick and non-negotiable list. Don't paste any of the following into an unapproved AI tool: Customer names, emails, or any personally identifiable information Employee performance records or HR issues Legal disputes, settlement details, or anything that's touched a lawyer Unreleased product details or confidential roadmap items Financial forecasts, board materials, or investor communications Security incidents, vulnerabilities, or access credentials Medical or safety information Sensitive client contracts or pricing If your stakeholder update involves any of these, either use an approved enterprise AI tool with appropriate data governance, or write that section yourself. The prompts below are built for sanitized project information. If you're not sure whether something is sensitive, assume it is. The reusable AI stakeholder update formula Before the 10 templates, here's the structure every good stakeholder update follows. Use this as your prompt foundation: Context: [Who this update is for, their role, what they care about] Source material: [Your raw notes, pasted in] What I need: [Type of update: weekly summary, exec brief, client-facing, etc.] Constraints: [Tone, length, what to include or exclude] What's true but messy: [The honest version of the status] What I don't want AI to do: [Fill gaps, invent progress, soften accountability, imply certainty] Every prompt below is built on this foundation. You can adapt it to any situation once you see the pattern. This formula also works in reverse. If you've already written a draft update and want to stress-test it, paste it in and ask the AI what looks unclear, vague, or missing. You'll often catch something you talked yourself out of noticing. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 copy-paste AI stakeholder update prompts These are ready to use. Replace the bracketed sections with your actual information. Prompt 1: Turn messy notes into a weekly stakeholder update Here are my raw project notes from this week: [paste notes] Turn these into a weekly stakeholder update for [audience: for example, my direct manager]. Structure: what was accomplished, what's in progress, blockers, next steps. Use only what's in my notes. If something is unclear, flag it as a question rather than filling it in. Do not invent progress or smooth over anything that looks like a problem. Keep it under 200 words. Prompt 2: Summarize what changed since last update Last update: [paste previous update or summary] Current status: [paste current notes] Write a "what changed" section for a stakeholder update. List only actual changes, not things that stayed the same. Flag anything that changed negatively. Do not frame setbacks as progress. Prompt 3: Explain a timeline shift without waffling Our timeline changed. Original deadline: [date]. New deadline: [date]. Reason: [your honest explanation]. Write a two-paragraph explanation for [audience]. Be direct about the reason. Do not minimize the impact. Include: what caused the shift, what we're doing about it, and what the new plan is. Do not use passive voice to avoid accountability. Prompt 4: Write an executive update Project: [name] Status: [on track / at risk / delayed / blocked] Key update this period: [one sentence] Decision needed from leadership: [yes/no, and what if yes] Risk to flag: [if any] Write a 150-word executive update. Lead with status and any decision needed. Executives don't need background. They need: status, risk, decision ask, next milestone. For more on writing updates executives actually read, the AI executive summary prompts page has additional templates built specifically for that format. Prompt 5: Prepare a client-safe update Internal status: [paste your real status] Client relationship context: [for example, nervous client, new client, long-standing relationship] What the client needs to know: [be honest] What stays internal: [list anything not appropriate for client view] Write a client-facing update that is honest, professional, and doesn't overpromise. If there's a problem, name it clearly. If there's good news, don't bury it. Do not invent timelines or make commitments we haven't confirmed internally. If you're navigating a difficult client situation, AI client communication prompts covers the harder conversations, not just the status updates. Prompt 6: Flag a blocker without creating drama Blocker: [describe it specifically, including what's stuck, who owns it, and how long it's been stuck] Impact: [what happens if it's not resolved, and by when] What I need from the reader: [specific ask: decision, escalation, resource, awareness only] Write a blocker escalation paragraph. Be calm, specific, and clear about the ask. Do not editorialize or assign blame. Do not soften the impact if it's real. The AI escalation plan prompts page has a fuller framework if the blocker needs a formal escalation path. Prompt 7: Turn decisions into a decision recap Decisions made this week: [list them with who made each one and when] Context for each: [brief] Write a decision recap section for a stakeholder update. For each decision include: what was decided, who made the call, the date, and any follow-on actions created. This is a record, not a sales pitch. Neutral tone. Prompt 8: Create a risk-and-next-steps table Current risks: [list them, with likelihood and impact if you have it] Mitigation actions: [what's being done about each] Owner: [who owns each mitigation] Next steps: [what happens next on each] Format this as a simple table with columns: Risk | Likelihood | Impact | Mitigation | Owner | Next Steps. Use the information I've provided. Do not add risks I haven't mentioned. Prompt 9: Adapt one update for different audiences Base update: [paste your draft update] Rewrite this update three ways: 1. For my direct manager (detail-oriented, wants specifics) 2. For a cross-functional stakeholder (needs context, cares about dependencies) 3. For an executive (needs status, risk, decision ask only, no background) Keep the facts identical across all three. Only adjust depth and framing. Prompt 10: Check whether the update is honest enough Here is my draft stakeholder update: [paste it] Review it for these specific problems: - Passive voice used to avoid naming who's responsible - Vague language where a specific number or date should be - Missing bad news that's implied but not stated - Commitments made without a named owner - Language that implies more certainty than actually exists List any problems you find. If the update is clean, say so. This last one is probably the most useful prompt in the set. Not because AI has great judgment about honesty, but because it forces you to look at your own draft with a more critical eye. Most people, if they're being honest, already know what they're softening. The prompt just makes it harder to ignore. What AI can help you do vs. what you still own Here's the honest split: TaskAI can helpYou own Organizing raw notes into a structureYesReviewing for accuracy Writing for different audiencesYesDeciding what to include Formatting risks into a tableYesAssessing actual severity Drafting a timeline explanationYesConfirming the reason is honest Flagging a blocker clearlyYesDeciding whether to escalate Checking your draft for vague languagePartiallyFixing the vagueness Deciding what a senior leader needs to knowNoYou Judging whether a problem is seriousNoYou The people who get this wrong think the polished output is the work. It's not. The work is everything in the right column of that table. For a broader look at where AI helps and where it doesn't, what AI can and can't do covers this in plain language without the jargon. Common mistakes people make with AI stakeholder update prompts Even with solid prompts, a few patterns keep showing up. Worth naming them so you can avoid them. Dumping raw notes without context. "Here are my notes, write an update" gives the AI nothing to work with in terms of audience, stakes, or tone. The output will be structurally fine and contextually useless. Always include who you're writing for and what they care about. Accepting the first draft. AI first drafts are a starting point, not a finished product. The first output will almost always be slightly too optimistic, slightly too vague on the hard parts, and missing at least one thing you actually need to say. Read it. Edit it. Send the second version. Using the update to avoid a conversation. This one's less about AI and more about human nature, but the tools make it worse. It's very easy to produce a professional-sounding update that technically mentions a problem without actually communicating its severity. "We are monitoring some resource constraints" can mean anything from "it's fine" to "we're about to miss the deadline and I haven't told anyone." If the situation requires a direct conversation, no update format fixes that. Forgetting the ask. A lot of AI-generated updates trail off. They summarize the situation and stop. Before you send anything, check: does the reader know what to do with this? Do they need to make a decision, approve something, escalate something, or simply acknowledge? If you're not sure, add a line at the bottom: "No action needed" or "I need [X] by [date]." Either answer is better than silence. Using polished language to paper over vague thinking. This is the most dangerous one. AI is exceptionally good at making half-formed thoughts sound authoritative. If you don't fully understand the situation you're reporting on, the AI will generate a confident-sounding paragraph anyway. The prompts above include explicit instructions to flag gaps rather than fill them. Use that instruction. When to escalate instead of update Some situations don't belong in a weekly update at all. If your status notes include any of the following, you're not writing an update, you're writing around a conversation you need to have directly: A financial exposure you haven't confirmed with finance A client situation that could affect the relationship significantly A security or data incident of any kind An HR issue involving a team member Legal risk, even a small one Something that will surprise a senior leader in a bad way AI can help you prepare for that conversation. But it doesn't replace having it. If you're not sure whether your situation crosses this line, assume it does and loop in the relevant person: your manager, legal, security, finance, HR, or whoever owns the exposure. The AI risk assessment prompts page has a useful framework for thinking through blast radius before you decide what to communicate and to whom. The thing polished updates can't fake There's a reason good stakeholder communication is hard. It requires you to know what you know, know what you don't, and be honest about both in a way that's useful to the reader without causing unnecessary panic. That's judgment. AI doesn't have it. The prompts above will save you time on the drafting. They won't save you from the one mistake that actually costs people: confusing a smooth update with an honest one. A senior leader who gets three polished updates in a row and then discovers a serious problem wasn't mentioned will remember that, regardless of how well-formatted the updates were. The tool is fast. You're the one who knows what's actually happening. That combination, when you use it right, is the thing no one can replace. Frequently asked questions What should I include in an AI stakeholder update prompt? The most useful prompts include: who the audience is, what they care about, your raw status notes, what type of update you need, and explicit instructions not to invent progress or smooth over problems. The more context you give, the more useful the output. Vague input produces confident nonsense. Can AI write stakeholder updates for me? AI can draft stakeholder updates from your notes, but you need to review every output for accuracy before it goes anywhere. AI will make uncertain situations sound more resolved than they are. It doesn't know the difference between a timeline slip you're handling and one that needs executive attention. That judgment is yours. What information should I never paste into an AI tool for a stakeholder update? Don't paste customer PII, employee records, legal matters, security incidents, unreleased product details, financial forecasts, board materials, medical or safety information, or sensitive client contracts into unapproved AI tools. If your update involves any of these, either use an enterprise-approved tool or write that section manually. How do I write a stakeholder update that flags a problem without panicking everyone? Be specific and calm. Name the problem, name the impact, and name what you need from the reader. "We have a blocker: the API integration is blocked on vendor approval, which has been pending 10 days. If not resolved by Friday, we slip one week. I need [name] to escalate to their vendor contact." That's more useful than "we're experiencing some challenges with the integration timeline." How is an AI stakeholder update different from an AI status report? A status report is usually a structured document covering the full project health: scope, schedule, budget, risks, issues, decisions. A stakeholder update is audience-specific communication, often shorter, filtered for what a particular group needs to know. For status report templates, AI status report prompts covers that format specifically. Should stakeholder updates always have a decision ask? Not always, but often. If you're sending an update and there's something you need from the reader, say so explicitly. If it's awareness only, say that too. "No action needed, awareness only" is a complete and useful sentence. Stakeholders who have to guess what you need from them tend to do nothing. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Presentation Prompts: 10 Templates That Work 10 copy-paste AI presentation prompts for slide outlines, speaker notes, executive Q&A prep, and more. No jargon, no fluff, just practical templates. ai presentation promptschatgpt for presentationsai at work AI Prioritization Prompts: 10 Templates That Work 10 copy-paste AI prioritization prompts for task lists, backlogs, roadmaps, and tradeoffs. Practical templates with real safety guardrails. AI prioritizationprompt templatestask management AI Prompts for Work: 12 Templates You Can Use Today 12 copy-paste AI prompt templates for emails, reports, and decisions. Real examples for normal people, with what to check before you send anything. ai prompts for workchatgpt promptsworkplace ai --- ## AI Status Report Prompts: 10 Templates That Work URL: https://dontreplace.me/articles/ai-status-report-prompts-10-templates-that-work ai status report promptsstatus report templateAI at workproject communicationChatGPT prompts AI Status Report Prompts: 10 Templates That Work 10 copy-paste AI status report prompts for weekly updates, exec summaries, client updates, and delay messages. Real templates, real results. June 20, 2026 Updated June 20, 2026 Status reports are the dark matter of office work. Everyone has to produce them. Nobody likes writing them. And the gap between what actually happened this week and what ends up in the report can be... creative. AI status report prompts can close that gap fast. Not by making things up, but by turning your messy, half-finished notes into something a human can actually read. That's the job. Structure, clarity, audience tuning. Not invention. The problem is most people either don't use AI for this at all, or they use it badly. They paste in a vague paragraph and ask for a "professional status report," then send whatever comes out without reading it. That's how you end up with a confidently written update claiming you're "on track" when you're two weeks behind and three people haven't responded to a single message. This guide gives you 10 copy-paste prompts, a reusable formula, and the rules for using all of it without embarrassing yourself. The formula behind every good AI status report prompt Before the templates, one rule: garbage in, garbage out. Your AI tool is fast, not clairvoyant. If you paste in bullet points that say "still waiting on thing from Sarah, budget maybe fine," you'll get a polished report that says... "we're monitoring budget status and awaiting stakeholder input." Which is nothing. A useful AI status report prompt has five ingredients: Context. What's the project, who's the audience, what's the format you need. Raw facts. What actually happened. Real progress, real blockers, real numbers. Names, dates, owners. What you need from the AI. A summary? An executive version? A client-friendly paragraph? Tone. Formal update for the board vs. casual standup message are different things. What to leave out. Any gaps in your notes the AI should flag rather than fill. Keep that formula in mind and every prompt below gets sharper. The formula matters more than the template. A great template with bad inputs gives you garbage. A mediocre template with specific, honest notes gives you something you can actually send. Put your time into the notes, not into finding the perfect prompt. What AI can and can't do in a status report AI can turn a messy list of notes into a readable, structured update in about 30 seconds. That's real and it's useful. It can adjust tone for different audiences, spot where your update is missing something obvious, build a follow-up checklist from action items, and write the same status five different ways depending on whether the reader is an exec, a client, or a developer. What it cannot do: know what actually happened. It will not invent problems if you don't give it any. But it will happily paper over missing information with plausible-sounding filler if you let it. It doesn't know your project. It doesn't know your client's sensitivities. It doesn't know that "we're exploring options" is a phrase your CEO specifically hates. That's your job. The AI handles structure. You handle truth and judgment. If you want to understand where those lines sit more clearly, the article on what AI can and can't do is worth five minutes of your time before you start. There's also a tone problem worth naming. AI defaults to a particular kind of calm, measured corporate voice. That voice is fine most of the time. But sometimes the situation calls for urgency, or directness, or even some controlled alarm. If your project is genuinely in trouble, AI will still write you a reassuring update unless you specifically tell it not to. You have to encode the emotional register, not just the facts. The rule: every AI-generated status report gets human eyes before it leaves your screen. No exceptions. What never goes into an AI status report prompt This section exists because people skip it and then have a bad day. Do not paste any of the following into an AI tool that hasn't been approved for sensitive data at your organization: Customer names, contract details, or anything that could identify a specific client Employee performance information or HR records Financial figures that are unreleased or board-level only Security incidents, vulnerabilities, or anything your IT or legal team owns Legal disputes or anything involving lawyers Credentials, passwords, or system access information Unreleased product details or roadmap specifics marked confidential Private strategy documents or M&A discussions When in doubt, anonymize. "Client A" instead of the actual name. "Q3 target" instead of the specific number. You can still write a useful status report from anonymized notes. You can't un-paste something that went to a model with a data retention policy you haven't read. Most consumer AI tools, including the free tier of ChatGPT, have data practices that your legal or IT team may not have reviewed. If your company has an approved enterprise AI tool, use that for anything sensitive. If you're not sure, ask. This is one of those situations where the five-minute check is worth it. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI status report prompts you can use today These are designed to be copy-paste ready. Replace the bracketed text with your actual information. Don't remove the instruction to flag gaps rather than fill them. That line does real work. 1. Turn messy notes into a weekly status report Here are my raw notes from this week on [project name]: [Paste your notes] Turn these into a structured weekly status report with four sections: Progress This Week, Blockers/Risks, Next Steps, and Decisions Needed. Use plain, professional language. If any section is unclear from my notes, write [NEEDS CLARIFICATION] rather than guessing. Audience: [project team / my manager / both]. 2. Write an executive summary Here is a detailed status update for [project name]: [Paste your full update] Write a 3-sentence executive summary for a senior leader who has no time to read the full update. Lead with the most important thing they need to know. Flag any decision they need to make. Do not add information that isn't in the original update. 3. Create a client update Here is the internal status update for [project name]: [Paste internal update] Rewrite this as a client-facing update. Keep the tone confident and professional. Include what we completed, what's coming next, and any open items we need from the client. Remove internal team names, internal jargon, and any information that shouldn't be visible to the client. Flag anything I should double-check before sending. 4. Explain a delay without destroying confidence I need to communicate a delay to [audience: client / exec / project team]. Situation: [Describe what was planned, what's delayed, why, and what the new expected date is]. Write a professional message that explains the delay clearly, takes accountability, gives the revised timeline, and explains what's being done to prevent further delays. Don't soften it to the point of being vague, but don't catastrophize either. I'll review before sending. For more on these harder conversations, the AI client communication prompts article has templates specifically built for delivering bad news. 5. Summarize risks and blockers Here are the current risks and blockers on [project name]: [List your risks and blockers with as much specificity as you have: what the risk is, likelihood, impact, owner, mitigation in progress] Write a concise risk summary for [audience]. For each item, include: the risk, current status, owner, and what decision or action would resolve it. Use [MISSING OWNER] or [MISSING DATE] if that information isn't in my notes. 6. Turn meeting notes into action items Here are the notes from a meeting on [date] about [project]: [Paste meeting notes] Extract all action items. For each, provide: what needs to be done, who owns it (if mentioned), and the due date (if mentioned). Where ownership or dates weren't discussed, flag as [UNASSIGNED] or [NO DATE SET]. Format as a clean list. The AI meeting notes prompts article has 10 more of these for different meeting types if this is something you do a lot. 7. Write a team standup update Here's what I'm working on: Yesterday: [What you completed or made progress on] Today: [What you're working on] Blockers: [Anything stopping you or slowing you down] Rewrite this as a clear, concise standup update. Keep it under 100 words. Professional but not stiff. Do not add detail that isn't here. 8. Prepare a stakeholder escalation I need to escalate an issue to [role: senior leadership / project sponsor / client exec]. Here's the situation: [Describe the issue, what's been tried, what the impact is, and what you need from them] Write a professional escalation message that clearly explains the problem, shows what's already been done to resolve it, describes the impact if it stays unresolved, and makes a specific ask. Tone should be direct without being alarmist. I'll review before sending. 9. Compare this week vs. last week Last week's status update: [Paste] This week's notes: [Paste] Write a brief comparison that shows: what moved forward, what didn't, what new risks or issues emerged, and whether the project is trending better or worse overall. Use only the information I've provided. Flag gaps with [NEEDS CLARIFICATION]. 10. Create a follow-up checklist Here is the status update or meeting summary for [project]: [Paste content] Extract a follow-up checklist of everything that needs to happen before the next update. For each item: what needs to be done, who should do it, and when it's due (if mentioned). Format as a simple checklist. Mark anything with no clear owner as [UNASSIGNED]. How to adapt these prompts for different audiences Same project, different readers, different reports. This is where people often underuse AI. Your weekly update for the project team can be detailed, technical, and full of names. The same information going to an exec needs to be three sentences with a single decision request. The client version strips out all internal context and reframes everything around their outcomes, not your process. AI handles these rewrites well. Once you have one solid version of the status, you can paste it back in with a simple instruction: "Rewrite this for [new audience]. Keep only what's relevant to them. Remove anything internal." You don't need a new prompt from scratch. You just need to change the audience parameter. A few audience-specific adjustments worth making explicit when you prompt: Executive audience: Lead with impact and decisions needed. Cut process detail. No internal team names unless a specific person needs to act. Client audience: Focus on their outcomes, their deliverables, their open items. Remove your internal drama. Sound confident even when things are complicated. Developer or technical team: More detail is fine. Be specific about blockers. Technical jargon is okay here, unlike everywhere else. Project team: Full picture. Blockers, owners, dependencies. Don't sanitize. They need the real version. The prompts in this article each specify an audience. That's not decoration. Changing that one word changes the output significantly. The safety checks that matter before you hit send Every one of these prompts produces a draft. A draft. Not a final document. Before you send anything the AI helped write, run through this list: Are all dates, deadlines, and timelines accurate? AI doesn't know your calendar. Are all owner names correct and current? People leave projects. Roles change. Does the risk section reflect actual risk levels, not AI-softened "monitoring" language? Is there anything in the update the AI invented that wasn't in your notes? Did anything confidential end up in the prompt that shouldn't have? Does the tone match the actual situation? AI tends toward calm even when calm isn't appropriate. Rule 13 in Don't Replace Me says it plainly: vague notes create polished nonsense. The tool is fast. That's its only advantage. The accuracy and the judgment are still yours. This connects directly to something the AI project management prompts article gets into: AI is a drafting assistant for project communication, not a replacement for the project tracker, the accountability conversation, or the call where you tell someone the launch is slipping. When AI status report prompts make things worse It's worth naming this. Some people use AI status reports as a way to avoid thinking about the project. The report comes out clean and organized. The underlying situation is a mess. The report creates a false sense of control. A polished update is not the same thing as a healthy project. If you're using AI to make a disaster sound like a delay, you're building debt. The client, the exec, or the team will figure it out eventually, and the gap between the report and reality will land on you. There's a specific pattern to watch for: the AI-smoothed update that removes all urgency from a genuinely urgent situation. You paste in notes that say "critical path at risk, three dependencies outstanding, no response from vendor in eight days." The AI produces a paragraph about "active coordination with key partners to resolve outstanding dependencies." You read it, it sounds fine, you send it. Six weeks later someone asks why nobody flagged the vendor issue earlier. AI is useful for expressing your status clearly. It's not useful for pretending you have status you don't. Before you generate the report, do a quick gut check: if I read this out loud in a meeting, would the words be true? If not, fix the inputs. The AI decision-making prompts article has a few prompts that help with the harder question underneath the status report: what actually needs to happen next. Frequently asked questions Can AI write a status report for me? AI can draft a status report from your notes, but it can't replace the notes themselves. Give it real facts, real names, real dates, and it'll organize them clearly and quickly. Give it vague inputs and you'll get a polished document that says nothing useful. What's the best AI prompt for a weekly status report? The most reliable structure is: context (project name, audience), raw facts (actual progress, blockers, decisions), what format you need, and an instruction to flag gaps rather than fill them. Prompt 1 in this article follows that structure and works well as a starting point. Is it safe to use AI for client status updates? It can be, with two conditions. First, don't paste anything identifying specific clients, contracts, or confidential project details into an unapproved AI tool. Anonymize names and sensitive figures first. Second, always review the output before sending. AI doesn't know your client relationship or their sensitivities. How do I use AI to explain a delay without sounding bad? Give the AI the full situation: what was planned, what slipped, why, what the new timeline is, and what's being done to prevent further delays. Ask for directness over vagueness. A good delay message takes accountability and gives a revised plan. Prompt 4 in this article is built specifically for this. Will AI invent progress or deadlines in a status report? It can if you let it. If your notes have gaps, a less careful prompt will produce filler. The prompts in this article include an instruction to mark missing information as [NEEDS CLARIFICATION] rather than guess. Always read the output against your actual project notes before sending. How is using AI for status reports different from using it for project management? Status reports are communication. Project management is coordination, tracking, and accountability. AI can help with the communication layer. It can't replace your project tracker, your named owners, your real dates, or the conversation where you tell someone they're behind. For the broader project management side, the AI project management prompts article covers that territory. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts AI Meeting Agenda Prompts: 10 Templates That Work 10 copy-paste AI meeting agenda prompts for decision meetings, kickoffs, 1:1s, retros, and more. Includes the reusable formula behind every good prompt. ai meeting agendameeting promptsai at work --- ## AI Support Handoff Prompts: Stop Losing Context URL: https://dontreplace.me/articles/ai-support-handoff-prompts-stop-losing-context-after-launch AI supportsupport handoffpost-launchcustomer successoperations AI Support Handoff Prompts: Stop Losing Context AI support handoff prompts for post-launch docs, known issues, escalation paths, support coverage, and customer-safe updates after go-live. August 01, 2026 Updated August 01, 2026 The launch team celebrates. Support inherits the haunted house. That is why AI support handoff prompts are useful. Not because AI can magically understand your customers, read private tickets, verify production health, or bless a messy launch into stability. It cannot. But it can turn scattered launch notes, known issues, ticket themes, owner lists, escalation paths, workarounds, and monitoring links into a support handoff that does not rely on one tired project manager remembering everything from a Slack thread called final-final-launch-stuff. A good support handoff keeps context alive after go-live. It tells support what changed, what might break, who owns each issue type, what customers can be told, where to escalate, and when the launch team is officially allowed to stop hovering. If you are still preparing for release, start with AI launch checklist prompts or AI smoke test prompts. If the thing is already live and the question is “who has the context now?”, this guide is for you. AI can organize the handoff. A human still owns the facts, customer impact, sensitive context, and final sign-off. What is a support handoff after launch? A support handoff is the transfer of operational context from the launch team to the people who will support the thing after the launch window ends. It answers the questions that become expensive when nobody wrote them down: What changed? Who is affected? What known issues exist? Which workarounds are approved? What should support say to customers? What requires engineering, customer success, security, legal, or leadership review? Who owns follow-up work after hypercare ends? Without a handoff, support gets vibes, screenshots, and “ping Alex if anything looks weird.” Alex is on vacation. The screenshot is from staging. The customer is not amused. AI support handoff prompts help because AI is good at structure. It can organize notes into sections, summarize sanitized ticket themes, draft known-issue tables, create escalation matrices, rewrite release notes for support teams, and flag missing information. That is useful, especially after a launch when everyone is fried. But this is still the same core rule from what AI can and can't do: use AI for speed and structure, not authority. It can make the handoff cleaner. It cannot make the handoff true. The reusable AI support handoff prompt formula Use this formula when you want AI to help with a post-launch support handoff: “You are a support handoff assistant. I am [your role] handing off [launch/project/change] from [launch team] to [support team]. Here is sanitized context: [scope, release notes, known issues, ticket themes, affected customers or user groups, owners, escalation paths, monitoring links, workarounds, open questions]. Create [specific handoff output]. Separate confirmed facts from assumptions. Flag missing information, risks, approvals needed, and anything support should not say yet. Do not invent customers, owners, production status, root causes, severity, commitments, or decisions.” The last sentence does real work. AI will often smooth over uncertainty because smooth answers look helpful. In support, smooth uncertainty is how you accidentally tell customers nonsense with confidence. Before using any prompt, sanitize the input. Do not paste customer PII, credentials, access tokens, raw production logs, private tickets, unreleased strategy, legal disputes, security vulnerabilities, financial records, HR issues, regulated data, medical information, private client conversations, or sensitive personal information into unapproved AI tools. If the prompt requires raw private data to work, rewrite the input. Use summaries, categories, counts, approved excerpts, and links to authorized systems instead. The robot does not need the customer’s home address to help you write a ticket-theme summary. Tiny mercy. What to bring before asking AI for a handoff AI cannot hand off context you never provide. It can only organize the ingredients you bring. Use this checklist before you ask for a support handoff: InputWhy it mattersHuman check Launch scopeKeeps support focused on what actually changedConfirm it matches the final release Release notesGives support the plain-language change summaryRemove internal-only claims before sharing Known issuesPrevents duplicate discovery and panicInclude severity, status, owner, and workaround Ticket themesShows what users are already experiencingUse privacy-safe summaries, not raw tickets Escalation pathsMakes urgent issues movableName real owners or approved rotations Support coverageAvoids dead zones after hypercareInclude time zones and backups Monitoring linksGrounds the handoff in evidenceUse approved dashboards and real thresholds Customer messagingKeeps everyone aligned externallyMark what is approved vs. draft Follow-up commitmentsPrevents launch leftovers from vanishingAssign owners and dates Sunset criteriaDefines when launch support becomes normal supportUse measurable signals, not “seems fine” If you do not have these inputs, ask AI to create a missing-information checklist first. Do not ask it to write a complete handoff from fragments unless you enjoy professionally formatted hallucinations. For the window before this point, AI hypercare plan prompts can help structure the first days after go-live. The support handoff is what keeps that temporary launch knowledge from evaporating when hypercare ends. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI support handoff prompts you can use today Prompt 1: Turn launch notes into a support handoff You are a support handoff assistant. I am handing off this launch to the support team: [sanitized launch summary]. Here are the release notes, known issues, support channels, ticket themes, owners, escalation paths, monitoring links, workarounds, and open questions: [paste details]. Create a support handoff doc with sections for what changed, who is affected, known issues, approved workarounds, customer messaging, escalation rules, owner matrix, monitoring notes, open risks, and follow-up actions. Separate confirmed facts from assumptions. Flag missing information and approvals needed. What to check after: Make sure every owner, workaround, and customer-facing statement is real. “Support can explain the workaround” is not a plan if nobody has approved the workaround. Prompt 2: Summarize known issues safely You are a support documentation assistant. Here are privacy-safe summaries of known issues after launch: [paste sanitized issue list]. Create a known-issues table with issue name, affected workflow, symptoms, severity, status, workaround, owner, next update date, and customer-facing wording if approved. If customer wording is not approved, mark it as "needs approval" instead of drafting a final statement. What to check after: AI can turn chaos into a table. It cannot decide whether an issue is safe to mention to customers, whether the workaround is acceptable, or whether severity is politically survivable. For cleaner issue writeups, pair this with AI bug report prompts. Prompt 3: Create an escalation matrix You are a support operations assistant. Here are the teams, roles, issue categories, support hours, customer tiers, and approval rules for this launch: [paste sanitized details]. Create an escalation matrix. Include issue category, first responder, backup, engineering owner, customer success owner, response target, escalation trigger, approval needed, and communication channel. Do not assign owners that are not provided. Mark gaps clearly. What to check after: A team name is not an owner. “Engineering” cannot join the call. “Customer Success” cannot approve a customer credit. Use names, rotations, or approved escalation queues. Prompt 4: Convert release notes into support guidance You are a support enablement assistant. Here are the approved release notes: [paste release notes]. Convert them into support guidance for frontline teams. Include plain-language summary, customer impact, likely questions, what changed, what did not change, troubleshooting notes, known limitations, approved links, and items support should not promise. Flag anything that needs product or legal review. What to check after: Release notes often explain what shipped. Support guidance explains what customers will ask when the shipped thing behaves like software, meaning confidently weird. If your release notes are still messy, use AI release notes prompts before turning them into support guidance. Prompt 5: Draft customer-safe talking points You are a customer communication assistant. Using only the approved facts below, draft customer-safe talking points for support and customer success. Context: [paste approved facts]. Include a short explanation of what changed, common questions, approved workaround language, what to say if the issue is unresolved, and when to escalate. Do not invent timelines, root causes, compensation, commitments, or private details. Mark uncertain items as "not approved for customer communication." What to check after: Customer-safe does not mean bland. It means accurate, approved, and not accidentally promising a fix by Friday because AI wanted the paragraph to have a satisfying ending. Prompt 6: Map support coverage and gaps You are a support operations assistant. Here is the post-launch support coverage plan: [paste teams, time zones, hours, channels, owners, backup owners, customer tiers, launch dates, and known risk periods]. Identify coverage gaps, unclear ownership, unsupported hours, missing backups, overloaded owners, and escalation risks. Return a table with gap, why it matters, recommended fix, owner needed, and urgency. What to check after: This prompt is especially useful when teams are spread across time zones. AI can spot obvious dead zones. Humans still need to decide whether the dead zone is acceptable or terrifying. Prompt 7: Turn ticket themes into follow-up actions You are a post-launch support analyst. Here are privacy-safe ticket themes from the first [time period] after launch: [paste sanitized themes]. Turn them into follow-up actions. Group by product fix, documentation update, support macro, training need, monitoring change, customer success follow-up, and backlog item. For each action, include evidence, owner type, urgency, and what must be verified before work starts. Do not invent root causes. What to check after: Ticket themes are clues, not verdicts. Three customers saying “slow” might mean a performance bug, a confusing UI, a regional issue, or everyone clicking the same cursed button. For broader post-launch watching, see AI post-launch monitoring prompts. Prompt 8: Compare hypercare support vs. steady-state support You are a launch operations assistant. Here is how support worked during hypercare: [paste coverage, owners, channels, cadence, escalation paths, dashboards, and issue categories]. Here is normal support coverage: [paste steady-state process]. Compare the two. Identify what can move to normal support, what needs a temporary extension, what needs documentation, what needs training, and what should remain with the launch team until resolved. What to check after: This is where teams accidentally dump launch leftovers onto support and call it “transition.” If the launch team still owns unresolved risk, the handoff should say that plainly. Prompt 9: Find documentation and training gaps You are a support enablement assistant. Here is the launch context, known issues, ticket themes, support questions, release notes, and current documentation links: [paste sanitized details]. Identify missing docs, outdated docs, confusing instructions, training gaps, support macro needs, internal FAQ questions, and customer-facing help article updates. Prioritize by customer impact and support volume. Do not invent documentation that does not exist. What to check after: Support handoffs fail when everyone assumes “the docs” exist. AI can find gaps in the material you provide. It cannot browse your internal wiki unless you give it approved excerpts or links. Prompt 10: Write the final handoff sign-off You are a support handoff assistant. Create a final handoff sign-off checklist for [launch/project]. Use this context: [paste sanitized support handoff]. Include required approvals, unresolved known issues, accepted risks, support owners, escalation paths, customer communication status, documentation status, monitoring ownership, follow-up actions, and criteria for closing launch-team involvement. Add a section called "Not ready to hand off if" with blockers. Do not mark anything complete unless the input says it is complete. What to check after: The handoff is not done because the document exists. It is done when the accepting team understands it, owners are real, risks are explicit, and someone accountable signs off. What AI should not do in a support handoff AI should not decide the truth of the launch. It should help organize the truth humans provide. Do not use AI to: invent root causes from ticket summaries; decide customer impact without verified data; approve workarounds; promise fix dates; write final customer statements from unapproved facts; replace engineering, support, security, legal, or customer success review; summarize raw private tickets in unapproved tools; treat missing owners as implied owners; decide that hypercare can end; hide uncertainty because the handoff would look nicer. That last one matters. A support handoff with honest gaps is better than a polished lie. Support teams can work with “unknown root cause, engineering owner needed, customer wording not approved.” They cannot work with fake certainty unless your business model is avoidable chaos. This is also why AI operational readiness prompts are useful before launch. The more readiness gaps you leave unresolved, the more your support handoff becomes a cleanup invoice. A simple support handoff structure If you only need one template, use this: Launch summary: What changed, when it changed, and who is affected. Support scope: What support owns now and what remains with the launch team. Known issues: Symptoms, severity, owner, workaround, and next update. Escalation rules: When to escalate, who receives it, and what evidence to include. Customer messaging: Approved talking points and what not to say. Monitoring and signals: Dashboards, ticket queues, feedback channels, and review cadence. Docs and training: Updated docs, missing docs, support macros, and training needs. Follow-up actions: Owners, dates, and accepted risks. Sign-off: Who accepted the handoff and what still blocks closure. That structure is boring. Good. Boring is the goal. Launch support should be boring. Exciting support handoffs are usually just incident reviews waiting their turn. If things have already gone sideways, use AI incident review prompts afterward to learn without turning the review into a blame piñata. If rollback is still possible, keep AI rollback plan prompts close enough to reach without spelunking through bookmarks. Frequently asked questions Can AI write a support handoff for me? AI can draft the structure, summarize sanitized inputs, and flag missing details. It cannot verify the launch, approve customer commitments, or decide whether support is ready. Treat the draft as a working document that humans review and correct. What should I never paste into AI for a support handoff? Do not paste customer PII, raw private tickets, credentials, access tokens, production logs, legal disputes, security vulnerabilities, regulated data, financial records, HR details, medical information, or confidential customer conversations into unapproved AI tools. Use approved systems and privacy-safe summaries instead. What is the difference between hypercare and a support handoff? Hypercare is the temporary intense support period after go-live. A support handoff is the transfer of context, ownership, known issues, and escalation rules from the launch or hypercare team to steady-state support. One keeps the launch stable; the other keeps knowledge from disappearing. Who should approve an AI-generated support handoff? At minimum, support leadership, the launch owner, product or engineering owners, and customer success should review it. Depending on the content, security, legal, compliance, finance, or executive stakeholders may need to approve customer-facing language or accepted risk. Can AI summarize support tickets for a handoff? Yes, if the input is privacy-safe and approved for the tool you are using. AI can group symptoms, repeated themes, likely affected workflows, and open questions. It should not invent root causes or expose private customer details. When is a support handoff actually done? A support handoff is done when support understands the change, owns the right processes, has approved messaging, knows escalation paths, sees unresolved risks clearly, and accepts the handoff. A document without ownership is just launch confetti in PDF form. The bottom line AI support handoff prompts are best used as a context-preservation tool. They turn messy launch leftovers into a document support can actually use: known issues, owners, escalation paths, talking points, docs, training gaps, and follow-up actions. They do not replace the people who know the product, customers, risk, politics, or consequences. That is the useful human work. If you want the broader survival framework, Don't Replace Me by Dmitry Kargaev is the field guide for staying valuable while everyone else mistakes faster documentation for better judgment. Use AI to make the handoff clearer. Do not use it to launder uncertainty into confidence. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Account Planning Prompts: Build a Better Plan 10 AI account planning prompts to organize buyer evidence, map stakeholders, expose risks, test strategy, and plan next actions without inventing intel. AI promptsaccount planningsales strategy AI Change Impact Analysis Prompts: Spot Damage AI change impact analysis prompts for mapping affected systems, teams, customers, QA checks, support gaps, and rollout risks. AI promptschange impactrisk AI Churn Analysis Prompts: Find Why Customers Leave 10 AI churn analysis prompts to investigate why customers leave, test assumptions, protect customer data, and turn evidence into accountable action. AI promptschurn analysiscustomer retention --- ## AI Survey Design Prompts: Ask Better Questions URL: https://dontreplace.me/articles/ai-survey-design-prompts-ask-better-questions AI promptssurvey designuser researchcustomer feedbackresearch methods AI Survey Design Prompts: Ask Better Questions 10 AI survey design prompts for objectives, neutral questions, response scales, pilots, analysis, and reporting—without turning biased feedback into facts. August 22, 2026 Updated August 22, 2026 Surveys look easy because everybody has answered one. Write a few questions, add some radio buttons, send a link, and wait for Truth to arrive in a spreadsheet. Then twelve coworkers interpret “often” twelve different ways and the team calls a 4% response rate “the voice of the customer.” AI survey design prompts can help turn a fuzzy business question into a research objective, draft neutral wording, critique answer options, organize a pilot, and summarize de-identified responses. They cannot make a convenience sample representative, remove bias by rewriting one adjective, infer what respondents secretly meant, or decide what the organization should do. AI can help edit the questionnaire. Humans still choose the research method, recruit responsibly, protect respondents, validate the analysis, and own the decision. These ten templates are for researchers, product managers, marketers, customer-experience teams, HR teams, nonprofit staff, and founders who need useful evidence rather than decorative percentages. For broader desk research, use AI research prompts. For quantitative cleanup after collection, start with AI data analysis prompts—inside an approved environment. What a survey can and cannot tell you A survey collects self-reported answers from people who were invited, noticed the invitation, chose to respond, understood the questions, and felt able to answer. Every step creates possible distortion. A useful survey needs: One decision: what someone will do differently after seeing the results. A defined population: exactly whose experience or opinion matters. A defensible sample: how people are recruited and who may be missing. A clear construct: what each question is actually trying to measure. Neutral wording: no pressure toward the answer the sponsor wants. Sensible response options: complete enough to fit real people. A short pilot: observation of how respondents interpret the instrument. An analysis plan: written before interesting answers tempt the team. Transparent limits: what the results do not support. Surveys are useful for reported attitudes, preferences, recollections, intentions, and experiences. They are weaker at explaining observed behavior, proving causality, or representing a population that was poorly sampled. If you need to see where people actually struggle, pair the survey with usability testing. If you need causal evidence about a product change, consider a properly reviewed A/B test. The reusable survey design prompt formula Use this instruction with any template below: “Act as a survey-design assistant. I am researching [decision, target population, research objective, recruitment method, constraints, and deadline]. Use only the verified information I provide. Produce [artifact] with assumptions, risks, exclusions, open questions, privacy considerations, accessibility checks, and required human review. Separate supplied facts, respondent reports, analysis, interpretations, and decisions. Do not invent respondents, response rates, sample sizes, quotes, benchmarks, statistical significance, or representative claims.” That separation matters. “Forty of 100 respondents selected option B” is an observation. “Customers prefer B” is an inference requiring a defensible sample, clean question, and appropriate analysis. “Build B” is a decision involving costs, strategy, ethics, and judgment. Never paste respondent names, email addresses, employee IDs, health or financial details, credentials, raw survey exports, confidential comments, or demographic combinations that could identify someone into an unapproved AI tool. Remove free-text details carefully; deleting a name does not make a story anonymous. Use synthetic examples, approved systems, access controls, documented retention, and human privacy review. What to collect before prompting “Make me a customer survey” gives the model room to manufacture your research plan. Provide a compact, verified brief instead. InputWhy it mattersHuman check Decision to informPrevents curiosity theaterAccountable owner confirms Research objectiveDefines what must be learnedResearch lead reviews Target populationSets the boundary of claimsDomain owner validates Recruitment methodReveals selection riskResearcher documents Known evidenceAvoids asking settled questionsEvidence owner verifies Sensitive topicsTriggers consent and safety reviewPrivacy/legal reviews Delivery channelAffects length and accessibilityOperations owner tests Analysis planReduces post-hoc storytellingAnalyst approves Timeline and budgetKeeps the design realisticProject owner confirms Decision thresholdExplains how evidence will be usedDecision owner signs off Do not ask for data merely because it might be interesting later. Every additional question costs attention and creates more personal information to protect. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI survey design prompts Replace brackets with sanitized, verified details. These templates draft artifacts; they do not certify the research. 1. Turn a business question into a research objective “Using this proposed decision, current evidence, stakeholder questions, target population, constraints, and unknowns: [paste], draft three possible research objectives. For each, state the construct to measure, why a survey is or is not suitable, required population, likely limitations, and what decision the result could inform. Reject objectives that ask a survey to prove causality or predict behavior with certainty.” “Do customers like our product?” is not an objective. Which customers, which experience, what does “like” mean, and what changes if the answer is no? A useful objective might be: “Estimate how recently activated administrators rate setup clarity, identify commonly reported obstacles, and decide which three issues deserve follow-up interviews.” It has a population, topic, output, and decision. A researcher still decides whether a survey is the right method. 2. Define the audience and screening criteria “Given this target population, decision, product eligibility rules, geography, time window, recruitment channels, exclusions, and known coverage gaps: [paste], draft a recruitment and screening brief. Distinguish the target population, sampling frame, invited sample, respondents, and analysis sample. Flag selection bias, nonresponse risk, duplicate responses, professional respondents, and groups likely to be underrepresented. Do not claim representativeness.” A customer email list excludes people who churned before joining it. An in-product intercept excludes people who cannot access the product. An employee Slack poll excludes silence, fear, leave, time zones, and anyone who muted the channel for emotional survival. Document who had a realistic chance to respond. If the sample is directional, say so. “Respondents to this survey” is honest. “Our customers” may be fiction. 3. Draft neutral questions “Using this research objective, constructs, audience language, reading-level goal, sensitive-topic rules, and prohibited assumptions: [paste], draft up to 12 concise survey questions. Use one concept per question. Avoid leading wording, loaded terms, absolutes, hypotheticals presented as facts, unnecessary jargon, and assumed experience. For each question, explain its purpose and list one possible misinterpretation for human review.” Questions should not contain the sponsor’s preferred answer. “How helpful was our improved dashboard?” smuggles in two claims: that it improved and that it helped. Ask about a specific task or experience instead. AI can spot textbook problems. It may miss political pressure, cultural meaning, product history, or why “simple” irritates people who struggled. Have humans from the target audience review the wording. 4. Choose consistent response scales “Review these draft questions and intended analyses: [paste]. Recommend a response format for each using only justified categories, such as frequency, agreement, satisfaction, ease, confidence, number, date, or open text. Provide balanced labels for every scale point, a suitable recall period, and options for not applicable, do not know, or prefer not to answer where needed. Flag mismatches between the question and scale.” Do not mix “strongly disagree” with a question asking frequency. Avoid unlabeled mystery numbers unless respondents understand the anchors. Keep scale direction consistent so people do not accidentally answer the opposite. More scale points do not automatically create more truth. The researcher should choose a scale that respondents can distinguish and that supports the planned analysis. 5. Find leading and double-barreled questions “Audit this questionnaire: [paste]. Create a table with question number, exact wording, issue, severity, likely effect, and a neutral rewrite. Check for leading language, double-barreled questions, hidden assumptions, absolutes, social-desirability pressure, emotionally loaded terms, unclear reference periods, jargon, negation, and questions that require knowledge respondents may not have. Do not silently rewrite; preserve the audit trail.” “How satisfied are you with the speed and accuracy of support?” asks two questions. A respondent may love the speed and hate the answer. One radio button cannot express both. An audit is a useful second pass, not validation. Pilot the revised instrument with real members of the audience and ask them to explain what each question means in their own words. 6. Improve answer options “For each closed-ended question and current answer list: [paste], check whether options are mutually understandable, sufficiently complete, consistently ordered, non-overlapping where appropriate, and aligned with the question. Identify missing ordinary answers, false binaries, unequal ranges, ambiguous ‘other,’ and cases needing ‘none,’ ‘not applicable,’ ‘do not know,’ or ‘prefer not to answer.’ Propose revisions and explain tradeoffs.” Age ranges that overlap at 35 are broken. Job-role lists that omit contractors distort the answer. “Other” can be useful, but a required free-text explanation can turn it into a privacy trap. Randomizing options can reduce order effects in some cases; it is wrong for ordered scales and may confuse respondents. Humans should approve the logic in the actual survey platform. 7. Create an accessible, mobile-friendly version “Review this survey for a respondent using a phone, keyboard, screen reader, magnification, limited bandwidth, or plain-language translation: [paste]. Produce an accessibility checklist covering question length, heading structure, labels, focus order, error messages, contrast, matrix questions, required fields, progress indicators, time estimate, save-and-return behavior, and alternatives to visual information. Mark every item unverified until tested.” A ten-column matrix may look efficient on a laptop and become thumb-operated punishment on a phone. Break complex grids into shorter questions when possible. Do not use color alone to carry meaning. Automated suggestions cannot replace testing with assistive technology and people with relevant access needs. Accessibility belongs in the design, not in an apology after launch. 8. Plan a small pilot “Using this draft survey, audience, delivery channel, risks, and deadline: [paste], create a pilot plan. Include recruitment criteria, a small purposeful range of participants, think-aloud or debrief questions, completion-time observation, comprehension checks, technical tests, privacy checks, stopping conditions, issue logging, owners, and revision rules. Do not invent a universally valid pilot size.” A pilot is not a miniature launch whose only metric is completion. Watch people interpret the questions. Ask what they thought each item meant and whether any answer option failed to fit. Fix severe confusion, broken logic, inaccessible controls, or unexpected disclosure risk before broader distribution. Record revisions so the team knows which version produced which data. 9. Build a de-identified analysis framework “Using this approved research objective, questionnaire, sampling notes, response coding, missing-data rules, and decision needs: [paste], draft an analysis plan before viewing outcomes. Map every question to an objective; define valid denominators, exclusions, weighting decisions requiring expert review, planned subgroup comparisons, open-text coding steps, uncertainty reporting, and limits. Ban post-hoc segment fishing and invented significance.” Write the plan early because dashboards encourage storytelling. If twenty subgroups are searched after collection, one surprising pattern will eventually appear by luck. Free-text responses need special care. Remove direct identifiers, review rare details, restrict access, and report themes without exposing respondents. AI-assisted coding requires human calibration, disagreement review, and documentation of the model and prompt used. 10. Summarize findings without overclaiming “Using only these approved aggregate tables, documented sample, field dates, recruitment method, analysis notes, limitations, and de-identified themes: [paste], draft a findings brief. Separate observed results, plausible interpretations, unanswered questions, and recommendations. State who responded, who may be missing, uncertainty, contradictory evidence, and what follow-up research is needed. Never fabricate quotes, causal claims, population claims, or confidence.” A readable report does not need fake certainty. Say “62% of 214 respondents recruited through the customer newsletter selected…” rather than “62% of customers believe…” when the sample does not justify the larger claim. Recommendations need accountable owners. Use a documented decision-making process and preserve the result in a decision log. The model can draft the memo. It does not absorb the consequences. A human review gate before launch Run a short review with research, privacy, accessibility, analytics, domain, and decision owners as appropriate. Does every question support the stated objective? Is a survey the right method for each claim? Can the target population realistically be reached? Are consent and sensitive-topic protections adequate? Could wording pressure, shame, confuse, or identify respondents? Do answer options fit ordinary real-world situations? Does skip logic work on mobile and assistive technology? Is the analysis plan written before results arrive? Are claims limited to what the sample supports? Is someone accountable for acting—or explicitly not acting—on the result? Use AI policy prompts if the organization lacks rules for approved tools, retention, or sensitive research data. For general safe operating habits, read the no-BS guide to using AI at work. Frequently asked questions Can ChatGPT write a whole survey for me? It can draft one, which is different from designing valid research. Give it a verified objective, audience, constraints, and review criteria. Then have a qualified human review the method, wording, privacy, accessibility, sampling, and analysis. Never launch raw model output because it looks polished. How many questions should a survey have? As few as needed to inform the decision. There is no universal number. Length depends on audience, topic, channel, burden, sensitivity, and value to the respondent. Pilot completion time and remove questions that are merely “nice to know.” Can AI calculate the required sample size? It can explain inputs or reproduce a calculation under supervision, but it should not choose assumptions by vibes. Sample planning depends on population, design, expected effect or precision, variability, clustering, subgroup needs, nonresponse, and analysis method. Use a qualified researcher or statistician. How do I stop AI from writing leading questions? Explicitly prohibit leading language and require an audit explaining each risk. Ask for alternative neutral wording. Then pilot with real people. Models can reproduce bias from the brief, especially when stakeholders describe their preferred conclusion as a fact. Is survey data safe to paste into an AI tool? Not by default. Raw exports can contain direct identifiers, sensitive free text, timestamps, rare demographics, and combinations that re-identify people. Use approved tools and data-handling rules. Prefer aggregated or synthetic data, minimize fields, restrict access, and involve privacy or security owners. Can AI analyze open-ended survey responses? It can assist with a documented coding process in an approved environment. Humans should build and test the codebook, review disagreements, inspect minority themes, check whether summaries erase criticism, and retain traceability to de-identified evidence. Do not present generated themes as objective truth. Does a large response count make a survey representative? No. A huge biased sample is still biased. Representation depends on who could be selected, who was invited, who responded, coverage gaps, nonresponse, weighting, and the claim being made. Report the recruitment method and limits alongside the count. When should I use interviews instead of a survey? Use interviews or observation when you need depth, context, language, motivations, or understanding of an unfamiliar problem. Use a survey when the constructs and answer space are sufficiently understood and structured responses can inform a defined decision. Often the best sequence is interviews first, survey second. The point is better evidence, not faster forms AI is genuinely useful at the boring parts: generating alternatives, checking consistency, building review tables, and turning notes into an organized draft. That speed matters. It also makes it easier to launch a bad survey before anyone asks whether the sample, question, or decision makes sense. The durable advantage is human judgment: choosing a worthwhile question, respecting respondents, noticing context, and resisting the convenient conclusion. If you need a broader field guide for keeping that advantage while using the tools, Don’t Replace Me by Dmitry Kargaev is the low-drama version. Use the machine to sharpen the instrument. Do not let it manufacture the evidence. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Customer Interview Prompts: Learn Without Leading 10 AI customer interview prompts for research goals, neutral questions, follow-ups, notes, synthesis, and findings—without manufacturing customer truth. AI promptscustomer interviewscustomer discovery AI Customer Journey Mapping Prompts: Find the Friction 10 AI customer journey mapping prompts to organize evidence, map touchpoints, find friction, plan research, and turn journey maps into accountable action. AI promptscustomer journey mappingservice design AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation --- ## AI Test Case Prompts: Turn Requirements Into Checks URL: https://dontreplace.me/articles/ai-test-case-prompts-turn-requirements-into-checks AI promptstest casesquality assurancesoftware testingrequirements AI Test Case Prompts: Turn Requirements Into Checks AI test case prompts for turning verified requirements into reviewable scenarios, expected results, test data, evidence, and human-owned decisions. August 09, 2026 Updated August 09, 2026 A requirement walks into a sprint and says, “Users should be able to update their profile.” It sounds harmless. Then somebody asks which users, which fields, which permissions, which validation rules, what happens during an outage, and whether changing an email address should trigger verification. Suddenly the one-line requirement has grown teeth. AI test case prompts can help turn verified requirements into structured candidate checks. They can suggest happy paths, negative scenarios, boundaries, role variations, data needs, and missing questions. They cannot inspect your product, discover undocumented behavior, execute a test, verify an outcome, or decide whether a risk is acceptable. AI can draft the test map. Humans still own the territory, the evidence, and the decision about whether anything actually works. If your team first needs an overall strategy, use AI test plan prompts. If the scope is clear and you need concrete, reviewable checks, the 10 templates below are built for that job. What a useful test case contains A test case is more than a sentence beginning with “verify.” It connects a source requirement to a repeatable action and an observable result. A reviewer should understand what is being checked, under which conditions, with which data, and what evidence distinguishes a pass from wishful thinking. Useful test cases usually include: a stable ID and source requirement or risk; scope, purpose, and priority assigned by humans; roles, permissions, preconditions, and environment; privacy-safe representative test data; clear steps or an unambiguous scenario; an observable expected result; evidence requirements and an owner; cleanup needs, dependencies, and open questions. “Test login” is not a test case. “For an active member with a verified email, submit valid credentials in the supported browser; confirm the account dashboard loads, a session is created according to the documented policy, and no other account’s data appears” is closer. It still needs real product rules, safe data, execution, and evidence. If the acceptance criteria are vague, fix that before multiplying them into cases. AI acceptance criteria prompts can help expose ambiguity, but the responsible product and technical people must settle the intended behavior. The reusable AI test case prompt formula Use this foundation before any specialized template: “Act as a test-case drafting assistant. I am [role] reviewing [feature, workflow, fix, integration, or change]. Use only the verified, privacy-safe sources below. Draft candidate test cases with source traceability, purpose, preconditions, role, safe data needs, steps, observable expected results, evidence, owner fields, and unresolved questions. Separate confirmed behavior from suggestions. Do not invent requirements, system behavior, data, results, severity, owners, dates, or approval. Label unsupported assumptions ‘needs human confirmation.’ Identify contradictory or untestable requirements instead of silently resolving them.” The guardrails matter. Models are excellent at making a guessed rule look like it arrived from a product specification carved into stone. Require source references and visible uncertainty. Never paste credentials, access tokens, customer PII, private tickets, raw production logs, security vulnerabilities, regulated data, unreleased strategy, legal disputes, HR records, or confidential client material into an unapproved AI tool. Use approved systems, redacted summaries, synthetic records, and references to controlled evidence. What to collect before generating test cases Do not ask AI to “write all the tests” from a feature name. A feature name is a filing label, not evidence. InputWhy it mattersHuman validation Approved requirementsDefines intended behaviorResolve conflicts and stale language Acceptance criteriaCreates observable conditionsConfirm each criterion is testable Workflow or process mapShows sequence and handoffsVerify current roles and exceptions Role and permission matrixPrevents authorization guessworkConfirm least-privilege rules Interface and dependency notesReveals integration boundariesVerify contracts with owners Validation and error rulesDefines negative behaviorConfirm exact messages only when required Supported environmentsBounds credible coverageRecord parity gaps and exclusions Test-data policyKeeps testing representative and safeApprove creation, retention, and cleanup Known risks and prior defectsFocuses valuable coverageConfirm relevance to current scope Evidence standardDefines a credible pass or failureMatch evidence depth to impact When information is missing, have AI produce questions. A blank cell is honest. An invented expected result is a future bug report wearing office clothes. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI test case prompts Replace bracketed text with verified, privacy-safe information. Treat every output as a draft requiring human review. 1. Convert acceptance criteria into traceable test cases “Review these approved requirements and acceptance criteria: [paste]. Create a traceability table with source ID, exact criterion, candidate test-case ID, purpose, preconditions, role, input category, steps, observable expected result, evidence, and open question. Preserve the source wording. Flag criteria that are ambiguous, contradictory, compound, or not observable. Do not invent behavior to make a criterion testable.” This is the cleanest starting point because it keeps every case attached to an authorized source. It also reveals the requirements that sound clear until somebody tries to observe the result. Do not measure quality by case count. One precise scenario can cover a coherent rule better than six nearly identical cases generated to make a spreadsheet look busy. 2. Draft happy-path checks “Using these verified requirements, supported workflow, and approved example data: [paste], draft the minimum happy-path test cases. For each, include user goal, valid preconditions, role, safe input, steps, expected system state, expected user-visible result, downstream record or event if documented, and required evidence. Cite the supplied source for every expected result. Do not assume success behavior that is not documented.” Happy paths confirm that the intended journey works under normal conditions. They are necessary, but they are not a testing strategy. A model naturally produces neat success stories because neat success stories are easy to write. Ask reviewers whether each case proves a meaningful outcome. Loading a confirmation page may not prove that inventory changed, an email was queued, or a downstream record was created. 3. Find negative and failure scenarios “Review this verified workflow, validation policy, dependency information, and failure-handling documentation: [paste]. Draft candidate negative cases for invalid input, missing input, unauthorized action, unavailable dependency, interrupted request, duplicate submission, stale state, and documented business-rule rejection. Include the source, trigger, expected safe behavior, user feedback, data-integrity expectation, retry or recovery rule if supplied, and evidence. Mark undocumented failure behavior as a question.” Negative testing asks whether the product fails safely and intelligibly. It should not become a carnival of random bad inputs. Prioritize realistic misuse, known failure modes, high-impact data risks, and boundaries supported by the product’s rules. For discovered failures, AI bug report prompts can structure observations without pretending the first theory is the root cause. 4. Identify boundary-value cases “Using these verified field rules, ranges, limits, date rules, file constraints, and business thresholds: [paste], create a boundary-analysis table. For each documented boundary, propose cases immediately below, at, and immediately above it where meaningful. Include units, inclusivity, format, locale or timezone dependency, expected result, source, and evidence. Do not invent numeric limits or treat an example value as a formal boundary.” Boundaries attract defects because code must switch behavior somewhere. Minimum and maximum lengths, cutoff times, quantity limits, pagination, money precision, file sizes, and date transitions all deserve attention when they actually apply. Human reviewers must confirm units and semantics. “30 days” can mean calendar days, business days, or a duration measured to the second. AI cannot settle a policy dispute by choosing the most statistically familiar answer. 5. Create role and permission cases “Using this approved role-permission matrix and workflow: [paste], draft positive and negative authorization cases. Include actor role, resource ownership, initial state, attempted action, permitted or denied expectation, user-visible response, audit requirement if documented, data-exposure check, source, and evidence. Include cross-account and changed-role scenarios only when relevant. Do not invent permissions or expose real account data.” Permissions deserve explicit cases because “the button is hidden” does not prove the action is forbidden. Coverage may need both interface and service-level evidence, reviewed by qualified technical or security owners. Use synthetic accounts and approved environments. Never paste real credentials or customer records into a prompt. The chatbot does not need production access to help format a matrix. 6. Check validation and error handling “Review these approved validation rules, error categories, content guidelines, and recovery rules: [paste]. Draft test cases for required fields, malformed values, conflicting values, expired state, duplicate actions, and recoverable versus nonrecoverable errors. Include input category without sensitive values, trigger, expected handling, preservation or rollback of data, user guidance, accessibility consideration, logging or alert expectation if supplied, and evidence. Quote exact error text only when the source requires it.” Good error handling protects state and helps people recover. A red message is not automatically useful, and a generic success response can conceal a partial failure. Avoid brittle assertions against punctuation unless exact copy is contractually, legally, or operationally significant. Test the behavior and meaning that matter, then add exact text checks where humans explicitly require them. 7. Cover integrations and data flows “Using this approved interface contract, dependency map, and data-flow summary: [paste], draft candidate integration test cases. Include trigger, source, destination, payload category, transformation, expected response, persistence, timeout or retry behavior if documented, duplicate handling, failure path, observability evidence, environment limitation, and owner fields. Never invent endpoints, schemas, status codes, retry rules, or downstream effects.” Integrations produce elegant diagrams and ugly incidents. Data can be accepted at one boundary, transformed incorrectly, duplicated by retry logic, or lost after a cheerful success message. Keep real payloads and secrets out of prompts. Use approved schemas and synthetic examples. When dependencies are unclear, AI dependency mapping prompts can organize known relationships, but owners must verify the map. 8. Prepare test data and preconditions “From these approved data rules, role definitions, environment notes, and scenario list: [paste], create a test-data and precondition matrix. Include case ID, required account state, role, record relationships, safe synthetic values, setup method, environment dependency, reset or cleanup method, collision risk, prohibited data, owner field, and readiness question. Do not generate real personal data, credentials, secrets, or production identifiers.” A case that depends on mysterious existing data is not reliably repeatable. Define setup and cleanup so another tester can recreate the conditions without borrowing an account named “final_test_REAL_7.” Representative data should preserve relevant shape and variation without copying sensitive production records. Privacy, security, and compliance owners decide what is allowed; AI does not grant exceptions. 9. Turn a defect into retest cases “Review this privacy-safe defect record, confirmed requirement, observed evidence, approved fix summary, and impact notes: [paste]. Draft a focused retest case for the reported failure and candidate adjacent regression cases. Separate observed facts from suspected cause. Include original conditions, fixed behavior to verify, unchanged behavior to protect, environments, data, expected results, evidence, and open questions. Do not claim the defect is fixed or assign root cause from the summary.” A retest reproduces the reported problem and checks the intended correction. Regression checks ask whether the correction damaged neighboring behavior. Keep those purposes visible instead of dumping both into “verify fix.” If several failures arrive together, AI defect triage prompts can help group evidence. Humans still assign severity, ownership, and release consequences. 10. Review a test-case set for gaps and duplication “Audit this draft test-case set against the supplied requirements, risk list, role matrix, workflow, interfaces, and supported environments: [paste]. Produce four sections: traced coverage, untraced cases, uncovered sources, and suspected duplicates. Also flag missing expected results, hidden assumptions, unsafe data, unclear preconditions, nonobservable outcomes, unsupported priorities, and cases that test implementation details rather than behavior. Do not claim complete coverage. End with questions for human reviewers.” This prompt is useful after drafting because generation creates clutter as easily as coverage. Duplicate cases waste execution time; untraced cases can preserve obsolete behavior; missing cases leave a requirement protected only by optimism. Use AI QA checklist prompts for the broader quality picture. Test cases are one layer among requirement review, exploratory testing, accessibility, security, performance, operations, monitoring, and human judgment. A practical human-owned workflow The prompts work best in sequence: Confirm the current approved requirements and release scope. Resolve contradictory or unobservable acceptance criteria. Identify critical workflows, roles, dependencies, and risks. Draft traceable happy-path, negative, and boundary cases. Add permission, validation, integration, and data coverage. Review cases with product, engineering, QA, and specialists. Prepare approved environments and synthetic test data. Execute in the real testing system and capture evidence. Triage failures without inventing causes. Update regression coverage and record residual risk. After a change ships, use AI regression testing prompts to connect that verified change to existing behavior that still needs protection. Common mistakes with AI-generated test cases Treating fluent output as product knowledge The model knows common software patterns, not your undocumented rule. If it says an account locks after five attempts, require a source. Familiar behavior is still invented behavior when your specification never said it. Generating hundreds of cases before fixing requirements Automation makes duplication cheap. It does not make ambiguity safe. Resolve the rule first, then generate only the coverage that reviewers can understand, execute, and maintain. Using vague expected results “Works correctly” and “shows an error” are not observable enough. Define the state, response, record, message category, or evidence that proves the outcome without overspecifying irrelevant implementation details. Pasting sensitive material into an unapproved tool Do not solve a testing problem by creating a privacy or security problem. Redact, summarize, synthesize, and keep controlled evidence in approved systems. Letting AI assign risk or release approval A model can format supplied risk labels. It cannot understand customer commitments, operational consequences, architecture, law, or organizational appetite well enough to accept risk. Named humans own those calls. Frequently asked questions Can AI write complete test cases from a requirement? It can draft candidate cases from a sufficiently detailed, verified requirement. It cannot know hidden behavior, inspect the product, confirm that the source is current, execute the case, or prove completeness. Human review and real evidence remain mandatory. What is the best prompt for generating test cases? The best prompt supplies approved requirements, roles, preconditions, data rules, risks, and expected behavior; requires source traceability; separates facts from suggestions; and forbids invented results. The reusable formula near the top of this article is a solid starting point. Should every acceptance criterion have one test case? Not necessarily. One criterion may require several scenarios, while one coherent end-to-end case may cover parts of several criteria. Preserve traceability, then choose the smallest maintainable set that gives credible risk-based evidence. Can ChatGPT generate negative test cases? Yes, as candidates. Give it verified validation, permission, dependency, and failure-handling rules. Review suggestions against real behavior and risk instead of accepting a generic parade of blank fields and invalid characters. How do I stop AI from inventing expected results? Require a source reference for every expectation, tell it to mark unsupported behavior as “needs human confirmation,” and reject any case whose result cannot be traced to an approved rule. Do not ask the model to fill missing requirements. Is it safe to paste bug tickets or logs into AI? Not by default. Tickets and logs may contain customer data, credentials, vulnerabilities, internal identifiers, or confidential context. Use approved tools, redact aggressively, and provide synthetic or summarized inputs whenever possible. Who should review AI-generated test cases? The appropriate mix of product, QA, engineering, operations, security, privacy, compliance, accessibility, and domain owners should review them based on risk. The person accountable for the behavior must remain visible. Do AI-generated test cases replace exploratory testing? No. Structured cases confirm known expectations. Exploratory testing uses human curiosity, product understanding, and observation to find risks the documented expectations missed. Both can matter, and neither proves absence of defects. The useful boundary AI is good at transforming messy verified inputs into tables, repeating a structure consistently, proposing ordinary variations, and pointing at blanks. It is bad at knowing whether the source is true, whether the environment is credible, whether an observed result matters, or whether the team should ship. Use it to accelerate drafting and review. Keep humans on product truth, sensitive context, execution, evidence, risk, specialist judgment, and accountability. That boundary is less glamorous than “autonomous testing,” but it is dramatically less likely to turn a polished spreadsheet into a production incident. For a broader field guide to using AI without outsourcing your judgment, Don’t Replace Me by Dmitry Kargaev follows the same principle: speed is useful; responsibility is still yours. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance AI API Testing Prompts: Catch Broken Integrations AI API testing prompts for contracts, edge cases, authorization, retries, error responses, and human-verified integration evidence. AI promptsAPI testingquality assurance AI Compatibility Testing Prompts: Catch Breakage 10 AI compatibility testing prompts for browser, device, OS, responsive, locale, network, defect triage, and evidence-based release decisions. AI promptscompatibility testingcross-browser testing --- ## AI Test Data Prompts: Build Safer Scenarios URL: https://dontreplace.me/articles/ai-test-data-prompts-build-safer-scenarios AI promptstest datasynthetic dataquality assurancesoftware testing AI Test Data Prompts: Build Safer Scenarios AI test data prompts for designing synthetic records, boundary cases, privacy guardrails, reusable datasets, and human-reviewed testing plans. August 10, 2026 Updated August 10, 2026 Your test environment needs a customer record. Somebody copies one from production, changes the name to “Testy McTestface,” and declares the privacy problem solved. It is not solved. The email, order history, support notes, device ID, and oddly specific delivery address are still waving from the spreadsheet. AI test data prompts can help teams design safer synthetic records from approved schemas and documented rules. They can organize fields, suggest categories, draft obviously fictional examples, and expose missing constraints. They cannot certify anonymization, inspect your database, know undocumented relationships, reproduce real-world distributions, or approve a privacy decision. AI can help draft the fake data. Humans must prove it is valid, useful, and actually fake. If you are still deciding what must be tested, start with AI test case prompts or build the broader strategy with AI test plan prompts. Once the scenarios are known, use the templates below to design data without casually feeding a chatbot your customer table. What useful synthetic test data actually does Good test data creates a controlled condition. It gives a tester the records, relationships, roles, states, and values needed to execute a scenario and observe a meaningful result. It should be repeatable, resettable, traceable to a documented rule, and safe for the approved environment. Useful test data usually includes: a scenario or test-case ID; the approved schema and field constraints; clearly fictional identifiers and contact details; valid, invalid, boundary, and missing-value categories; required relationships between records; role, permission, account, and lifecycle state; setup, reset, retention, and deletion instructions; a data owner and privacy or security reviewer where needed; validation evidence and unresolved questions. A random list of names is not a test-data strategy. Neither is a production export with the obvious columns hidden. Representative data preserves the shape and rules needed for testing without preserving the identity or confidential substance of real people. That distinction matters. Replacing dee@example.com with alex@example.com may be pseudonymization at best, and only if every linked identifier and free-text field is handled correctly. It is not automatically anonymous. Qualified humans must assess the actual data, system, threat model, and legal obligations. The reusable AI test data prompt formula Use this as a foundation before the specialized prompts: “Act as a synthetic test-data planning assistant. I am [role] preparing data for [approved environment, feature, workflow, or test set]. Use only the verified, privacy-safe schema, rules, and scenarios below. Create candidate synthetic data specifications with field purpose, type, allowed values, constraints, relationships, scenario coverage, setup method, reset method, and validation checks. Use obviously fictional placeholders. Separate confirmed rules from suggestions. Do not reproduce real people, infer production values, invent system behavior, claim anonymization, or approve privacy, security, legal, compliance, or release decisions. Label unsupported assumptions ‘needs human confirmation.’” The phrase “obviously fictional” is doing work. A generated address that happens to belong to a real person is not a charming coincidence. Use reserved domains such as example.com, documented fictional ranges where available, non-routable or approved test numbers, and internal conventions reviewed by the people responsible for privacy and security. Never paste real customer PII, credentials, access tokens, production exports, payment details, health information, HR records, private tickets, raw logs, vulnerability details, regulated data, or confidential client material into an unapproved AI tool. Redaction is a process, not a vibe. What to collect before prompting A model cannot infer the safe dataset from a feature name. Gather approved inputs first. InputWhy it mattersHuman validation Schema or field dictionaryDefines available fields and typesRemove secrets and confirm current version Business rulesDefines valid combinations and statesResolve contradictions and exceptions Test cases or scenariosConnects records to actual checksConfirm scope and priority Relationship mapPreserves parent, child, and dependency rulesVerify cardinality and deletion behavior Role and permission matrixSupports access-control scenariosConfirm least-privilege expectations Data classificationIdentifies sensitive categoriesPrivacy and security owners approve handling Environment constraintsDefines setup and available integrationsConfirm isolation and production boundaries Reset and retention policyKeeps tests repeatable and cleanName an owner and deletion schedule Localization requirementsCovers formats, scripts, and regionsVerify supported locales and rules Validation methodProves records match documented constraintsKeep evidence in the approved system When a rule is missing, ask AI to produce a question instead of a value. “Unknown maximum length” is useful. A confident 255-character guess is a bug seed with excellent formatting. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI test data prompts Replace bracketed text with verified, privacy-safe information. Every output is a draft for human review. 1. Map the required data fields “Review these approved test cases, schema notes, and business rules: [paste]. Create a field inventory with field name, business purpose, type, required or optional status, allowed category, relationship, sensitivity classification supplied by us, scenarios that use it, and unresolved question. Preserve source references. Do not invent fields, defaults, classifications, or constraints.” Start here when the team has test cases but no shared view of the data they require. The inventory exposes a common mess: the scenario expects a suspended account, but nobody documented which field or workflow creates that state. Keep sensitive fields visible as categories, not values. “Payment token required from approved sandbox fixture” is safer and more useful than asking a public model to imitate a real token. 2. Create happy-path synthetic records “Using this approved schema, valid-value rules, and happy-path scenarios: [paste], draft a minimal set of obviously fictional records. For each record, include scenario ID, field-value category, relationship, reason the value is valid, setup instruction, expected lifecycle state, and validation check. Use reserved domains and approved fictional conventions. Do not claim the records have been inserted or accepted by the system.” Happy-path records should prove the normal workflow without creating ten decorative variations. Prefer the smallest dataset that covers meaningful combinations and is easy to understand. The human reviewer still checks every field against the actual validation rules. A plausible postal code or date format can be wrong for the supported locale, even when it looks perfectly respectable in a table. 3. Generate boundary-value records “From these verified numeric limits, text lengths, date rules, file limits, and business thresholds: [paste], create a boundary-data matrix. For each documented boundary, propose safe synthetic values immediately below, at, and immediately above it where meaningful. Include units, inclusivity, locale or timezone dependency, expected validity category, source, and test-case link. Do not invent a limit or treat an example as a formal rule.” Boundaries are where software changes its mind. Minimum quantities, maximum lengths, age thresholds, cutoff times, pagination limits, money precision, and date windows all deserve intentional records when they apply. Humans must confirm semantics. “One month” might mean a calendar transition, 30 days, or a billing period. The model cannot settle that dispute by choosing the most common implementation. 4. Draft invalid and negative records “Using these approved validation and business rules: [paste], draft candidate invalid data categories for missing required values, malformed formats, disallowed combinations, stale states, duplicate identifiers, broken references, and prohibited transitions. Include the violated rule, synthetic example pattern, expected rejection category if documented, data-integrity expectation, cleanup need, and evidence. Do not include attack payloads or invent exact error text.” Negative data should target real rules and credible mistakes. Random corruption creates noise; documented invalid states create evidence. Keep security testing separate and controlled. Do not ask a general-purpose model to generate exploit payloads from private vulnerability details. Qualified security owners choose the approved tools, environment, scope, and evidence requirements. 5. Build role and permission datasets “Using this approved role-permission matrix and resource-ownership model: [paste], design a synthetic account and resource dataset for positive and negative authorization tests. Include fictional account ID, role, ownership relationship, resource state, permitted action category, denied action category, setup, reset, and evidence. Do not generate credentials, real identities, or permissions absent from the matrix.” Authorization testing needs more than an admin and a user. It may need resource owners, non-owners, changed roles, disabled accounts, tenant boundaries, and records shared under documented rules. A hidden button does not prove authorization. The dataset supports testing; the real application and service-level evidence determine whether access is actually enforced. 6. Cover dates, time zones, and lifecycle states “Review these documented date, time, timezone, expiration, scheduling, and lifecycle rules: [paste]. Create synthetic scenarios for ordinary dates, supported timezone differences, daylight-saving transitions if relevant, leap dates if supported, cutoff boundaries, expired records, future records, and invalid state transitions. Include the exact supplied rule, fictional timestamp pattern, expected category, environment assumption, and question. Do not invent timezone behavior.” Time data is where “works on my machine” starts collecting passports. Store and display rules, user timezone, server timezone, business calendar, and external integrations can all differ. Use only scenarios relevant to the product. Generating every famous date edge case is not rigor; it is a trivia collection. Trace each record to a rule, risk, or prior failure. 7. Model relationships and dependencies “Using this approved entity relationship map, interface contract, and test scope: [paste], draft a compact synthetic dataset that preserves required parent-child, many-to-many, ordering, and dependency relationships. Include record aliases, relationship type, creation order, deletion or cleanup order, dependency owner, integrity validation, and unsupported assumptions. Never invent endpoints, foreign keys, or downstream effects.” Many failures live between records. An order without a customer, a child created before its parent, a deleted object still referenced downstream, or a duplicate event can reveal more than another standalone valid row. If ownership is unclear, use AI dependency mapping prompts before pretending the graph is complete. AI can format known relationships; system owners verify them. 8. Create localization and accessibility variants “Using these approved locales, character support rules, display constraints, input methods, and accessibility requirements: [paste], propose synthetic data variants for supported scripts, name lengths, address formats, decimal and date formats, right-to-left content if supported, screen-reader labels, and meaningful empty states. Include source, scenario, safe fictional pattern, expected category, and reviewer. Do not invent supported locales or use stereotypes as data.” Localization is not changing “color” to “colour.” Data shape varies across names, addresses, currencies, scripts, reading directions, and calendars. Accessibility can also depend on the quality and length of labels, errors, and state descriptions. Human reviewers with relevant expertise decide which variants are representative and respectful. A model trained on internet patterns can reproduce bias with astonishing confidence. 9. Prepare a resettable test-data pack “Turn these approved synthetic records and setup rules into a resettable test-data pack specification. Include pack version, covered scenarios, prerequisite fixtures, creation order, validation queries or checks supplied by us, ownership, environment, reset sequence, retention, deletion evidence, collision prevention, and rollback instructions. Mark every command or query requiring technical review. Do not claim execution.” Reusable packs prevent each tester from building a private universe of mystery accounts. Version them, document their purpose, and make reset behavior explicit. The model can draft the runbook. Engineers and environment owners review commands, execute them, and confirm that cleanup does not remove shared fixtures or leave sensitive debris. 10. Audit a dataset for gaps and privacy risk “Audit this candidate synthetic dataset against the supplied schema, scenarios, business rules, data classification, and environment policy: [paste]. Report traced coverage, uncovered scenarios, invalid combinations, missing relationships, duplicates, accidental realism, direct identifiers, free-text leakage risk, unsupported assumptions, reset gaps, retention gaps, and required human reviews. Do not certify anonymization, privacy compliance, security, or complete coverage.” Run this after drafting, not as a magic safety stamp. It can catch an example.com email paired with a real phone number, a supposedly fictional record containing copied support notes, or a dataset that cannot be reset. For the wider release picture, AI QA checklist prompts can organize checks beyond data: requirements, behavior, accessibility, security, operations, monitoring, and sign-off. A practical human-owned workflow Use the prompts in a controlled sequence: Confirm the current test scope and approved environment. Collect schemas, rules, relationships, and data classifications. Remove secrets and production values before prompting. Map fields and scenario requirements. Draft the smallest useful synthetic dataset. Add boundaries, negative states, roles, dates, and relationships. Review with QA, engineering, privacy, security, and domain owners as risk requires. Validate records against documented rules in the real test system. Execute tests and capture evidence in approved tools. Reset or delete data according to policy, then record proof. When a change modifies existing behavior, pair the dataset with AI regression testing prompts so the team protects what should remain unchanged. Common mistakes with AI-generated test data Calling realistic data anonymous Changing a name does not remove identity from linked fields, rare combinations, timestamps, notes, or external IDs. Never let a model certify anonymization. Use approved synthetic-data practices and qualified review. Generating values before confirming constraints A model will happily fill every blank. Require source references and questions instead. Unknown allowed values, relationships, defaults, and lifecycle rules must remain visibly unknown until humans resolve them. Optimizing for dataset size A million generated rows can still miss the one relationship or threshold that matters. Start with scenario coverage and risk. Add volume only when the test purpose requires it and the environment can handle it safely. Forgetting reset and ownership Synthetic data becomes operational clutter when nobody knows who created it, which tests depend on it, or whether it can be deleted. Every reusable pack needs an owner, version, setup path, and cleanup path. Pasting production examples into the prompt “Use this real record as inspiration” defeats the point. Describe approved structure and constraints without supplying customer substance. If the model does not need the value, it should not receive the value. Treating generated output as execution evidence A CSV in a chat window proves that text was generated. It does not prove the application accepted the records, enforced relationships, rejected invalid states, or cleaned them up. Real execution and observed evidence remain human-owned. Frequently asked questions Can ChatGPT generate synthetic test data? Yes, it can draft candidate fictional values and structured datasets from an approved schema and explicit rules. Humans must verify validity, privacy, relationships, representativeness, and suitability for the actual environment. Is AI-generated test data automatically anonymous? No. “Synthetic” and “anonymous” are not magic labels. Generated content can resemble real identities, preserve copied production details, or create combinations that identify somebody. Privacy and security reviewers must assess the real process and output. What should I include in an AI test data prompt? Include privacy-safe schemas, field rules, relationships, scenarios, environment constraints, fictional-value conventions, reset requirements, and validation checks. Require source traceability and forbid invented rules, real identities, credentials, and claims of compliance. Can I paste a production sample with names removed? Not by default. Other fields, free text, timestamps, device identifiers, account history, and rare combinations can remain sensitive or identifying. Follow your organization’s approved data-handling and de-identification process instead of improvising in a chatbot. How do I make synthetic test data representative? Start from documented categories, constraints, distributions, workflows, and known risks that humans are allowed to use. Cover meaningful variations and boundaries. Do not ask AI to guess the production population or manufacture demographic realism. What domains and contact details should fictional records use? Use organization-approved conventions and reserved resources such as example.com where appropriate. Avoid random plausible addresses, phone numbers, and identifiers that might belong to real people or route to real systems. Who should review AI-generated test data? QA and engineering should verify technical validity. Product and domain owners should verify business meaning. Privacy, security, legal, compliance, accessibility, or data specialists should review when the risk and data categories require them. Does synthetic data replace production-like testing? It can reduce exposure and support repeatable scenarios, but it may not reproduce every production distribution, integration, performance condition, or historical oddity. Humans decide which additional controlled testing is necessary and what data is permitted. The useful boundary AI is useful for turning approved rules into tidy tables, producing ordinary fictional variations, repeating schemas, and spotting blanks. It is unreliable at knowing whether a dataset reflects the real product, whether a value is safe, whether anonymization succeeded, or whether a privacy risk is acceptable. Use it for drafting. Keep humans on data classification, sensitive context, validation, execution, specialist review, evidence, cleanup, and accountability. That boundary is less exciting than “generate production-like data instantly,” but it is much less likely to create a reportable incident while the team is trying to prevent a bug. For a broader field guide to using AI without outsourcing judgment, Don’t Replace Me by Dmitry Kargaev follows the same durable rule: let the machine provide speed; keep responsibility attached to people. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance AI API Testing Prompts: Catch Broken Integrations AI API testing prompts for contracts, edge cases, authorization, retries, error responses, and human-verified integration evidence. AI promptsAPI testingquality assurance AI Compatibility Testing Prompts: Catch Breakage 10 AI compatibility testing prompts for browser, device, OS, responsive, locale, network, defect triage, and evidence-based release decisions. AI promptscompatibility testingcross-browser testing --- ## AI Test Plan Prompts: Turn Risk Into Strategy URL: https://dontreplace.me/articles/ai-test-plan-prompts-turn-risk-into-a-real-strategy AI promptstest planningquality assurancesoftware testingrisk AI Test Plan Prompts: Turn Risk Into Strategy AI test plan prompts for turning requirements, release risks, workflows, and known constraints into a reviewable testing strategy with clear evidence. August 07, 2026 Updated August 07, 2026 A test plan should explain what you are testing, why it matters, how you will test it, who owns the work, and what evidence is required before release. Instead, many teams get a spreadsheet named FINAL_test_plan_v7_REAL.xlsx, three vague tickets, and a meeting where everyone says, “QA has it.” AI test plan prompts can help turn scattered requirements, risk notes, acceptance criteria, and release scope into a reviewable testing strategy. They are useful for organizing supplied facts, drafting scenarios, exposing missing inputs, and formatting coverage matrices. They cannot inspect your product, execute a test, know undocumented behavior, verify a result, or decide whether a release is safe. AI can draft the testing map. Humans still own the environments, evidence, defects, risk decisions, and sign-off. If the work is too vague to test, start with AI acceptance criteria prompts. If you have a real change and need a disciplined plan for testing it, keep reading. What is a test plan? A test plan is a shared operating document for validating a product, system change, workflow, migration, campaign, or release. It connects scope and risks to test scenarios, owners, environments, data, evidence, and release criteria. A useful test plan answers: What is in scope and explicitly out of scope? Which user journeys and business operations matter most? What can fail, and how serious would each failure be? Which test types and environments are required? What data, accounts, integrations, devices, or permissions are needed? Who executes, reviews, fixes, retests, and approves? What counts as a pass, a failure, or an acceptable known issue? Where will evidence and defects be recorded? What blocks release, and who can accept residual risk? A bad plan says, “Test checkout.” A useful plan says which payment methods, customer states, currencies, tax rules, failure responses, retries, receipts, refunds, environments, and expected results must be checked—and which accountable human reviews the evidence. AI can accelerate the first draft. It can normalize inconsistent notes, group scenarios, suggest questions, and flag blank fields. But fluent text is not product knowledge. As the plain-language guide to what AI can and cannot do explains, plausible output is not verified output. The reusable AI test plan prompt formula Use this as a starting pattern: “Act as a test-planning assistant. I am [role] preparing to test [product, release, migration, integration, workflow, campaign, or policy change]. Using only the privacy-safe information below, draft a risk-based test plan. Include scope, exclusions, assumptions, requirements traceability, user journeys, test types, scenarios, expected results, environments, test data, owners, dependencies, evidence, defect workflow, entry criteria, exit criteria, and unresolved questions. Separate confirmed facts from suggestions. Flag contradictions, missing requirements, untestable language, unsupported assumptions, inaccessible environments, privacy or security concerns, and risks requiring specialist review. Do not invent product behavior, credentials, owners, dates, test results, or approvals.” That last sentence prevents a common failure: an AI sees “payment integration” and cheerfully invents card types, retry rules, and refund behavior. Those details may look professional while being completely wrong. Never paste credentials, access tokens, customer PII, private tickets, raw production logs, security vulnerabilities, regulated data, unreleased strategy, legal disputes, HR records, or confidential client material into an unapproved AI tool. Redact sensitive content. Use safe summaries, synthetic test data, ticket references, and links to controlled systems. What to collect before drafting a test plan Do not give AI a feature name and ask it to create “complete test coverage.” That is how you get a beautiful inventory of assumptions. Bring the strongest available inputs: InputWhat it contributesHuman check Approved scope and requirementsThe boundaries of testingConfirm current version and exclusions Acceptance criteriaObservable expected behaviorRemove vague or contradictory language User journeys and operating proceduresReal ways people use the changeValidate with users and operators Architecture and integration notesSystem boundaries and dependenciesReview with technical owners Risk assessmentPriority and potential harmConfirm severity and likelihood judgments Supported devices, browsers, roles, and localesCoverage dimensionsConfirm the actual support policy Environment inventoryWhere tests can runVerify access, parity, and limitations Test data rulesSafe, representative inputsConfirm privacy and retention controls Prior defects and incident notesKnown failure patternsCheck whether they still apply Release and rollback planTiming and recovery constraintsConfirm decision owners and gates When information is missing, ask AI to create questions rather than fill blanks. “What should happen when the payment provider times out?” is useful. A fabricated timeout rule is not. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI test plan prompts you can use today Replace bracketed text with verified, privacy-safe context. Require source references wherever possible. The goal is a review artifact, not a robot-written permission slip. 1. Extract testable scope from requirements “Review these approved requirements and acceptance criteria: [paste privacy-safe text]. Build a traceability table with requirement ID, source, intended behavior, testable condition, expected result, priority, proposed test type, dependency, owner field, and unresolved question. Put vague, conflicting, or untestable requirements in a separate section. Do not rewrite ambiguity as certainty or invent missing behavior.” This prompt shows whether the source material can support a plan at all. Phrases such as “loads quickly,” “works seamlessly,” and “supports enterprise users” are aspirations wearing fake mustaches. They need measurable definitions from accountable people. Traceability also protects against accidental omission. Every high-priority requirement should connect to at least one test, and every test should connect to a requirement, risk, incident, or explicit exploratory goal. 2. Build a risk-based coverage matrix “Using this release scope, risk register, customer impact summary, and known constraints: [paste], draft a coverage matrix. For each risk, include affected journey, possible failure, impact, likelihood label supplied by humans, detectability, proposed test coverage, environment, data needs, owner field, evidence required, and escalation path. Rank attention using only the supplied ratings. Flag risks with no proposed coverage and coverage with no stated risk or requirement.” Testing everything equally is impossible and usually dishonest. Risk-based planning concentrates effort where failure would hurt customers, revenue, operations, privacy, safety, or trust. AI may help organize the matrix, but people must judge impact and acceptable risk. Pair this with AI risk assessment prompts when the blast radius is still unclear. 3. Draft happy-path and failure scenarios “For this verified workflow: [paste], draft candidate happy-path, alternate-path, and failure-path scenarios. For each scenario include preconditions, actor, steps at a useful level, expected result, evidence, cleanup, and source requirement. Mark every behavior not explicitly supported by the source as ‘needs confirmation.’ Include interruptions, retries, cancellation, duplicate submission, unavailable dependencies, permission failures, and recovery only when relevant.” Happy paths prove that the demo can work. Failure paths help prove the product can survive contact with reality. Do not accept a giant pile of mechanically varied cases. Review whether each scenario represents a meaningful behavior or risk. Ten versions of the same button click are not sophisticated coverage. They are paperwork cardio. 4. Identify boundary conditions and edge cases “Analyze these fields, rules, limits, states, and transitions: [paste]. Propose boundary and edge-case tests for empty values, minimum and maximum allowed values, just-inside and just-outside limits, unusual but valid characters, duplicate actions, stale state, concurrent updates, timezone boundaries, locale differences, partial completion, and interrupted sessions where applicable. Cite the source for every limit. Do not invent undocumented limits or claim an edge case is supported.” This prompt is good at generating possibilities, which is different from knowing what matters. A human should remove irrelevant suggestions and add domain-specific cases the model cannot infer. For financial, medical, legal, accessibility, security, or safety-sensitive behavior, involve qualified reviewers. Generic AI suggestions are not specialist validation. 5. Plan integration and data-flow testing “Using this approved architecture and data-flow summary: [paste], draft an integration test worksheet. For each interface include source, destination, trigger, payload summary, authentication category without secrets, expected response, timeout behavior if documented, retry or idempotency rule if documented, failure handling, observability evidence, test environment, safe test data, owner fields, and unresolved questions. Mark inferred connections as unverified.” Integrations fail in the seams: mismatched formats, delayed events, duplicated messages, stale caches, permissions, and assumptions about which system owns truth. Never paste secrets or raw production payloads. Use schemas, redacted examples, synthetic records, and approved test accounts. AI should help prepare the review, not become another data leak with excellent grammar. 6. Define environments and test data “Create an environment and test-data readiness table from this approved information: [paste]. Include environment, purpose, version or build, configuration differences, integration availability, account roles, safe data sets, refresh rules, access owner, monitoring, known limitations, readiness status, and validation step. Separately list scenarios that cannot be trusted in this environment and why. Do not include or request credentials, tokens, customer records, or production secrets.” A test can pass because the test environment quietly removed the hard part. Maybe email is stubbed, permissions are relaxed, data volumes are tiny, or the third-party sandbox behaves nothing like production. Document those differences before execution. “Passed in staging” means very little if nobody can explain staging. 7. Assign owners, evidence, and review points “Turn this draft test inventory into an ownership and evidence matrix: [paste]. For every test group include executor role, reviewer role, technical contact, business contact, planned window, evidence format, evidence location, defect destination, retest owner, and approval checkpoint. Leave unconfirmed people and dates blank. Flag single points of failure, conflicting ownership, missing reviewers, inaccessible evidence locations, and approval steps with no accountable role.” “QA owns quality” is organizational fan fiction. QA may coordinate testing, but product owners clarify intent, engineers fix defects, security reviews security risk, operators validate runbooks, and release owners make release decisions. Evidence should be useful and proportionate: test run records, screenshots where they prove something, logs stored in approved systems, query results, accessibility reports, or signed review notes. A green cell with no source is interior decoration. 8. Create entry and exit criteria “Using this scope, risk level, environment status, defect policy, and release process: [paste], draft measurable entry and exit criteria. Entry criteria should cover approved requirements, build availability, environment readiness, access, test data, dependencies, and known blockers. Exit criteria should cover required coverage, evidence review, defect thresholds supplied by humans, regression status, unresolved-risk decisions, documentation, rollback readiness, and named approval. Flag subjective language and missing decision owners.” Entry criteria stop teams from “starting testing” before there is anything stable to test. Exit criteria stop release conversations from becoming mood rings. Do not ask AI to choose defect thresholds or accept risk. Those are governance decisions. The model can format agreed rules and point out contradictions. 9. Turn defects into retest and regression coverage “Review these privacy-safe defect summaries: [paste]. For each defect, draft a retest outline and candidate regression impact map. Include original expected behavior, observed behavior, affected component, likely adjacent journeys to review, required environment, safe data setup, evidence, fix version, retest owner field, and unresolved questions. Do not diagnose root cause, claim the fix works, or infer affected systems without evidence.” A retest checks whether the specific defect was fixed. Regression testing checks whether the fix broke something nearby. They are related, not interchangeable. Use AI bug report prompts to improve vague defect reports and AI defect triage prompts to structure the queue. Humans still reproduce, prioritize, fix, retest, and accept outcomes. 10. Write a release test summary “Using only these verified test results, defect statuses, evidence links, scope changes, and signed risk decisions: [paste], draft a release test summary. Include tested scope, excluded scope, environment limitations, execution totals if supplied, passed and failed areas, open defects, retest status, residual risks, evidence links, rollback or monitoring dependencies, required decisions, and named approvers. Distinguish facts from recommendations. Do not change counts, invent results, declare the release safe, or imply approval.” A test summary is not a victory speech. It is a decision packet. A good one helps a release owner understand what was tested, what was not, what remains broken, and who accepted which risk. For the final release gate, AI go/no-go decision prompts can organize evidence. They cannot make the decision. How to review AI-generated test cases Before adding generated cases to a test management system, review them against five filters. 1. Source Can you point to the requirement, risk, incident, design, policy, or exploratory charter that justifies the test? If not, label it as a suggestion and validate it. 2. Value Would failure teach the team something that changes a decision? Delete duplicates and ornamental cases. More rows do not automatically mean more confidence. 3. Executability Are preconditions, environment, role, data, steps, and expected results clear enough for a capable tester? If the result says “works correctly,” the case is not ready. 4. Safety Does the case require sensitive data, risky production actions, destructive operations, or specialist review? Move it into the correct controlled process. 5. Ownership Who executes it, reviews evidence, fixes failure, approves exceptions, and decides release impact? Unowned tests become archaeological artifacts immediately after the planning meeting. A practical AI QA checklist can help review the plan itself before execution. Test plan versus QA checklist, UAT, and smoke testing These artifacts overlap, but they solve different problems. Test plan: the overall strategy connecting scope, risk, coverage, resources, environments, evidence, and decisions. QA checklist: a repeatable list of checks, often useful for a feature type or release routine. User acceptance testing: validation by business users or representatives that the change supports intended real-world work. Smoke testing: a narrow, fast check that critical functions are alive enough for deeper testing or release confidence. Regression testing: checks that existing behavior still works after a change. Exploratory testing: structured investigation where learning guides the next test. Use AI user acceptance testing prompts for business scenarios and AI smoke test prompts for critical-path launch checks. The test plan explains how those activities fit together. Common mistakes when using AI for test planning Asking for complete coverage No tool can prove complete coverage from incomplete context. Ask for candidate coverage, gaps, and validation questions. Treating generated expected results as requirements Expected results must come from approved behavior. If AI invented them, the test may enforce the wrong product. Optimizing for case count A thousand shallow cases can hide one untested critical failure. Optimize for risk reduction and decision-quality evidence. Ignoring environment differences A passing result in a simplified sandbox may not transfer to production. Record limitations and compensating checks. Using sensitive data for convenience Do not feed real customer records into an AI tool because synthetic data takes effort. Privacy debt is still debt. Letting the summary overstate confidence If only happy paths ran, say so. If an integration was unavailable, say so. If risk was accepted, name the human decision and source. Frequently asked questions Can ChatGPT write a complete test plan? It can draft a structured plan from the information you provide. It cannot know undocumented product behavior, inspect environments, verify integrations, execute tests, or prove coverage. Treat the output as a reviewable first draft. What should I include in an AI test plan prompt? Include approved scope, requirements, acceptance criteria, important user journeys, known risks, supported configurations, environment constraints, test-data rules, roles, evidence expectations, defect workflow, and release gates. Tell the model not to invent missing facts. Are AI-generated test cases reliable? They are candidates, not evidence. Their quality depends on the source material and human review. Every case should be checked for source, relevance, expected behavior, executability, safety, and ownership. Can AI replace a QA engineer or tester? No. AI can accelerate drafting, formatting, grouping, and gap-question generation. Human testers investigate behavior, understand context, operate environments, notice weirdness, evaluate risk, communicate defects, and remain accountable. How do I protect confidential data while using AI for testing? Use an approved tool and follow your organization's policy. Redact names and sensitive details, use synthetic data, reference controlled tickets instead of pasting them, and never include credentials, tokens, private logs, vulnerabilities, regulated data, or customer PII. How many test cases should a test plan contain? There is no universal number. The right amount depends on change scope, risk, complexity, supported configurations, evidence needs, and available time. Coverage tied to meaningful risks beats a target case count. Who approves an AI-assisted test plan? The same accountable humans who would approve any test strategy: typically relevant QA, product, engineering, operations, security, compliance, accessibility, business, or release owners. AI does not approve plans or accept risk. Should AI-generated tests be automated immediately? No. First confirm that the behavior is real, stable, valuable, and suitable for automation. Automating a fabricated or low-value case only makes the mistake run faster. The point is better questions, not automated confidence The best use of AI in test planning is not pretending a model knows your product. It is making your team confront the gaps earlier: Which requirement is actually approved? Which failure would hurt most? Which environment cannot represent reality? Which evidence will support the release decision? Which risk has no coverage? Which test has no owner? Which open defect requires a named decision? That work is less glamorous than asking a chatbot for “200 comprehensive test cases.” It is also much more useful. Use AI for speed: organizing notes, drafting matrices, producing review questions, and keeping summaries consistent. Keep humans responsible for facts, judgment, access, execution, evidence, and consequences. That is the durable pattern for using AI at work without the nonsense. If you want the broader field guide for working with AI without outsourcing your brain, Don’t Replace Me by Dmitry Kargaev carries the same rule throughout: use the machine for leverage, then bring human judgment to the part that matters. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance AI API Testing Prompts: Catch Broken Integrations AI API testing prompts for contracts, edge cases, authorization, retries, error responses, and human-verified integration evidence. AI promptsAPI testingquality assurance AI Compatibility Testing Prompts: Catch Breakage 10 AI compatibility testing prompts for browser, device, OS, responsive, locale, network, defect triage, and evidence-based release decisions. AI promptscompatibility testingcross-browser testing --- ## AI Training Prompts: Teach Your Team Better, Faster URL: https://dontreplace.me/articles/ai-training-prompts-teach-your-team-better-faster AI training promptsemployee trainingonboardingAI at worklearning and development AI Training Prompts: Teach Your Team Better, Faster 10 copy-paste AI training prompts for onboarding, role-plays, quizzes, and coaching guides. Includes safety rules and a reusable prompt formula. June 06, 2026 Updated June 06, 2026 Most training content is bad. Not bad-interesting. Bad like a microwaved chicken breast: technically food, technically edible, something you'd only consume if someone made you. You've sat through the slidedecks. You've clicked through the compliance modules with the progress bar that won't let you skip. You know the feeling. AI training prompts won't fix bad training automatically. But they'll get you from zero to usable draft in 20 minutes, which is 20 minutes less time you spend staring at a blank slide deck resenting your job. The catch is that "fast" doesn't mean "right." AI is good at structure and bad at judgment. You bring the judgment. It brings the speed. That split matters, and we'll come back to it. Here are 10 copy-paste prompts that actually work, plus the formula behind all of them. The formula behind every AI training prompt that works Vague prompts get vague training. "Write a lesson about customer service" will produce the training equivalent of elevator music: technically present, completely forgettable. The prompts below follow a tighter structure. Every good AI training prompt has five parts: role (who the learner is), task (what they need to do), standard (what good looks like), context (tools, constraints, company-specific details you can share safely), and output format (quiz, job aid, checklist, scenario, etc.). When you give AI all five, it stops inventing filler and starts producing something you can actually edit. When you leave one out, especially the standard, you get corporate oatmeal. Filling in those five fields takes three minutes. It saves you thirty. The formula: Role + Task + Standard + Context + Output Format = a first draft worth editing. AI training prompts: 10 templates you can use today These are designed to copy, paste, and adjust. Replace the bracketed fields with your specifics. Do not paste in customer data, employee records, security incidents, unreleased strategy, legal matters, or anything else that belongs behind a firewall. 1. Turn an SOP into a plain-English lesson I have a standard operating procedure for [process name]. My audience is [role, e.g., new frontline staff with no prior experience]. Here is the SOP: [paste the SOP text here, redacted of any confidential data]. Turn this into a plain-English lesson that: - Explains why this process matters (1 short paragraph) - Lists the steps in simple numbered format - Flags the 2-3 most common mistakes - Ends with a single "check yourself" question to confirm understanding Keep the language at a level a smart 16-year-old could follow. 2. New-hire onboarding checklist Create a 30-day onboarding checklist for a new [job title] joining a [type of company, e.g., mid-size e-commerce brand]. Organize it by week. Include: - Systems access and tools setup (Week 1) - Key people to meet and what to ask them - Core processes to observe and then attempt supervised - First independent tasks with clear success criteria - A simple "what did you learn?" reflection prompt at the end of each week Format as a checklist with checkboxes. Keep it practical, not motivational. 3. Role-play scenario for practice Write a realistic role-play scenario to help [role] practice [skill, e.g., handling an upset customer who received the wrong order]. Include: - A customer opening line that sets up the challenge - 3 variations (mild, medium, difficult) of how the situation might escalate - What a good response looks like at each level - What a common poor response looks like (and why it makes things worse) - A debrief question: "What did you learn about your instinct here?" Use realistic, plain language. No corporate jargon in the customer's dialogue. 4. Manager coaching guide for a specific skill Write a one-page coaching guide for managers to help their team members improve at [skill, e.g., writing clear status updates]. Include: - What good looks like (3 observable behaviors) - What common gaps look like (with specific examples, not vague labels) - 3 coaching conversation starters the manager can use in a 1:1 - A simple practice prompt to send before the next meeting - How to measure improvement over 4 weeks Tone: direct, practical. No management theory. This kind of guide pairs well with the broader manager 1:1 and feedback prompts if you're building a full coaching toolkit. 5. Quick quiz to check understanding Write a 5-question multiple-choice quiz to check whether someone understands [topic, e.g., our returns policy]. For each question: - Write the question clearly - Provide 4 answer options (one clearly correct, two plausible wrong answers, one obviously wrong) - Include the correct answer and a one-sentence explanation of why it's right Base the questions on this content: [paste the relevant policy or lesson text, redacted of anything confidential]. Do not invent policy details. Only write questions based on what I've given you. 6. Before/after examples for quality standards I'm training [role] on [standard, e.g., how to write a professional customer email response]. Create 3 before/after examples: - BEFORE: a realistic weak version (not a strawman, make it something a real person might genuinely write) - AFTER: a stronger version that meets the standard - WHAT CHANGED: a 2-sentence explanation of the improvement Standard of good: [describe what good looks like, e.g., clear subject line, one clear ask per email, polite but not sycophantic tone, under 150 words]. 7. Translate expert notes into a job aid A subject-matter expert gave me these rough notes about [topic]: [paste notes here, redacted of confidential details]. Turn these into a one-page job aid that someone can use in the moment while doing the task. Format: - Title (clear, task-specific) - When to use this guide (one sentence) - Step-by-step instructions (numbered, max 10 steps) - Common mistakes to avoid (bullet list, 3-5 items) - Who to ask if something goes wrong Plain language. No jargon unless the expert's specific terminology is necessary for the role. 8. 30-minute workshop agenda Design a 30-minute workshop agenda to teach [role] how to [skill or task]. Include: - Opening (5 min): how to get people focused and explain why this matters - Core content (10 min): the fewest concepts needed to do this well - Practice activity (10 min): something people actually do, not just watch - Debrief (5 min): one question that surfaces what they're unsure about Also suggest: one physical prop or visual aid that would help, and one follow-up task participants can complete within 48 hours to reinforce the lesson. 9. Feedback rubric for a task or output Create a feedback rubric for evaluating [output, e.g., a customer service chat response] produced by [role]. The rubric should have 4-5 criteria. For each criterion: - Name it clearly - Describe what "meets standard" looks like (specific, observable) - Describe what "below standard" looks like (with a concrete example) - Describe what "exceeds standard" looks like Format as a simple table: Criterion | Below Standard | Meets Standard | Exceeds Standard. Base the standard on this: [paste your quality standard or example of excellent work]. For anything touching HR performance processes, run this by HR before using it formally. AI can draft a rubric, but it can't know your company's performance management policies. 10. Turn training feedback into the next version I ran a training session on [topic] and collected this feedback from participants: [paste anonymized feedback here]. Identify: - The 3 most common gaps or confusions - What changes to the content would address each one - What changes to the format or delivery would help - One element participants said was actually useful (keep this) Then suggest a revised agenda or lesson structure that incorporates the improvements. Be specific. Don't give me generic advice like "add more examples." Tell me what kind of examples, in what part of the lesson. What AI will get wrong in training content (and how to catch it) This is where Rule #13 from Don't Replace Me applies directly: garbage in, garbage out. AI drafts fast. It also invents confidently. It will write a quiz question based on a policy that doesn't exist. It will invent compliance requirements. It will describe a process that sounds reasonable and is completely wrong for your context. The antidotes are simple. Have a subject-matter expert review every draft before it goes to learners. Never let AI invent policy language, legal requirements, safety procedures, or compliance standards. If you're building training that touches regulated areas (finance, healthcare, HR, security), the AI produces the skeleton. A qualified human fills in the requirements. If you're not sure where to draw that line, what AI can and can't do is a good 5-minute read before you start. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → The things you must not paste into AI tools at work This is not optional. Before you build training content with AI, memorize this list. Do not paste into unapproved AI tools: Customer names, emails, or account data Employee performance records or HR documentation Active legal disputes or settlement information Security incidents or vulnerability details Unreleased product strategy or financial data Login credentials or API keys Proprietary datasets or trade secrets Anything your company's AI policy says is off-limits If your company doesn't have an AI use policy yet, that's a conversation worth having. There are some useful starting points in the AI policy prompts guide for whoever needs to draft one. Anonymize everything before it goes into a prompt. Real examples are useful for training. Real customer names are not. Good training still needs a human standard-setter Speed is the easy part. The hard part is knowing what good looks like in your specific context, with your specific team, doing your specific work. AI can write a role-play scenario about a difficult customer conversation. It cannot know that your company's tone is deliberately warmer than the industry default, or that your returns policy changed last month, or that a certain type of complaint escalates to legal and the script is different. That knowledge lives with your team. It doesn't live in a language model. This is what Dmitry Kargaev calls the taste moat in Don't Replace Me: the combination of judgment, context, and standards that a person who actually does the work holds, and that no AI can replicate from a blank prompt. Training is one of the places where that moat is most visible. Use the AI for the scaffold. Build the walls yourself. The practical version of this is a review owner. Every piece of AI-generated training content should have one named human responsible for checking it against reality before it goes live. Not a vibe check. A real review against the actual standard. Customer-facing teams have an extra consideration here: training that touches how your team handles complaints, escalations, or policies should be checked against your actual protocols. The customer service prompts guide has examples of how that specificity looks in practice. Frequently asked questions Can I use ChatGPT or Claude to create employee training materials? Yes, with conditions. AI tools are good at turning rough notes into structured lessons, generating practice scenarios, and drafting checklists. They're unreliable for anything involving company policy, compliance requirements, or safety procedures. Always have a subject-matter expert review the output before learners see it. What's the biggest mistake people make when using AI for training content? Skipping the standard. If you don't tell the AI what good performance looks like, it invents a generic version. The result is training that sounds plausible and teaches nothing specific. Every prompt should include a concrete description of the target skill or behavior. See the five-part formula above. Is it safe to paste our company's training materials into AI tools? It depends on the tool and your company's policy. Paste only what you'd be comfortable putting on a whiteboard in a public meeting. Anonymize examples. Never include customer data, employee records, security information, or unreleased strategy. When in doubt, check with IT or whoever owns your AI use policy. The AI policy prompts article has a starting framework. How do I make sure AI-generated training is accurate? Assign a review owner before you start. Decide who is responsible for checking the draft against your actual processes, policies, and standards. That person should review every piece before it goes to learners. Don't let "we ran it through AI" function as a quality check. For a broader look at where AI gets things right and wrong, what AI can and can't do is worth five minutes. Do I need to tell my team the training was made with AI? Probably not required, but reasonable to be transparent if asked. What matters more is whether the content is accurate and useful. Focus on that. The disclosure question is less interesting than the quality question. Can AI replace the L&D or training function? No. AI can produce first drafts faster. It cannot replace the person who understands what learners already know, where they get stuck, why previous training didn't work, and how to create conditions where people actually practice. Those are judgment calls that require context. If you want the broader picture on what roles are genuinely at risk, jobs AI can't replace lays it out. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Automation Prompts: Automate the Boring Stuff 10 copy-paste AI automation prompts for mapping workflows, writing SOPs, adding approval gates, and QA testing. Practical templates with safety rules included. ai automation promptsworkflow automationSOP templates AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts --- ## AI UI Testing Prompts: Catch Interface Bugs URL: https://dontreplace.me/articles/ai-ui-testing-prompts-catch-interface-bugs-before-users-do AI promptsUI testingquality assuranceaccessibilitysoftware testing AI UI Testing Prompts: Catch Interface Bugs AI UI testing prompts for responsive layouts, forms, accessibility, browser coverage, visual changes, and human-verified interface evidence. August 13, 2026 Updated August 13, 2026 A button can look perfect while doing nothing. A form can work on a developer's laptop and become interpretive dance on a phone. A modal can be technically visible but unreachable by keyboard. Interfaces fail in ways that screenshots, tickets, and confident demos conveniently hide. AI UI testing prompts can help turn verified requirements and designs into a useful test plan. They can organize states, propose scenarios, expose missing questions, draft accessibility checks, and summarize sanitized observations. They cannot see or operate your interface unless explicitly connected to approved tools, know what the designer intended, verify behavior across real devices, judge whether friction is acceptable, or approve a release. AI can draft the test map. A human still has to use the interface, observe what happened, and decide whether it is good enough. If you need broader strategy first, start with AI test plan prompts. For requirement-level coverage, use AI test case prompts. UI testing needs both. A huge generated checklist is not proof that a customer can complete the flow. What good UI testing actually checks UI testing is not clicking the happy path once and announcing victory. A useful test examines what the user sees, what they can operate, what state the system enters, and what happens when reality refuses to cooperate. Depending on the feature, that can include: content, hierarchy, controls, labels, and instructions; layout across supported viewport sizes and zoom levels; keyboard navigation, focus order, and visible focus; screen-reader names, roles, states, and announcements; valid, invalid, incomplete, and unexpectedly long input; loading, empty, success, warning, offline, and error states; permissions, account states, and role-specific views; browser, operating-system, device, and input-method coverage; navigation history, refresh, deep links, and interrupted flows; visual changes against an approved baseline; state changes and downstream outcomes behind the interface; evidence, cleanup, severity criteria, and human ownership. The exact coverage comes from the real requirement and the risk. A decorative preference toggle is not the same as a payment confirmation. A public marketing form is not the same as an employee health-data workflow. The model should arrange verified facts, not flatten every screen into the same generic checklist. A test also needs an oracle: the source used to decide whether behavior is correct. That may be approved acceptance criteria, a current design, a content specification, an accessibility standard adopted by the organization, a browser-support policy, or a confirmed product decision. “The model thought this looked wrong” is not an oracle. It is autocomplete wearing a tiny QA badge. The reusable AI UI testing prompt formula Use this foundation, then add one of the specialized prompts below: “Act as a UI testing planning assistant. I am testing [named user flow] for [user role] in [approved environment]. Use only the supplied requirements, approved design, supported browsers and devices, accessibility expectations, known risks, and sanitized examples. Produce [requested artifact] with setup, actions, expected results traced to a source, observable evidence, state changes, cleanup, and unanswered questions. Separate confirmed requirements from hypotheses. Do not invent behavior, test results, accessibility compliance, browser support, or release decisions.” That last sentence matters. Models complete patterns. If the design shows a familiar date picker, a model may assume it supports arrow keys, date formatting, and range selection exactly like another component. Familiar is not verified. Never paste credentials, customer PII, production screenshots, private logs, payment details, health data, HR records, unreleased strategy, vulnerability details, regulated data, or confidential client material into an unapproved AI tool. Use approved test environments, synthetic records, redacted screenshots, placeholder secrets, and human privacy and security review. What to collect before prompting “Test the settings page” produces generic sludge. Gather the smallest safe set of verified context first. InputWhy it mattersHuman check Named flow and user goalDefines what success meansProduct owner confirms scope Current requirement and designSupplies expected behaviorResolve conflicts and stale versions User roles and account statesExposes permission differencesConfirm approved test accounts Supported browsers and devicesBounds compatibility coverageCheck current support policy Viewports, zoom, and input methodsShapes responsive and access checksInclude keyboard and touch where relevant Content and validation rulesMakes form expectations preciseConfirm source and localization needs Loading, empty, and error behaviorCovers non-happy pathsIdentify safe failure simulations Synthetic fixturesEnables repeatable testingPrivacy review and cleanup plan Evidence and severity rulesTurns findings into useful defectsName owners and storage location Keep unknowns visible. If nobody has decided what happens when a title wraps to four lines, ask the model to create a clarification question. Do not let it invent a truncation rule because ellipses feel professional. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI UI testing prompts Replace bracketed text with verified, non-sensitive information. Every output is a draft artifact, not proof of execution. 1. Turn requirements and designs into a UI test matrix “Review this verified requirement, approved design, content specification, user role, and risk note: [paste]. Create a test matrix covering the happy path, alternate paths, key states, navigation, validation, accessibility, responsive behavior, persistence, and cleanup. For each expected result, cite the supplied source or label it ‘needs human confirmation.’ Do not invent interactions.” This is the useful first pass when truth is scattered across a ticket, a design file, and three messages everybody pretends are not part of the specification. Ask for traceability. Each row should connect to a requirement, design annotation, content rule, or risk. Rows without a source are questions. A designer, product owner, and tester should review the matrix before generated breadth becomes fake certainty. 2. Test responsive layouts without worshipping screenshots “Using these supported viewports, zoom levels, content constraints, breakpoints, and approved designs: [paste], draft responsive UI checks. Cover narrow and wide screens, orientation changes, long text, localization expansion, zoom, scrolling, sticky elements, overlays, touch targets, clipped content, and preserved task completion. Mark unspecified behavior as a question.” Responsive testing is about usability, not merely matching pixels. Check whether controls remain reachable, text stays readable, content order still makes sense, and the user can finish the task. A card shifting by two pixels matters less than a submit button escaping below an unscrollable modal. Use real supported devices or trustworthy emulation for execution. The model can suggest a matrix; it cannot prove what Safari, a mobile keyboard, or a high-zoom desktop actually renders. 3. Challenge forms and validation “From these verified field rules, formats, dependencies, and submission outcomes: [paste], create tests for blank required fields, optional fields, whitespace, minimum and maximum lengths, invalid formats, pasted content, autofill, duplicate submission, server rejection, corrected input, and preserved values. Include exact expected messages only when supplied.” Forms are tiny bureaucracies. They fail by rejecting valid names, deleting input after one mistake, placing an error where nobody can find it, or showing “Something went wrong” after the submission actually succeeded. Test both client and server outcomes. Confirm focus moves sensibly, errors are associated with fields, summaries link to problems where applicable, and resubmission does not create duplicates. Use synthetic data. Do not feed a general chatbot real customer records for the noble purpose of checking a dropdown. 4. Review keyboard navigation and focus “Using this verified interaction design and component inventory: [paste], draft keyboard tests for entry focus, logical tab order, visible focus, native controls, custom widgets, menus, dialogs, escape behavior, focus trapping, focus restoration, skip links, and completion without a pointer. Cite supplied requirements and flag uncertain key behavior.” A flow is not keyboard-accessible because pressing Tab eventually reaches something. Focus must be visible, ordered sensibly, and managed when content appears or disappears. Dialogs should not strand focus behind themselves or fling it to the top of the page when closed. Run these checks with an actual keyboard. Component semantics and behavior need human observation, particularly when a custom widget imitates a native control but forgets half the contract. 5. Plan screen-reader and accessibility checks “From these approved accessibility requirements, semantic structure, component states, labels, error rules, and supported assistive-technology combinations: [paste], create a manual review plan. Cover page title, landmarks, headings, accessible names, roles, values, status announcements, errors, instructions, dynamic updates, reading order, and task completion. Do not claim conformance from supplied text.” AI can help organize an accessibility review. It cannot certify accessibility. Automated checks also catch only part of the problem. A button may have a name and still be confusing; a heading structure may pass a rule and still misrepresent the page. Use qualified human review, real assistive technology, and the organization’s adopted standard. Treat generated suggestions as preparation, not a compliance statement. False confidence is particularly rotten when it excludes people. 6. Cover loading, empty, error, and interrupted states “Using this flow, dependency map, approved state designs, retry rules, and safe simulation methods: [paste], draft tests for initial loading, slow loading, empty results, partial data, offline behavior, timeout, permission failure, server error, retry, refresh, back navigation, interrupted submission, and recovery. Include evidence and stop conditions.” Teams lovingly design the populated state and leave everything else to chance. Users spend plenty of time in everything else. Check whether loading indicators explain what is happening, empty states distinguish “nothing exists” from “nothing loaded,” retries are safe, and error recovery preserves work. Confirm the underlying state. A spinner disappearing does not prove a save succeeded. 7. Test permissions and role-specific views “Using this verified role-permission matrix, account states, ownership rules, and flow requirements: [paste], create UI tests for allowed actions, hidden or disabled controls, direct navigation, stale sessions, changed roles, cross-owner records, and useful denial messages. Include the service-side outcome to verify. Do not treat hidden UI as authorization.” The interface should communicate permissions clearly, but hiding a button is not security. A user may still navigate directly or call the underlying service. Coordinate with API and security testing where risk warrants it. Use approved synthetic accounts. Never paste live session cookies, tokens, customer identities, or screenshots containing private records into the prompt. Humans verify both the visible experience and the actual authorization response. 8. Build a sensible cross-browser and device matrix “Given this current browser-support policy, usage data, feature risk, device constraints, input methods, and test budget: [paste], propose a prioritized compatibility matrix. Map high-risk flows to browser, operating system, viewport, touch or pointer, keyboard, and assistive-technology combinations. Explain prioritization without inventing market-share data.” Testing every combination is impossible. Testing only the developer’s favorite browser is lazy. Use policy, audience evidence, technical risk, and change scope to choose coverage. AI can structure the decision. Humans must confirm the support data and execute on real engines. “Chromium worked” does not prove Safari worked, no matter how strongly the generated table believes in web standards. 9. Review visual-regression differences “Compare these approved baseline notes and sanitized visual-difference observations: [paste]. Classify differences for human review by affected component, viewport, content, likely intentionality, task impact, and evidence needed. Separate cosmetic differences from possible functional or accessibility problems. Do not approve or reject changes.” Visual regression tools produce differences, not decisions. Font rendering, dynamic content, animation, timestamps, and operating-system behavior can create noise. Meanwhile, a small shift can hide a checkbox or change reading order. Have a human compare against the approved change, inspect the live interaction, and update baselines only after review. Otherwise visual testing becomes an elaborate machine for blessing whatever appeared most recently. 10. Turn observations into a reproducible UI bug report “Turn these sanitized observations into a draft UI bug report: [paste]. Include environment and build, browser or device, viewport and zoom, user role, safe setup, exact steps, actual behavior, expected result traced to a source, reproducibility, task impact questions, redacted evidence, cleanup, and missing information. Do not assign severity or claim a root cause.” “Button broken on mobile” is not a report. The reader needs the starting state, exact environment, action, outcome, expectation, and evidence needed to reproduce it. Use AI bug report prompts for more reporting patterns. A human still confirms the defect, strips sensitive information, judges impact, routes ownership, and decides whether it blocks release. A practical workflow that avoids checklist theater Use these prompts inside a human-owned process: Confirm the source of truth. Resolve stale designs and conflicting requirements. Name the user and goal. Test a real flow, not an abstract collection of screens. Define risk and coverage. Choose browsers, devices, roles, and access needs deliberately. Sanitize the input. Replace secrets and personal data with safe fixtures. Generate a draft artifact. Require traceability and explicit unknowns. Review before execution. Product, design, engineering, QA, and accessibility owners correct assumptions. Run real tests. Use approved environments, accounts, browsers, devices, and assistive technology. Inspect actual state. Verify persistence and service outcomes, not just visible messages. Capture evidence. Record versions, settings, steps, redacted media, and timestamps. Triage with humans. Confirm defects, severity, ownership, and release impact. For less scripted discovery, use AI exploratory testing prompts. For change-focused coverage, use AI regression testing prompts. AI organizes possibilities. It does not create evidence by describing what evidence might look like. Common mistakes Treating the design file as the whole product A design may omit delays, long content, permission failures, browser behavior, validation details, and state persistence. Pair it with verified requirements and the running system. Letting AI invent expected behavior A plausible interaction is still fiction. Expected results need a source. Unknown behavior should become a question, not an assertion with excellent formatting. Checking only appearance Inspect keyboard operation, semantics, state changes, history, service outcomes, and recovery. A polished screen can conceal a failed transaction. Pasting production screenshots into a chatbot Do not. Screenshots can contain names, messages, account numbers, internal URLs, or unreleased work. Redact them or use approved synthetic environments. Confusing generated checks with executed tests A model can write “verify that focus returns to the trigger.” It has not verified anything. Observed results belong to humans and authorized test systems. Automating unstable assumptions Review generated checks before placing them in CI. An invented expectation becomes a dependable false alarm, which is a very efficient way to teach a team to ignore testing. Frequently asked questions Can ChatGPT test a user interface for me? It can help design tests and draft automation from supplied requirements. Unless it has an explicitly configured, authorized browser connection, it cannot operate the interface or observe results. Even with tools, humans must control access, evidence, coverage, and release decisions. Can AI replace manual UI testing? No. It can accelerate planning, variation, documentation, and code drafts. Humans still evaluate clarity, usability, accessibility, visual intent, unexpected behavior, and business impact on real systems. Should AI generate UI automation code? It can create a useful first draft. Review selectors, assertions, waits, fixtures, privacy, cleanup, and failure messages. Run the code in a controlled environment and keep human ownership of what the test proves. How do I stop AI from inventing interface behavior? Supply current requirements and designs, require each expectation to cite a source, instruct the model to label unknowns, and reject claims that cannot be traced. Then conduct a human review. Can AI certify accessibility compliance? No. It can organize checks and explain supplied criteria. Conformance requires the appropriate standard, qualified review, real assistive-technology testing, and organizational accountability. What is the best first UI testing prompt? Start by turning one important user flow into a traceable matrix of roles, states, interactions, expected outcomes, accessibility checks, responsive coverage, and unknowns. Narrow, reviewed scope beats a generic hundred-item checklist. Can AI decide whether a UI bug blocks release? No. It can summarize supplied evidence and identify missing questions. Product, design, engineering, QA, accessibility, security, and business owners judge severity and release risk. How should I use UI prompts in an existing QA process? Use them to accelerate preparation, gap review, documentation, and reporting. Keep execution, evidence collection, defect confirmation, and sign-off in the existing human-owned process. For the wider boundary, read what AI can and cannot do. The bottom line AI UI testing prompts are useful when they turn verified context into sharper questions and executable drafts. They are dangerous when polished output is mistaken for a requirement, a test run, an accessibility judgment, or a release decision. Give the model safe facts. Demand traceability. Keep unknowns unknown. Then operate the interface, inspect the real outcomes, and make the decision with humans who understand the users and consequences. That is the wider point of Don’t Replace Me: use AI for speed and structure without handing it facts, access, taste, judgment, or accountability it does not possess. A fast assistant can improve UI testing. It cannot own whether your interface works. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance AI API Testing Prompts: Catch Broken Integrations AI API testing prompts for contracts, edge cases, authorization, retries, error responses, and human-verified integration evidence. AI promptsAPI testingquality assurance AI Compatibility Testing Prompts: Catch Breakage 10 AI compatibility testing prompts for browser, device, OS, responsive, locale, network, defect triage, and evidence-based release decisions. AI promptscompatibility testingcross-browser testing --- ## AI Usability Testing Prompts: Find Friction Fast URL: https://dontreplace.me/articles/ai-usability-testing-prompts-find-friction-before-users-quit AI promptsusability testingUX researchuser testingproduct design AI Usability Testing Prompts: Find Friction Fast 10 AI usability testing prompts for research plans, neutral tasks, moderator guides, observation notes, evidence synthesis, and human-reviewed priorities. August 20, 2026 Updated August 20, 2026 Usability testing is the part where a team watches a real person miss the giant button everyone in the design review called “obvious.” It is useful, humbling, and occasionally devastating to a favorite idea. AI usability testing prompts can help structure a study, draft neutral tasks, check a moderator guide for leading language, organize de-identified observations, and turn evidence into a reviewable brief. AI cannot become a representative participant, notice every hesitation, understand what someone meant, or decide which product tradeoff is right. AI can prepare the clipboard. Humans still recruit people, earn consent, observe behavior, interpret evidence, protect participants, and own the decision. This guide gives you ten copy-paste templates without pretending a chatbot is a research department. If you need broader visual checks, use AI UI testing prompts. For disability-specific barriers, pair this work with AI accessibility testing prompts and qualified disabled participants. What usability testing actually measures Usability testing examines whether intended users can complete realistic tasks with a product. It is not a vote on whether they like the color blue. Useful studies look for: Task success: whether participants reached the intended outcome, partly succeeded, or failed. Errors and recovery: what went wrong, whether people noticed, and how they recovered. Comprehension: whether labels, instructions, choices, and system feedback made sense. Navigation: whether people could form a useful mental model and find the next step. Efficiency: unnecessary loops, repeated reading, dead ends, and avoidable effort. Confidence: moments when participants proceeded while unsure or feared causing damage. Accessibility: barriers involving input, perception, cognition, language, motion, or assistive technology. Expectations: mismatches between what people predicted and what the interface did. A study does not need a laboratory or a hundred participants. It does need a clear question, appropriate participants, realistic tasks, ethical handling of data, and someone willing to observe rather than explain the product into passing. Do not confuse usability testing with an opinion survey. “Would you use this?” predicts very little. Watching someone attempt a relevant task reveals much more. Analytics can show where people leave; usability research can help explain what happened, but only when the team separates observation from interpretation. The reusable usability testing prompt formula Use this base prompt with any template below: “Act as a usability research planning assistant. I am studying [product, audience, workflow, maturity, and decision]. Use only the supplied research goals, participant criteria, prototype details, consent rules, known constraints, and de-identified evidence. Produce [artifact] with assumptions, exclusions, risks, open questions, and human review points. Separate observed behavior, participant statements, researcher interpretation, hypotheses, and product decisions. Do not invent participants, quotes, sessions, findings, accessibility claims, or statistical confidence.” The separation clause prevents polished nonsense. “Participant selected Help after two failed attempts” is an observation. “Participant did not trust the navigation” is an interpretation. “Rename the tab” is a design response. They belong in connected but distinct fields. Never paste participant names, emails, faces, voices, recordings, account details, health information, financial data, customer PII, credentials, private strategy, unreleased prototypes, or raw research repositories into an unapproved AI tool. Use consented, de-identified notes and approved systems. Check vendor retention, model-training, access, residency, and deletion rules before uploading research data. What to collect before prompting Give the model a compact evidence packet, not “test my app.” InputWhy it mattersHuman check Research questionKeeps the study decision-focusedProduct and research agree Intended audienceDefines whose experience mattersRecruiting criteria are defensible Participant screenerReduces obvious sample mismatchResearcher checks fairness and bias Critical task flowGrounds realistic scenariosProduct confirms current behavior Prototype/build versionPrevents evidence driftDesigner or engineer fingerprints it Known constraintsAvoids impossible recommendationsTeam confirms scope and deadlines Consent and privacy rulesProtects participantsResearch/privacy owner approves Moderator guideCreates consistencyFacilitator checks neutrality Observation schemaSeparates fact from inferenceResearchers calibrate together Decision criteriaExplains how findings will be usedProduct owner accepts accountability Record the build, prototype link version, feature flags, account state, device, browser, assistive technology, and session date. Otherwise an old prototype can quietly become “evidence” against a design that no longer exists. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI usability testing prompts Replace bracketed text with verified, sanitized details. These prompts make planning and analysis artifacts. They do not perform research. 1. Turn a vague concern into a research question “Using this product decision, target audience, workflow, existing evidence, known risks, constraints, and deadline: [paste], draft three narrow usability research questions. For each, state what behavior would be observed, which participants are relevant, which method fits, what the study cannot answer, and what decision the evidence could inform. Flag questions that are actually market, preference, analytics, accessibility, or technical-performance questions.” “Test the onboarding” is not a research question. “Can first-time administrators connect a data source and understand whether the connection succeeded without assistance?” is closer. It names a user, task, and observable outcome. Ask humans to choose the question. AI tends to broaden scope because comprehensive documents look impressive. A small study that informs one real decision beats a majestic plan nobody can recruit or analyze. 2. Draft a fair participant screener “Using this verified audience definition, study goal, required experience, exclusion rationale, accessibility needs, recruiting channels, incentive policy, and privacy rules: [paste], draft a short screener. Include neutral questions, disqualifying logic, quotas that require human approval, an accessibility/accommodation invitation, and a field explaining why each criterion is necessary. Do not infer protected traits or recommend deceptive screening.” Recruiting only coworkers, power users, or people who already understand the team’s vocabulary can make a broken workflow look healthy. On the other hand, recruiting random internet humans for specialist software can produce noise rather than insight. Review every exclusion. Criteria should connect to the research question, not convenience or stereotypes. Participants should know what data is collected, how incentives work, and whether sessions are recorded before consenting. 3. Write neutral task scenarios “Using this research question, participant context, critical workflow, starting state, prototype limits, and forbidden hints: [paste], draft realistic task scenarios. State the participant goal without naming the exact control, menu, label, or preferred path. Add setup requirements, completion signals, allowed facilitator responses, and reasons each task matters. Flag wording that teaches the interface.” A bad task says, “Click Billing, choose Export, and download the CSV.” That tests obedience. A better scenario says, “Your manager needs last month’s charges for a reconciliation. Show me what you would do.” Avoid fantasy context. Tasks should resemble work participants might actually perform. Check that accounts contain enough realistic synthetic data and that no task exposes another customer, triggers a real payment, sends a message, or alters production. 4. Build a moderator guide “Using these approved task scenarios, session length, consent script, recording policy, prototype limitations, accessibility accommodations, and research questions: [paste], draft a moderator guide. Include welcome, consent confirmation, think-aloud explanation, warm-up, task transitions, neutral probes, distress/withdrawal handling, technical-failure handling, closing questions, and post-session data steps. Mark all language requiring legal, privacy, or research review.” A guide supports consistency; it should not turn the moderator into a robot. Good facilitation pays attention to confusion, fatigue, embarrassment, and power dynamics. Participants are not failing an exam. The product is being examined. Do not rescue someone the instant they pause. Also do not let them suffer to protect methodological purity. Define when the facilitator may clarify the scenario, remind them the prototype is incomplete, or move on. 5. Create useful follow-up probes “Given these tasks and possible observable moments: [paste], draft neutral follow-up probes for hesitation, unexpected navigation, repeated actions, apparent success, failure, recovery, uncertainty, and comments that need clarification. Avoid ‘why’ questions that sound accusatory, assumptions about emotion, and questions that suggest the intended answer. Explain when each probe should and should not be used.” Useful probes include “What are you looking for right now?”, “What did you expect to happen?”, and “Tell me more about that.” Ask them after observing enough behavior; interrupting every pause changes the task. Never manufacture a participant’s intent from cursor movement. A long pause may mean confusion, distraction, reading, motor effort, network delay, or something else. Ask when appropriate and preserve uncertainty when the answer remains unclear. 6. Audit a study plan for leading language “Review this screener, introduction, tasks, probes, and closing questions: [paste]. Identify leading language, product jargon, social-pressure cues, double-barreled questions, assumed emotions, hidden solutions, and wording that reveals the preferred path. Return a table with original text, risk, neutral rewrite, and human-review note. Do not silently rewrite consent language.” Teams leak answers constantly: “How easy was it to…,” “Did you notice the convenient filter?”, or “Most users prefer…” A model can catch obvious cues, but a researcher must consider context, tone, sequence, and the relationship between facilitator and participant. Audit the prototype too. Seed data, tooltips, empty states, and account history may reveal the expected answer before the participant starts. 7. Plan an unmoderated usability test “Using this research goal, platform, audience, tasks, prototype state, expected completion time, support route, consent requirements, and analysis plan: [paste], draft an unmoderated test plan. Include participant instructions, task order, success signals, optional post-task questions, abandonment handling, technical checks, fraud/quality flags, accessibility considerations, privacy safeguards, and limits compared with moderated research.” Unmoderated testing can reach people quickly, but nobody is present to clarify a broken setup, distinguish a prototype failure from confusion, or notice that instructions exclude a participant using assistive technology. Pilot the study with humans before launch. Test every link, account, device assumption, completion code, and recording behavior. Keep the session short enough that the incentive remains fair. 8. Organize de-identified observations “Using these de-identified session notes and the approved task list: [paste], organize evidence by participant code and task. Preserve exact participant wording only where consent allows. Separate observed action, direct statement, outcome, error/recovery, researcher interpretation, confidence, and open question. Do not merge contradictory observations or invent missing timestamps, quotes, demographics, or causes.” Organization is where AI can save real time. It is also where it can erase minority experiences by compressing them into a majority theme. Keep a path back to source notes in the approved repository, and have a researcher verify every summary used for a decision. Do not treat frequency as severity. One participant encountering an account-deletion trap can matter more than six people disliking a label. 9. Separate evidence from interpretation “Review this de-identified findings draft: [paste]. Label each sentence as observation, participant statement, interpretation, hypothesis, recommendation, or unsupported claim. Identify causal language not supported by the study, universal claims from a narrow sample, missing contradictory evidence, and accessibility conclusions without appropriate participants or expertise. Suggest cautious rewrites without weakening documented critical failures.” “We watched four of five participants return to the dashboard before finding invoices” is bounded evidence. “Users cannot find invoices” overgeneralizes. “Move invoices into the dashboard” is one possible response, not the finding itself. This prompt is especially useful before stakeholder readouts, where tidy certainty tends to spread. Preserve sample details and limitations without burying the team in disclaimers. 10. Build a human-reviewed priority brief “Using these verified findings, task criticality, affected audiences, severity rubric, frequency notes, business constraints, accessibility impact, technical dependencies, and unresolved questions: [paste], draft a priority brief. For each issue include evidence, affected task, consequence, confidence, reach limits, possible response options, owner questions, validation needed, and decision status. Do not calculate fake ROI, choose a final design, or mark an issue resolved.” A priority brief should help a team decide, not turn research into a decorative list of complaints. Connect friction to user outcomes: lost work, blocked access, accidental commitment, misunderstood cost, failed recovery, or avoidable support demand. Use AI prioritization prompts to structure the conversation, but keep accountability with the product team. If a release decision is involved, AI go/no-go prompts can expose missing evidence without approving the launch. Common ways teams misuse AI in usability research Inventing users instead of meeting them Synthetic personas can help brainstorm questions. They are not research participants. A model predicting what a nurse, warehouse worker, teenager, blind user, or small-business owner might do is generated text—not observed behavior. Uploading raw recordings because summarization is convenient Recordings can contain faces, voices, screens, names, notifications, disabilities, financial details, health details, and confidential work. Convenience does not cancel consent, contracts, retention rules, or privacy law. Compressing away disagreement A summary that says “participants found navigation easy” may hide a participant who could not finish. Review source evidence, exceptions, and critical failures. Averages are excellent places for painful experiences to disappear. Treating recommendations as findings Research describes evidence and implications. Design responses need exploration and validation. The first AI-generated recommendation is often the most generic familiar pattern, not the best fit for your users and system. Measuring the participant instead of the product Words such as “failed,” “confused,” or “incorrect” can subtly blame people. Describe what happened and where the interface did not support the task. Participants are lending you their time to reveal weaknesses your team could not see. A lightweight workflow that stays honest Name one decision. Write the product decision the study may inform. Choose participants deliberately. Explain why their experience matches the question. Prepare realistic tasks. Remove hints and protect production data. Pilot with humans. Fix instructions, timing, access, and recording problems. Run and observe. Facilitate neutrally; record facts separately from interpretations. De-identify before AI use. Follow consent and approved-tool rules. Use AI for structure. Organize notes, audit language, and expose gaps. Verify against sources. Researchers review summaries and contradictory evidence. Decide with accountable humans. Product, design, engineering, accessibility, privacy, and research owners make tradeoffs. Test the response. A changed interface creates a new hypothesis, not proof of improvement. This is the same sane rule as any practical AI-at-work workflow: use the machine for speed and structure, not fabricated certainty. Understanding what AI can and cannot do is part of research quality now. Frequently asked questions Can AI replace usability testing with real users? No. AI can draft materials and organize evidence, but it cannot provide representative lived experience or observed use of your product. Simulated feedback is useful for brainstorming possible risks, not validating usability. Can I paste interview transcripts into ChatGPT? Only if participant consent, organizational policy, contracts, privacy requirements, and the specific approved tool allow it. De-identify transcripts and minimize data first. When approval is unclear, do not upload them. How many participants do I need? There is no magic number. It depends on the research question, audience diversity, risk, method, and decisions involved. Small studies can reveal important friction, but they do not justify universal claims. Plan additional rounds for distinct audiences and accessibility needs. Should AI write my usability tasks? It can draft them. A researcher should check realism, neutrality, reading level, accessibility, prototype limits, and whether the task accidentally names the solution. Pilot every task with humans. Can AI calculate a usability score from notes? It can apply a defined rubric to structured, verified inputs, but the output is only as sound as the rubric and evidence. Do not let a generated score hide severe failures, sample limitations, or disagreement. Is AI useful for analyzing usability sessions? Yes, mainly for organization: grouping de-identified notes, applying a schema, checking claims, and finding missing fields. Researchers must verify summaries against source evidence and preserve contradictions and minority experiences. How should I use AI for accessibility research? Use it to prepare checklists or organize approved evidence, never to simulate disabled participants or certify accessibility. Include disabled people, accessibility specialists, assistive technology, and real product testing. Who owns the final product decision? The accountable human team. Researchers explain evidence and limits; design, product, engineering, accessibility, privacy, legal, and other owners make and document the decision. AI does not carry the consequences. The point Usability research is valuable because real people behave differently from the team’s internal story. Do not replace that corrective signal with a model trained to produce plausible stories. Use AI to remove clerical drag: shape the question, audit the script, organize de-identified notes, and make claims easier to inspect. Then bring human attention back to the parts that require it—consent, context, observation, judgment, care, and accountability. That division of labor is the broader argument in Don’t Replace Me: let AI be fast. Keep the consequential thinking human. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Customer Persona Prompts: Build Useful Profiles 10 AI customer persona prompts to turn verified research into useful profiles, expose assumptions, reduce stereotypes, and plan human validation. AI promptscustomer personasbuyer personas AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management --- ## AI User Acceptance Testing Prompts: 10 Templates URL: https://dontreplace.me/articles/ai-user-acceptance-testing-prompts-10-templates user acceptance testingUAT promptsAI testing promptsQA planninggo/no-go launch AI User Acceptance Testing Prompts: 10 Templates 10 copy-paste AI user acceptance testing prompts for scenarios, edge cases, defect triage, and go/no-go decisions. With safety rules and a reusable formula. July 22, 2026 Updated July 22, 2026 User acceptance testing is the part of a project where things get real. You've had the requirements meetings, the sprint reviews, the "it works in dev" moments. Now actual humans with actual jobs have to use the thing. And they will find failure modes nobody in the room imagined. AI user acceptance testing prompts won't replace that process. But they'll help you structure it faster, map the edge cases you forgot, and stop shipping go/no-go summaries that are really just optimism dressed up in a table. Here's how to use them without fooling yourself. Why AI UAT prompts exist, and what they can't do The honest answer: AI is fast at structure. Give it a requirement and it'll spit out a test case format in 30 seconds. That's useful. What it can't do is know whether your warehouse staff actually read error messages, whether your checkout flow breaks on a cheap Android on slow 4G, or whether that technically-correct approval workflow is operationally humiliating for the people who have to use it every day. That's the gap. Speed is not coverage. A tidy spreadsheet of 47 test scenarios is not evidence that users can do their jobs. The prompts below help you build the structure. You still have to run the tests with real people, capture real outcomes, and make judgment calls that no AI has the context to make. One more thing before we get into the prompts. Do not paste customer PII, credentials, access tokens, private tickets, production secrets, unreleased strategy, legal disputes, financial records, HR issues, regulated data, medical information, private client conversations, or security vulnerabilities into any AI tool that hasn't been approved by your security team. Strip your inputs down to sanitized examples. Every time. The reusable AI UAT prompt formula Before the templates, here's the pattern everything below follows. If you write your own prompts, start here. [Role] You are a QA planning assistant helping a product or project team structure user acceptance testing. [Context] We are testing [what the feature/system does] for [user role]. The verified requirement is: [specific requirement]. The acceptance criterion is: [measurable pass/fail condition]. [Task] Generate [specific output: scenarios, test cases, edge cases, etc.]. [Constraints] Do not invent expected results. Flag any assumption you've made. Use [X] as a placeholder where I need to supply real data. That's it. Context, task, constraints. The more specific your input, the less fictional the output. If you paste vague requirements, you get vague test cases. Garbage in, garbage out is not a metaphor here. 10 AI user acceptance testing prompts you can copy and use today Prompt 1: Turn requirements into UAT scenarios You are a QA planning assistant. I have a verified requirement: [paste sanitized requirement]. The intended user is [role]. The acceptance criterion is [measurable condition]. Generate 5-8 UAT scenarios that test whether this requirement is met in real use. For each scenario, include: scenario name, user role, starting state, steps, expected result, and pass/fail condition. Flag any assumption you've made about the requirement. When to use it: Right after requirements sign-off, before you write individual test cases. Gets you a structure to review with stakeholders rather than starting from blank rows. Prompt 2: Map user roles and workflows I'm planning UAT for [system or feature]. The user roles involved are: [list roles]. For each role, describe the primary workflow they'll test, the key actions they'll take, and the failure scenarios most likely to affect them. Flag any role that might be missing from this list based on the feature description. When to use it: Before you recruit testers. Stops you discovering on test day that you forgot to include the one person whose workflow breaks everything. Prompt 3: Draft individual test cases You are a QA planning assistant. Using this UAT scenario: [paste scenario], generate a detailed test case including: test case ID, objective, preconditions, test data required [use X as placeholder for real values], step-by-step instructions, expected result for each step, and overall pass/fail criteria. Do not invent expected results. Flag where I need to verify the expected result against the actual system. When to use it: When you have scenarios approved but need to turn them into step-level instructions testers can follow without guessing. Prompt 4: Identify edge cases and boundary conditions I'm writing UAT for [feature description]. The normal happy-path scenario is: [describe it]. Generate 8-10 edge cases and boundary conditions a real user might trigger accidentally. Include: the triggering action, the user type most likely to do it, and the potential failure mode. Do not assume the system handles these correctly. When to use it: After your happy-path cases are written. This is where AI earns its keep. It's faster at generating "what if they do this weird thing" scenarios than most teams are in a meeting. Prompt 5: Define test data rules and boundaries I need to define test data rules for UAT of [feature]. The data fields involved are: [list fields]. For each field, generate: valid boundary values, invalid boundary values, null/empty handling, special character handling, and maximum length scenarios. Use [PLACEHOLDER] instead of real values. Do not use real customer data, credentials, or PII in any example. When to use it: Before testers start. Prevents half your defects being caused by bad test data rather than actual bugs. Prompt 6: Write tester instructions for non-technical participants I need to brief non-technical users on UAT for [feature]. They are [role description] and their goal is to test [what]. Write a one-page tester briefing that includes: what UAT is and what they're being asked to do, how to log what they find, what counts as a pass vs. a fail, how to escalate blockers, and what not to do [e.g., don't use production data, don't share login credentials]. Use plain language. No jargon. When to use it: When your testers aren't developers or QA engineers. Which is the entire point of user acceptance testing. Prompt 7: Capture and categorize tester feedback I have raw tester feedback from UAT: [paste sanitized feedback notes, removing any PII or sensitive content]. Categorize each item as: defect, usability issue, missing requirement, question/clarification needed, or out of scope. For defects and usability issues, assign a severity: critical, high, medium, or low based on user impact. Do not mark any item as resolved unless I explicitly tell you it's been verified as fixed. When to use it: After each UAT session. Turns a wall of tester notes into something a product owner can actually triage. See also: AI QA checklist prompts for catching mistakes before they ship. Prompt 8: Triage bugs and blockers for go/no-go Here is the current UAT defect log: [paste sanitized defect list]. For each open item, flag: whether it blocks launch (yes/no), which user roles are affected, whether a workaround exists, and the estimated user impact if shipped as-is. Do not recommend closing or accepting any defect. That decision belongs to the product owner and stakeholders. Only provide the structured analysis. When to use it: In the final days before a launch decision. The AI doesn't make the call. It just structures the evidence. Prompt 9: Summarize UAT results for stakeholders I need to write a UAT results summary for stakeholders. Here is the testing data: total scenarios: [X], passed: [X], failed: [X], blocked: [X]. Open defects by severity: critical [X], high [X], medium [X], low [X]. Key findings: [paste sanitized summary]. Write a one-page results summary with: testing scope, pass rate, open issues, risks if launched now, and recommended next steps. Do not present incomplete evidence as approved. Flag any area where testing coverage was insufficient. When to use it: Before your go/no-go meeting. Gives stakeholders a clean read of where things stand, without you typing "everything looks good!" when it doesn't. For getting this to the right people fast, AI stakeholder update prompts can help. Prompt 10: Write a go/no-go launch recommendation I need to write a go/no-go recommendation based on UAT results. Testing status: [pass rate, open defects by severity]. Rollback plan: [describe or note if one doesn't exist]. Support readiness: [is support trained, do they have runbooks?]. Accessibility check: [completed/not completed]. Compliance sign-off: [completed/not completed]. Stakeholder approvals received: [list who has approved]. Write a structured launch recommendation that includes: recommendation (go/no-go/conditional go), rationale, outstanding risks, conditions that would change the recommendation, and who needs to provide final sign-off. Note clearly: this recommendation is a draft for human review and does not constitute approved sign-off. When to use it: Before any launch. The final word belongs to a human with decision authority. The AI's job is to make sure nothing got glossed over. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What AI genuinely can't do in UAT A neat test plan that covers every scenario on paper is not the same thing as actual test coverage. This is the gap most teams fall into, and AI makes it easier to fall into because the output looks so organized. Real UAT failures come from things like: a user who doesn't read the screen the way a developer assumes, an integration that only breaks when two things happen simultaneously, an accessibility issue on a screen reader that nobody tested, a data edge case that only exists in production environments. None of those show up in an AI-generated test case unless a human put that specific scenario in the prompt. Don't Replace Me has a rule for this: taste is the moat. Knowing which failure is embarrassing, which friction is tolerable, and when a technically passing test is still an operationally broken experience, that's human judgment. AI can scaffold the structure. It can't tell you which bug matters most to the person who has to use this system for eight hours a day. The other thing AI will happily do if you let it: make missing coverage sound like completed coverage. Vague inputs produce confident-sounding summaries. A prompt that says "summarize the UAT results" based on thin notes can produce a paragraph that sounds thorough and isn't. You have to verify everything. Especially the stuff that reads well. For related guidance on where AI-generated acceptance criteria can quietly fail you, read this breakdown of acceptance criteria prompts. The safety checklist you should read before pasting anything This bears repeating because people skip it every time. Before using any of the prompts above, strip your inputs of: Customer PII (names, emails, IDs, anything identifiable) Access tokens, API keys, credentials of any kind Production secrets or environment configuration Private Jira/Linear/GitHub tickets with sensitive context Legal dispute details or HR matters Financial records or regulated data Medical information Security vulnerabilities or penetration test findings Unreleased product strategy or roadmaps Private client conversations Use placeholders. Use sanitized examples. If your organization has an approved AI tool for internal work, use that instead of a public interface. If you're unsure whether something is safe to paste, it isn't. Frequently asked questions What are AI UAT prompts and how do they work? AI UAT prompts are structured instructions you give to a tool like ChatGPT or Claude to help you generate test scenarios, draft test cases, identify edge cases, or summarize test results. They work by giving the AI a clear context (the requirement), a specific task (write test cases), and constraints (don't invent expected results). The output is a starting point for human review, not a finished QA artifact. Can AI replace a UAT tester or QA analyst? No. AI can help you structure the planning, brainstorm edge cases, and organize tester feedback faster. It can't run tests, observe how real users interact with a system, identify usability issues through lived experience, or make judgment calls about which failures matter most. UAT requires real humans in the roles who actually test the thing. What's the biggest mistake people make when using AI for UAT? Mistaking a well-formatted test plan for evidence of coverage. AI produces confident-looking outputs from vague inputs. If your requirements are incomplete or your acceptance criteria aren't defined, AI will fill the gaps with plausible-sounding fiction. Read more about this problem at AI requirements gathering prompts. Should I use AI for go/no-go launch decisions? AI can help you structure the evidence and draft the recommendation document. It should never be the decision-maker. Go/no-go requires a human with decision authority who has reviewed the actual test evidence, the open defect list, the rollback plan, and the support readiness status. The AI's output is a draft for that human to approve, not the approval itself. What data should I never paste into AI tools for UAT? Never paste customer PII, access credentials, production secrets, regulated data (financial, medical, HR), security vulnerabilities, legal dispute details, or private client information into any AI tool that hasn't been approved by your security or IT team. Use sanitized inputs and placeholders every time. How do I make AI UAT prompts more useful? Give them more specific inputs. The prompt formula at the top of this article shows the pattern: role, context (including the verified requirement and acceptance criterion), task, and constraints. Vague prompts produce vague test cases. Specific prompts with real acceptance criteria and defined user roles produce output you can actually review and use. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance --- ## AI Vendor Evaluation Prompts: Pick Tools Wisely URL: https://dontreplace.me/articles/ai-vendor-evaluation-prompts-pick-tools-wisely ai vendor evaluationprocurement promptsai tools for workvendor selectionai at work AI Vendor Evaluation Prompts: Pick Tools Wisely 10 copy-paste AI vendor evaluation prompts for scorecards, demo questions, security review, and recommendation memos. With safety rules built in. June 18, 2026 Updated June 18, 2026 Someone sends you a Slack message: "Hey, we're evaluating three AI vendors. Can you help pull together a recommendation by Friday?" It's Tuesday. You have a half-finished spreadsheet, four browser tabs with marketing copy, and a demo recording you haven't watched yet. This is where AI vendor evaluation prompts actually earn their keep. Not because AI can do the procurement work for you. But because it can help you organize the chaos fast enough that you actually have time to do the thinking that matters. The ten templates below cover the full evaluation arc: from defining what you actually need, through scoring vendors and reviewing their security docs, to writing the recommendation memo your stakeholders will read. Each one is copy-paste ready. Each one has a boundary clearly marked: where the AI assistant stops and where your legal, security, or finance team takes over. What a good AI vendor evaluation prompt actually does Most people approach vendor evaluation like they're shopping for a laptop. Compare the spec sheet, check the price, pick the one with the best-looking demo. Then they're surprised when the integration breaks three months later or the contract auto-renews at twice the price. A well-built prompt doesn't fix that problem by being cleverer. It fixes it by forcing you to be specific about what you actually need before you talk to anyone. The formula is simple: Role + Context + Task + Constraint + Output format "You are a procurement analyst. We are evaluating [tool type] for [use case] in a [company size/industry] organization. Here are our requirements: [list]. Here are our constraints: [budget, timeline, compliance needs]. Generate [scorecard / question list / comparison summary] in [format]." That's it. The specificity is the work. Vague inputs produce confident-sounding outputs that mean nothing. One more thing before the templates: do not paste customer PII, employee records, contract terms under NDA, private financial data, credentials, unreleased product plans, or anything your security team would wince at into a public AI tool. Use anonymized or synthetic data for vendor evaluation work. If your organization has an approved internal AI environment, use that. Prompt 1: Define the actual business need Half of bad vendor decisions start before the first demo, because nobody wrote down what problem they were solving. "Help me write a one-page business requirements document for evaluating [tool category, e.g., AI meeting transcription software]. Our team is [describe team: size, function, workflow]. The problem we're trying to solve is [describe pain point]. Our must-have requirements are [list]. Our nice-to-have requirements are [list]. Our constraints include [budget range, deployment timeline, compliance requirements like HIPAA/SOC2/GDPR, existing tech stack]. Format this as a structured requirements doc with sections for problem statement, requirements, constraints, and evaluation criteria." Owner: The person signing the budget approval should sign off on this doc before you talk to a single vendor. Prompt 2: Build a vendor scorecard "Create a weighted scorecard for evaluating [tool type] vendors. Our highest-priority criteria are [list top 3]. Our secondary criteria are [list 3-5 more]. Include columns for: vendor name, score per criterion (1-5), weight per criterion, weighted score, and notes. Format as a table. Add a row for 'minimum threshold' so we can flag any vendor that scores below acceptable on a dealbreaker criterion." Safety note: The weights you assign to criteria are a judgment call. AI can suggest criteria; you and your stakeholders decide what matters. Don't let a generated scorecard become a rubber stamp. This is a good moment to link to the broader AI decision-making prompts guide if you want to structure the decision logic before you build the scorecard. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Prompt 3: Compare vendor claims side by side You've got three pitch decks and a set of follow-up emails. Here's how to make sense of them fast. "I'm going to paste anonymized notes from three vendor presentations below. For each vendor, extract: (1) their stated key features, (2) any specific claims they made about performance, reliability, or ROI, (3) what they didn't answer when asked about [integration / security / pricing], and (4) any red flags or vague language. Format as a comparison table. Label unverified claims clearly." Important: When AI flags a claim as "unverified," your job is to verify it, not to assume the AI did. Ask vendors for documentation, customer references, and third-party audits. AI can spot the gaps in the notes; it can't call the reference. AI vendor evaluation prompts for the hard stuff The easy part of vendor evaluation is building the scorecard. The hard part is the questions nobody wants to ask out loud: What happens to our data? What does the contract actually say? What's the exit plan if this doesn't work? These three prompts help you get there. Prompt 4: Review security and privacy documentation "I'm going to paste anonymized excerpts from a vendor's security whitepaper and privacy policy. Summarize: (1) what data they collect and retain, (2) how long they retain it, (3) what subprocessors they use, (4) what certifications they claim (SOC 2, ISO 27001, HIPAA, GDPR, etc.), and (5) what's missing or unclear. Flag anything that contradicts standard data minimization practices. Format as a bulleted summary." Non-negotiable: This output is a starting checklist for your security and legal team, not a clearance. AI cannot verify whether a vendor's SOC 2 certification is current, what it covers, or whether their subprocessors are acceptable under your agreements. Your security and privacy teams make those calls. If you don't have those teams, get the vendor's security report and have a lawyer read the data processing agreement. Prompt 5: Map data exposure risk Before you sign, you need to know exactly what data would touch this tool. This prompt is paired well with the AI risk assessment prompts for mapping the wider blast radius. "We are considering using [tool type] for [use case]. The data types that would flow through this system include [list: e.g., employee names, customer emails, financial records, internal documents]. For each data type, help me assess: (1) sensitivity level (public / internal / confidential / restricted), (2) relevant compliance requirements (GDPR, HIPAA, CCPA, etc.), (3) whether this data type should be excluded from the tool entirely, and (4) questions to ask the vendor about handling this data type. Format as a table." Owner: Legal and compliance sign off on the data classification. You're using AI to organize the conversation, not to approve the risk. Prompt 6: Prepare demo questions Vendor demos are theater. Everyone knows it. The demo shows you the three features that work perfectly, in a controlled environment, with no messy edge cases and no real user data. Your job is to break it. "We are attending a demo for [tool type] from [vendor name or 'Vendor A']. Our use case is [describe]. Generate 15 specific demo questions designed to: reveal integration complexity, expose gaps in claimed features, pressure-test performance claims, clarify pricing and contract terms, and surface support and escalation processes. Include at least three questions that start with 'What happens when...' to probe failure modes." Dmitry Kargaev covers this exact instinct in Don't Replace Me: knowing which demos are theater and which tradeoffs will hurt later is one of the genuinely human skills AI can't replicate. Speed is not due diligence. Prompt 7: Check integration and workflow fit A tool that doesn't connect to your existing systems is just an expensive extra tab. Before you get excited about features, check the pipes. "We use the following tools in our current workflow: [list your stack, e.g., Salesforce, Slack, Google Workspace, Jira]. We are evaluating [new tool] for [use case]. Help me create a list of integration questions to ask the vendor, covering: native integrations vs. API-only, authentication and SSO requirements, data sync frequency and limitations, workflow steps that would need to change, and what manual workarounds would be required if an integration breaks. Also list the internal teams who would need to be involved in integration work." For a deeper look at mapping your workflows before adding new tools, the AI workflow audit prompts are worth running first. Prompt 8: Design a pilot A pilot without success criteria is a vibe check with a contract attached. Don't do that. "Help me design a 30-day pilot program for [tool type] with [vendor name or 'Vendor A']. The pilot group would be [number] users from [team/function]. Define: pilot objectives tied to our business requirements, three to five measurable success criteria with specific thresholds, a week-by-week milestone structure, what data we'll collect, who owns evaluation, and what a failed pilot looks like. Include a go/no-go decision framework at the end of 30 days." Named owner: Someone specific is accountable for running this pilot and making the recommendation. If everyone owns it, nobody owns it. Prompt 9: Summarize stakeholder feedback You ran the demo, collected feedback from five people, and now you have a mix of Slack messages, a voice note, and two emails. Here's how to make that usable. "I'm going to paste anonymized feedback from [number] stakeholders who attended the [vendor] demo or pilot. Summarize: (1) common themes in positive feedback, (2) common concerns or objections, (3) any dealbreaker comments that would block adoption, (4) which stakeholder groups had the most friction (e.g., technical vs. non-technical users), and (5) open questions that need answers before a decision. Format as a structured summary with a section for 'what we still need to know.'" Strip any personally identifying information from the feedback before pasting. Use role labels instead of names. Prompt 10: Write the recommendation memo "Write a vendor recommendation memo for [tool type]. The recommendation is [proceed with Vendor X / do not proceed / request additional information]. The format should include: executive summary (3 sentences), problem we were solving, evaluation methodology, summary of vendors considered, scorecard results, key risks and mitigations, total cost of ownership over 12 months, integration requirements, pilot plan summary, and recommended next steps with owners and deadlines. Tone should be direct and factual. Mark any section where additional information is needed before the memo is final." Before you send this: Every number in the TCO section needs to come from actual vendor quotes. Every compliance claim needs sign-off from legal or security. Every integration timeline needs validation from your IT team. AI drafted the structure; humans are accountable for every assertion in it. This is also a good time to run the memo through the AI policy prompts framework if your organization needs to document how you'll govern the tool after adoption. And if you're still building out your automation thinking, the AI automation prompts cover what happens after you pick the tool. What AI can't do in vendor evaluation It can't verify that a SOC 2 certification covers what you think it covers. It can't tell you whether a vendor's pricing will hold after you sign. It can't read the contract and identify the auto-renewal clause buried in Section 14. It can't tell you whether a customer reference is real or curated. It can't approve a vendor for HIPAA compliance. It can't promise an ROI. What it can do is help you organize your requirements, structure your questions, compare notes at speed, and write the first draft of a memo that would otherwise take you three hours on a Friday afternoon. The rule is simple: AI does the organization. Humans make the decisions. Named humans, with actual authority, who can be held accountable if it goes wrong. If you want the full framework for staying useful while everyone else mistakes a polished demo for a good decision, that's exactly what Don't Replace Me is built for. Rule #7 is about taste: knowing which promises matter and which ones are just a good slide deck. That's not a skill AI can do for you. It's the skill that justifies having you in the room. Frequently asked questions Can I use AI to evaluate AI vendors? Yes, and it's one of the better uses of AI in procurement. Use it to build scorecards, organize notes, prepare demo questions, and draft recommendation memos. Don't use it to verify security certifications, approve compliance, or sign off on legal or financial terms. Those decisions need humans with real accountability. What data should I never paste into an AI tool during vendor evaluation? Don't paste customer PII, employee records, NDA-covered contract terms, credentials, private financial data, unreleased product plans, or anything your security policy classifies as confidential or restricted. Use anonymized or synthetic data in your prompts. If your organization has an approved internal AI environment, use that instead of a public tool. What's the most common mistake in AI vendor evaluation? Confusing a good demo with a good product. Vendor demos are controlled environments. The prompts in this guide are specifically designed to surface the questions that break demos: integration failure modes, data handling edge cases, pricing traps, and what happens when something goes wrong. See the broader guide to using AI at work for more on using AI to pressure-test instead of just validate. How do I know if an AI-generated vendor recommendation is trustworthy? It isn't, on its own. An AI-generated recommendation memo is a first draft and a structure. Every factual claim in it, including cost figures, compliance status, integration timelines, and customer references, needs to be verified by the people accountable for those domains: legal, security, finance, IT. The AI organizes the thinking. The humans sign off. What should a pilot program always include? Measurable success criteria with specific thresholds, a named owner who is accountable for the result, a defined timeline, a go/no-go decision framework, and a documented rollback plan if the pilot fails. A pilot without those elements is not a pilot. It's a slow path to an awkward contract renewal. How do I handle lock-in risk when evaluating vendors? Ask the vendor explicitly: what does data export look like, in what formats, and how long does it take? What happens to our data if we cancel? Is there an API? What's the migration path to a competitor? If they're vague on any of these, that's your answer. Also check contract terms for auto-renewal clauses, minimum commitment periods, and price escalation provisions. AI can help you build the question list; a lawyer should read what comes back. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Automation Prompts: Automate the Boring Stuff 10 copy-paste AI automation prompts for mapping workflows, writing SOPs, adding approval gates, and QA testing. Practical templates with safety rules included. ai automation promptsworkflow automationSOP templates AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts --- ## AI Win-Loss Analysis Prompts: Learn Why Deals Move URL: https://dontreplace.me/articles/ai-win-loss-analysis-prompts-learn-why-deals-move AI promptswin-loss analysissales analysisrevenue operationsproduct marketing AI Win-Loss Analysis Prompts: Learn Why Deals Move 10 AI win-loss analysis prompts to examine deal evidence, challenge sales assumptions, protect buyer data, and turn patterns into accountable action. August 29, 2026 Updated August 29, 2026 Win-loss analysis often starts with a CRM export and a question that sounds simple: “Why did we win some deals and lose the others?” Then everyone discovers that half the loss reasons say “competition,” three say “price,” and one says “vibes.” Excellent database. No notes. AI win-loss analysis prompts can organize evidence, code interview notes, compare segments, expose contradictions, and draft a research plan. They cannot read a buyer’s mind, establish causation from a CRM field, or determine that a salesperson failed because a model summarized incomplete records with impressive punctuation. AI can make deal evidence easier to inspect. Humans still own consent, definitions, interviews, interpretation, coaching, commercial judgment, and final decisions. These ten templates help sales leaders, revenue operations teams, product marketers, founders, and account executives learn from won, lost, and no-decision deals without turning autocomplete into the vice president of imaginary certainty. What win-loss analysis should actually answer A useful analysis clarifies what happened, who reported it, when it happened, which deals are comparable, and what the evidence can support. It separates direct buyer statements from seller interpretations and measured events from theories. Start by defining outcomes: Won: the buyer signed under a documented scope and commercial structure. Competitive loss: the buyer selected a named alternative, confirmed by credible evidence. No decision: the buyer paused, canceled, or kept the current process. Disqualified: the opportunity did not meet agreed qualification criteria. Unknown: the outcome or reason is not supported by enough evidence. Do not quietly combine no-decisions, budget freezes, procurement failures, product gaps, and competitor wins into “lost.” Those outcomes have different causes and demand different responses. Likewise, do not compare a small inbound deal with a complex enterprise procurement cycle and call the resulting average “the buyer journey.” CRM data records a sales process. It does not automatically record buyer truth. Seller notes can be useful and biased at the same time. Buyer interviews provide direct evidence, but participants are selected and may rationalize decisions afterward. Call recordings capture words, not every stakeholder discussion that happened elsewhere. For better primary research, use AI customer interview prompts. For broader quantitative work, AI data analysis prompts can help structure checks while humans verify the calculations. The reusable evidence-first win-loss prompt formula Add this instruction to any template below: “Act as a win-loss analysis assistant. Use only the sanitized evidence I provide for [deal population] during [date range]. The outcome definitions are [definitions], and the business decision is [decision]. Produce [artifact]. Separate observed events, direct buyer statements, seller statements, calculated measures, interpretations, hypotheses, and unknowns. For every important finding, include the source, segment definition, sample size, date range, confidence note, contradictory evidence, competing explanation, owner, and next validation step. Do not invent quotes, motives, probabilities, causal claims, competitor facts, or sensitive attributes. Flag missing records, selection bias, taxonomy changes, unfair comparisons, and conclusions requiring statistical, privacy, legal, HR, sales-leadership, or customer review.” This formula makes the output show its homework. Without it, “discount requested before loss” may become “pricing caused the loss.” Maybe. Or perhaps the buyer lacked authority, the project was canceled, requirements changed, procurement stalled, or the discount request was merely the final recorded event. Never paste buyer names, emails, call recordings, contracts, pricing exceptions, payment details, credentials, confidential roadmaps, private performance notes, or legally sensitive information into an unapproved AI tool. Use approved systems, minimum necessary fields, aggregation, de-identification, access controls, retention limits, consent, and human privacy review. What to collect before prompting Build a documented evidence pack before asking for conclusions. InputWhy it mattersHuman check Outcome definitionsPrevents category soupRevenue operations confirms Population and exclusionsDefines which deals countAnalyst reviews filters Date and observation windowsExposes timing and seasonalityBusiness owner approves Segment dimensionsEnables fair comparisonAnalyst checks sample sizes CRM stage historyShows recorded process eventsSystem owner verifies Buyer interview evidencePreserves direct statementsResearch owner checks consent Call and email summariesAdds interaction contextPrivacy owner approves use Product requirementsDistinguishes fit from processProduct owner verifies Pricing and termsAdds commercial contextFinance or legal reviews Competitor evidencePrevents rumor launderingProduct marketing validates Missing-data notesLimits false confidenceAnalyst documents gaps Decision and ownerMakes findings actionableAccountable leader confirms Keep source links in approved systems. Record taxonomy versions and query logic. If one team uses “no budget” for a temporary delay while another uses it for permanent rejection, fix the classification before inviting AI to write a strategy memo about it. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → 10 AI win-loss analysis prompts Replace the brackets with sanitized facts. Treat every output as a draft for human review. 1. Prepare a clean deal evidence brief “Turn this sanitized data dictionary, stage logic, outcome taxonomy, date range, source inventory, exclusions, and known limitations into a win-loss evidence brief: [paste]. Include the unit of analysis, outcome event, observation window, segment fields, source owner, refresh timing, missingness, and privacy classification. Flag ambiguous definitions, duplicate opportunities, changing account identifiers, post-outcome leakage, and fields that should be removed or aggregated.” Run this before requesting patterns. Duplicate opportunities can make one loud account look like several independent losses. Post-outcome notes can contaminate an analysis that claims to describe signals available before the decision. Have revenue operations verify joins, stage timestamps, reopened deals, subsidiaries, currencies, discounts, and outcome definitions. A neat table built on bad identity resolution is still bad evidence—just aligned nicely. 2. Audit evidence quality “Audit these sanitized win-loss evidence sources: [paste]. For each source, list coverage, freshness, collection method, reliability, likely bias, sensitive content, contradictory evidence, and claims it can and cannot support. Check for missing interviews, seller self-reporting, inconsistent loss reasons, stage skipping, changed taxonomy, small samples, and selection effects. Return a prioritized repair list with an owner.” Buyer interviews may overrepresent large or unusually engaged accounts. CRM notes may be richer for diligent sellers. Lost buyers may decline interviews more often than won buyers. None of this makes the sources useless; it makes their limits part of the result. Do not reward teams for better documentation by mistakenly concluding they experience more problems. Sometimes the spike is in reporting, not reality. 3. Separate observations from assumptions “Classify these sanitized deal claims and notes: [paste]. Put each into observed event, direct buyer statement, seller statement, calculated measure, contextual fact, interpretation, hypothesis, unsupported assumption, sensitive inference, or unknown. Cite the source and date. Rewrite unsupported conclusions as validation questions. Do not infer intent, emotion, authority, budget, protected characteristics, or probability.” “Legal review lasted 23 days” is an observed duration. “The buyer said implementation capacity was unavailable” is a direct statement. “They chose the competitor because our champion was weak” is a theory unless stronger evidence supports it. This exercise is especially useful before an executive review, where caveats tend to disappear while charts travel upward. 4. Code buyer interviews without flattening them “Code these de-identified buyer interview notes: [paste]. Preserve each excerpt ID. Propose a hierarchical theme codebook, apply primary and secondary codes, mark ambiguous cases, and report counts with denominators. Separate stated reason, contributing factor, trigger event, desired outcome, decision criterion, and unresolved unknown. Do not fabricate quotes or merge contradictory accounts into one narrative.” “Price” can mean unaffordable total cost, unclear value, an unfavorable contract structure, missing budget, or a negotiating tactic. Treating those as one theme creates a large bucket and a tiny amount of insight. Have two humans review a sample, revise the codebook, and check agreement. The model can accelerate coding; it should not silently decide what buyers meant. 5. Compare won, lost, and no-decision cohorts “Compare these sanitized aggregate measures for won, lost, and no-decision cohorts: [paste]. Confirm outcome definitions, date windows, denominators, sample sizes, and missingness first. Report meaningful differences, uncertainty, contradictory measures, and plausible confounders. Do not call any difference causal. Suggest checks for deal size, segment, region, source, seller tenure, sales-cycle length, product version, discounting, and procurement complexity.” Won deals often progress farther, generate more activities, and contain more completed fields because they survived longer. Raw activity totals can therefore describe opportunity duration rather than effective selling. Use comparable windows and rates. Check whether segments entered through different channels or faced different product versions. If you need help defining useful groups, use AI customer segmentation prompts. 6. Investigate no-decision outcomes “Analyze this de-identified evidence for no-decision opportunities: [paste]. Separate confirmed project cancellation, budget delay, priority change, internal build, status quo, procurement failure, lost contact, and unknown. Create a timeline for each evidence-backed pattern. List competing explanations and the next evidence needed. Do not relabel silence as a buyer motive.” No-decision is not a softer competitive loss. The buyer may agree that the problem exists and still choose not to change. That can indicate timing, internal capacity, unclear urgency, implementation risk, weak sponsorship, or missing evidence. The correct response is not always a new sales script. It may be qualification discipline, a smaller implementation path, better proof, or accepting that the project was never real. 7. Analyze competitor mentions carefully “Review these sanitized buyer statements, seller notes, and verified public competitor facts: [paste]. Separate direct buyer comparison, seller interpretation, public fact, rumor, and unknown. Build a comparison matrix by decision criterion with source and date. Flag stale claims, contradictory evidence, unsupported feature assertions, and conclusions requiring product, legal, or competitive-intelligence review.” Competitor analysis easily becomes fan fiction. A seller’s note that “they liked Competitor X’s security” does not establish which control mattered, whether it was decisive, or whether the buyer accurately understood either product. Never ask a model to invent weaknesses or generate deceptive battlecards. Validate claims against current evidence. Use findings to improve discovery and product understanding, not to produce confident mud-slinging. 8. Challenge a win-loss hypothesis “Pressure-test this hypothesis: [paste]. First state the exact claim and evidence required. Then list supporting evidence, contradictory evidence, missing data, alternative explanations, possible confounders, and ways the analysis could be biased. Propose the smallest ethical validation steps. Return confirmed facts, plausible hypotheses, rejected claims, and unresolved questions separately.” Use this when the room has already decided that losses come from price, onboarding, one competitor, or a specific seller behavior. A hypothesis should earn confidence; repetition in meetings is not a validation method. For a more general challenge process, AI decision-making prompts can help expose assumptions. Root-cause analysis prompts are useful when the issue crosses product, process, and commercial systems. 9. Create a human-owned interview plan “Create a human-owned win-loss interview plan for this decision: [paste]. Propose a balanced sampling frame across outcomes, segments, deal sizes, regions, and time periods. Draft neutral questions that distinguish decision criteria, process experience, alternatives, implementation concerns, commercial factors, and unresolved needs. Include consent language, data-handling rules, interviewer guidance, non-leading probes, refusal handling, and a plan for documenting uncertainty.” Do not let account teams hand-pick only friendly buyers. Do not offer incentives that pressure participation or ask questions designed to confirm the preferred story. Decide whether sellers should attend; their presence can change what buyers say. Interview notes should identify direct quotes, paraphrases, and interviewer interpretation separately. If recording is permitted, document consent and retention. If it is not, do not improvise because transcription would be convenient. 10. Turn findings into an accountable action brief “Convert these human-reviewed win-loss findings into an action brief: [paste]. Include the evidence-backed finding, affected segment, source, sample size, confidence, competing explanation, recommended action, owner, due date, success measure, risk, guardrail, validation method, and stop condition. Separate immediate data repairs, research tasks, sales-process changes, product questions, enablement needs, and decisions requiring leadership approval.” A finding without an owner becomes trivia. An action without a measure becomes theater. “Improve messaging” is not a plan; “test a revised implementation-risk explanation with this segment, review recordings with consent, and compare qualified progression for six weeks” is at least inspectable. Run a risk assessment before changing pricing, qualification, incentives, or rep evaluation. Never use thin win-loss analysis as an automated employee score. Coaching and performance decisions require complete context, fair process, and accountable human review. Common ways AI win-loss analysis goes wrong It treats CRM labels as buyer truth A required dropdown can improve consistency, but the selected value may be a seller’s interpretation, a convenient close reason, or the least-wrong option. Preserve the distinction between recorded label and verified buyer statement. It invents causation Deals with executive sponsors may win more often. That does not prove that adding an executive contact causes victory. Deal quality, account size, urgency, and sales effort may influence both. It turns incomplete evidence into rep scoring Documentation quality, territory, segment, product fit, inherited pipeline, and manager practices all affect records. AI summaries are not a fair basis for ranking people. Use them to locate questions, never to automate blame. It ignores no-decision A competitor is emotionally satisfying because it gives the loss a face. Status quo is often more complicated and strategically useful. Keep no-decision visible. It leaks confidential data Contracts, call recordings, pricing exceptions, and roadmap discussions are sensitive. “We removed the buyer’s name” is not sufficient when the remaining details identify the company or person. It produces a dashboard nobody acts on Require an owner, validation step, measure, review date, and stop condition. Otherwise the analysis becomes another polished artifact living peacefully beside the problems it describes. A practical review checklist Before using an AI-assisted finding, confirm: Outcome definitions and exclusions are documented. Every claim links to an approved source and date. Direct buyer statements are separated from seller interpretations. Won, lost, and no-decision deals are not mixed carelessly. Sample size, segment, missingness, and selection bias are visible. Correlation is not described as causation. Competitor claims are verified and current. Buyer and employee data was handled under approved rules. A human owner reviewed calculations and interpretations. Every proposed action has a measure, guardrail, and validation step. If several boxes fail, the answer is not a prettier prompt. Repair the evidence process first. Frequently asked questions Can ChatGPT perform a complete win-loss analysis? No. It can organize sanitized evidence, code text, draft comparisons, and identify missing information. Humans must define outcomes, verify data, conduct buyer research, assess bias, interpret results, and own decisions. What data should I give an AI win-loss analysis prompt? Use the minimum approved, sanitized evidence needed for the question: documented outcomes, aggregate stage history, de-identified interview excerpts, verified requirements, segment definitions, and known limitations. Keep source records in approved systems. How many deals do I need for win-loss analysis? There is no universal number. It depends on segmentation, outcome mix, variation, evidence quality, and the decision. Small samples can support qualitative learning when labeled honestly; they should not support precise population claims. Should won and lost deals be compared directly? Only after making windows, segments, definitions, and exposure reasonably comparable. Won opportunities often survive longer and accumulate more activity. Include no-decision outcomes rather than forcing every result into win or competitive loss. Can AI identify the real reason a deal was lost? It can summarize recorded reasons and propose hypotheses. It cannot access unspoken buyer deliberations or establish the “real reason” from incomplete notes. Direct research and triangulation remain necessary. Can these prompts analyze competitors? They can organize verified buyer statements and current public facts. They should not invent competitor weaknesses, repeat rumors, or turn seller impressions into product truth. Human competitive-intelligence and legal review still matter. Is it appropriate to use AI win-loss analysis for salesperson performance reviews? Not as an automated scoring system. Deal records are incomplete and shaped by territory, segment, product fit, inherited pipeline, and documentation habits. Performance decisions require fair process, complete context, and accountable managers. How often should a win-loss analysis be updated? Review it when enough new evidence accumulates and when pricing, positioning, product, market conditions, or sales processes change. Keep dated snapshots so teams can distinguish a real shift from a rewritten taxonomy. The useful boundary AI is fast at restructuring text and finding candidate patterns. It is not the buyer, the research lead, the sales manager, or the person accountable when a bad conclusion changes pricing, product priorities, or someone’s career. Use these prompts to make evidence easier to inspect and assumptions harder to hide. Then do the human work: talk to buyers, check the numbers, protect the data, compare competing explanations, and own the decision. That boundary is the point of what AI can and can’t do and the broader no-BS guide to using AI at work. For more practical rules about keeping judgment and accountability human, Don’t Replace Me by Dmitry Kargaev is the field guide—not a permission slip to outsource your brain. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Customer Renewal Prompts: Prepare With Evidence 10 AI customer renewal prompts to verify outcomes, expose account risks, prepare options, and protect trust without inventing value or customer sentiment. AI promptscustomer renewalcustomer success AI Deal Desk Prompts: Review Complex Deals 10 AI deal desk prompts for checking discounts, terms, risks, approvals, and complex sales deals without inventing policy, evidence, or authority. AI promptsdeal deskrevenue operations AI Sales Enablement Prompts: Build Useful Materials 10 AI sales enablement prompts for grounded messaging, training, coaching, role-play, objection practice, and seller resources without invented claims. AI promptssales enablementsales training --- ## AI Workflow Audit Prompts: 10 Templates URL: https://dontreplace.me/articles/ai-workflow-audit-prompts-find-the-bottlenecks ai workflow auditworkflow automationAI promptsprocess improvementAI at work AI Workflow Audit Prompts: 10 Templates 10 copy-paste AI workflow audit prompts for ChatGPT and Claude. Map processes, find bottlenecks, and sort tasks by automation potential safely. June 15, 2026 Updated June 15, 2026 Before you touch a single automation tool, you need AI workflow audit prompts. Not a consultant's PowerPoint. Not a 40-slide "digital transformation roadmap." A real look at how work actually moves through your team, where it gets stuck, and which parts are worth automating versus which parts will explode faster if you automate them. Most workflow problems aren't automation problems. They're clarity problems. People don't know who owns what. Handoffs happen over Slack DMs that nobody saves. The "process" lives in one person's head and they're on parental leave. Automating that isn't efficiency. It's just a faster way to lose the same stuff. That's where AI actually helps. Not as a replacement for operational judgment, but as a fast thinking partner for mapping, questioning, and structuring what you already know. Used right, it cuts the prep work for a workflow audit from days to hours. Here's how to use it without breaking anything. What an AI workflow audit actually is It's not magic. It's a structured way to examine a business process: what triggers it, who does what, where time gets lost, and what could reasonably be handed to a machine. An AI tool can help you build the questions, map the steps, draft the SOP, and spot the obvious failure points. It cannot tell you what matters to your customers, whether your security team will approve an integration, or what the real story is behind that one step everyone says takes 20 minutes but actually takes four hours because the legacy system crashes every Tuesday. That's still your job. Which is good news, by the way, if you've been reading what AI can and can't do. Before you open ChatGPT or Claude, one rule that never changes: do not paste customer PII, employee records, login credentials, contracts, legal disputes, private financial data, confidential strategy, security architecture, or proprietary workflows into any AI tool that hasn't been approved by your IT or legal team. Describe the process in general terms. Use placeholders. Keep the sensitive details in your head or in your secure systems. The AI workflow audit prompt formula Every good workflow audit prompt follows the same shape. Think of it as giving AI the role, the context, and the specific output you need. Role: You are a [operations analyst / process consultant / workflow mapper]. Context: I'm auditing a [type of process] in a [team/industry] context. [Brief, sanitized description of the process.] Task: [Specific deliverable: map the steps, find bottlenecks, draft interview questions, etc.] Constraints: [What to avoid, what format to use, what level of detail.] That's it. The more specific your context, the less generic the output. Garbage in, garbage out, as Don't Replace Me puts it in Rule #13. A prompt that says "help me audit my workflow" will get you a Wikipedia article. A prompt that describes the actual process, the team size, and the failure you're trying to solve will get you something useful. 10 AI workflow audit prompts you can use today These are copy-paste ready. Adjust the bracketed sections for your situation. All assume you're working with sanitized, non-confidential process descriptions. Prompt 1: Map the current workflow "You are a process mapping analyst. I'll describe a workflow in plain language. Your job is to turn it into a structured step-by-step map with: the trigger that starts the process, each step in sequence, who is responsible for each step, what the input and output of each step is, and how the process ends. Here's the workflow: [describe it in general terms]. Format as a numbered list. Flag any steps where the owner is unclear." Use this first. Before you can fix anything, you need to see what actually exists, not what the wiki says exists. Prompt 2: Find the bottlenecks "Based on this workflow description: [paste your step map from Prompt 1], act as an operations consultant and identify the three most likely bottleneck points. For each one, explain: what typically causes that bottleneck, what the downstream impact is, and what questions I should ask the team to confirm whether this is actually a problem. Do not make up specific metrics or timeframes. Just give me the diagnostic questions." The "do not make up metrics" constraint matters. AI will confidently invent a "typical delay of 3-5 business days" if you let it. You want the questions, not the invented answers. Prompt 3: Identify handoff failures "Review this workflow: [describe it]. Focus specifically on the handoffs between people or teams. For each handoff, tell me: what information needs to transfer, what format it's usually in, what the most common failure mode is when handoffs go wrong, and what I'd see in the output if the handoff failed. Give me a list of handoff risk questions I can bring to a team interview." Handoffs are where most workflows actually die. This prompt forces AI to think about the seams, not just the steps. Prompt 4: Sort tasks by automation potential "Here is a list of tasks in this workflow: [list them]. For each task, rate its automation potential as High, Medium, or Low based on these criteria: High = repetitive, rule-based, no judgment required, clear inputs and outputs. Medium = mostly rule-based but requires occasional human review. Low = requires judgment, context, relationship, or physical presence. Then flag any High-rated tasks that might involve personal data or compliance requirements that need review before automating." This is the most practically useful prompt in the list. It stops you from automating the interesting stuff and ignoring the grind. Pair it with the AI automation prompts guide for the actual implementation step. Prompt 5: Turn messy notes into an SOP draft "I have rough notes from a workflow walkthrough. Turn them into a structured SOP draft with: a one-sentence purpose statement, a list of who this SOP applies to, step-by-step instructions in plain language, a notes section for common exceptions, and a section for 'what to do when it breaks.' Here are my notes: [paste your notes]. Mark any steps where I've left a gap or contradiction with [REVIEW NEEDED]." The [REVIEW NEEDED] flag is your friend. It stops AI from silently papering over gaps with confident-sounding nonsense. For more on turning notes into clean documentation, see the AI documentation prompts article. Prompt 6: Create a stakeholder interview guide "I'm interviewing frontline team members and process owners about a [type of workflow]. Generate a stakeholder interview guide with: five questions for people who execute the process daily, five questions for the manager or process owner, three questions about the customer or end-user experience, and two questions about what they'd change first if they could. Keep the questions open-ended. Avoid leading questions that assume there's a problem." Don't skip the frontline interviews. AI can map a process from a description. It cannot tell you that the form everyone submits is actually filled out wrong half the time because the instructions are confusing. That comes from the people doing the work. Prompt 7: Spot data and privacy risks "Review this workflow description: [describe it]. Identify any steps where personal data, customer information, employee records, or sensitive business information might be processed, stored, or transmitted. For each step you flag, list the type of data involved, the potential risk if that step goes wrong, and the questions I should bring to my security or privacy team before automating it. Do not give legal advice or make compliance determinations." That last constraint is not optional. AI will write convincing-sounding GDPR guidance. It is not a lawyer. Your legal team is. Prompt 8: Compare manual vs automated options "For this workflow step: [describe one specific step], give me a side-by-side comparison of doing it manually versus automating it. Include: estimated effort for manual (describe the activities, not invented time estimates), what could go wrong with each approach, what skills or tools each requires, and what questions I'd need to answer before choosing automation. Frame this as a decision aid, not a recommendation." "Decision aid, not a recommendation" is the key phrase. You want the thinking structure, not a confident answer based on zero knowledge of your actual team and tools. Prompt 9: Build a small experiment plan "I want to run a small pilot to test whether automating [specific step] is worth pursuing. Help me design a 2-4 week experiment plan that includes: what we're testing, what success looks like in measurable terms, what we'll keep doing manually as a control, what could go wrong and how we'd catch it early, and what we'd need to see to decide whether to expand or stop. Keep the scope small and reversible." Small pilots before big commitments. Every time. The teams that automate carefully are the ones who can actually roll it back when it goes sideways. Prompt 10: Write a before/after workflow summary "Based on this current workflow description [paste it] and these proposed changes [describe them], write a before/after summary for stakeholders. Include: what the process looked like before, what's changing and why, what stays the same, what risks we're managing, and what success looks like in the first 30 days. Write it at a level a non-technical manager can read and approve. Flag any claims about time savings or efficiency gains with [VERIFY] since I need real data to support those." That [VERIFY] flag again. Do not let AI write "this will save 40% of processing time" in a document that goes to leadership. That number will become a promise. Promises become expectations. Expectations become performance reviews. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What to do with the outputs AI output from these prompts is a starting draft, not a finished product. Every single one needs a human review pass that checks against: The source of truth. Does this match what's actually documented, or what people told you in interviews? The process owner. Have the person responsible for this workflow confirmed the map is accurate? If they haven't seen it, it's fiction. Frontline reality. Does what's written match how the work actually gets done, or how it's supposed to get done in theory? Privacy and security review. Before anything gets automated, someone qualified needs to check whether that automation touches data it shouldn't. Realistic success metrics. "Reduce errors" is not a metric. "Reduce form rejection rate from X% to Y% within 60 days" is a metric. Make the AI flag where you need real numbers. For anything that touches project planning and resourcing, the AI project management prompts have templates for the next step. And if your audit uncovers data that needs analysis, the AI data analysis prompts can help you structure that work. What AI should never decide This needs saying plainly. AI should not invent your compliance requirements. It does not know your regulatory context, your industry-specific obligations, or what your contracts actually promise customers. AI should not set headcount plans. "This automation could replace two FTEs" is a real decision with real consequences for real people. That comes from operational judgment and HR, not a language model. AI should not make ROI claims. It will make them confidently if you ask. They will be based on nothing. Mark them for verification or remove them. AI should not approve its own outputs. Every workflow map, SOP draft, and experiment plan needs a human with actual authority to review and sign off. The book Don't Replace Me covers this tension directly: AI is fast, which is genuinely useful. But speed is not wisdom. The people who build durable processes are the ones who know which corners are dangerous to cut, and no prompt will teach a model that. That judgment is still yours to exercise. Frequently asked questions What is an AI workflow audit? An AI workflow audit uses AI tools to help map a business process, identify bottlenecks, flag handoff failures, and sort tasks by automation potential. AI assists with structuring and drafting, but human review from process owners and frontline staff is required to confirm accuracy. The outputs are drafts, not decisions. Can I paste my team's workflow documents into ChatGPT? Only if those documents have been cleared by your IT or legal team. Don't paste customer PII, employee records, contracts, credentials, confidential financial data, or proprietary processes into tools that haven't been approved for that use. Describe workflows in general terms and use placeholders for sensitive details. How do I know which workflow steps to automate? Rate each step by three criteria: how repetitive it is, whether it requires human judgment or context, and whether it involves sensitive data that needs special handling. High-automation-potential steps are rule-based, have clear inputs and outputs, and require no judgment calls. Anything requiring relationship, context, or compliance review stays human-led for now. Will AI workflow audit prompts replace a consultant? No. AI can help you structure your thinking, draft questions, and map processes faster than starting from scratch. It can't interview your team, understand your company's political context, or assess the real risk of a change. A good consultant brings operational judgment. AI brings speed. Use them for different things. How do I make sure AI doesn't invent fake metrics in workflow documents? Add explicit constraints to your prompts: "Do not include specific time or cost estimates without flagging them for verification." Use [VERIFY] or [REVIEW NEEDED] placeholders throughout. Never let AI-generated efficiency claims go into a leadership document without real data behind them. What should I do after running these prompts? Get the outputs in front of your process owner, run at least one frontline interview to pressure-test the map, have your security or privacy team review anything that touches sensitive data, and design the smallest possible pilot before committing to full automation. See the guide on using AI at work for how to build that habit across your team. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Automation Prompts: Automate the Boring Stuff 10 copy-paste AI automation prompts for mapping workflows, writing SOPs, adding approval gates, and QA testing. Practical templates with safety rules included. ai automation promptsworkflow automationSOP templates AI Customer Service Prompts: 10 Templates That Work 10 copy-paste AI customer service prompts for ChatGPT and Claude. Draft replies, de-escalate complaints, and summarize tickets without sounding like a robot. AI promptscustomer serviceChatGPT at work AI Launch Checklist Prompts: Ship Without Gaps 10 copy-paste AI launch checklist prompts for ChatGPT and Claude. Turn messy plans into structured checklists with owners, risks, and rollback steps. ai promptslaunch checklistproject management --- ## AI Workload Management Prompts: Stop Drowning URL: https://dontreplace.me/articles/ai-workload-management-prompts-stop-drowning ai workload managementworkload promptsai at worktask prioritizationcapacity planning AI Workload Management Prompts: Stop Drowning 10 copy-paste AI workload management prompts to make invisible work visible, rank priorities, prep capacity talks, and draft status updates. June 26, 2026 Updated June 26, 2026 Your inbox has 47 unread messages. You have three "quick syncs" on the calendar that will each run 40 minutes. Someone just Slacked you about a deliverable you forgot existed. And somewhere in a shared doc, there's a project that's two weeks behind and nobody has officially said so yet. That's invisible work. The stuff that doesn't show up on a roadmap but eats your week anyway. AI workload management prompts can help you drag it into the light, sort it into something usable, and stop your brain from running the same anxious loop at midnight. But first, the honest version: AI won't tell you what actually matters. It doesn't know your boss's real priorities, which stakeholder is about to go sideways, or that the Q3 launch is already doomed. That's still your job. What AI can do is help you think out loud faster, structure the mess, and spot the category of problem before it becomes a fire. Here are 10 prompts to do exactly that. What these AI workload management prompts are actually for These aren't magic. They're a thinking aid. You give AI the raw material, it gives you a structured draft. You review, adjust, and decide. That's the loop. The prompts below work in ChatGPT, Claude, or any similar tool your company has approved. Before you paste anything: no customer PII, no employee health information, no private performance records, no credentials, no confidential financial data, no legal disputes, no security incidents, no HR issues, no board materials, no sensitive client information. Use anonymized or generalized descriptions when in doubt. If your company hasn't approved a tool for work data, don't use it for work data. Also: AI will make things up if you leave gaps. It'll invent deadlines, assign owners it doesn't know, and produce a tidy workload plan that has nothing to do with your actual capacity. The prompts below are built to minimize that, but you still need to review everything before you act on it or send it to anyone. The reusable formula behind every prompt Every prompt in this list follows the same structure. Once you see it, you can build your own. [Role] + [Task inventory or situation] + [Specific output format] + [Constraints and assumptions to use] That's it. The more specific your input, the less the AI fills gaps with nonsense. Vague task dumps produce polished nonsense, which looks useful and isn't. If you want better outputs, treat context like fuel. If you're new to working with AI this way, the no-BS starter guide is a good place to calibrate expectations before you start. 10 copy-paste prompts for workload management Prompt 1: Make invisible work visible Use this when you feel overwhelmed but can't explain why. Brain-dump everything, including the undocumented stuff. "I'm going to give you a list of everything I'm currently responsible for, including things that aren't formally tracked anywhere. For each item I mention, help me identify: (1) what category of work it is (project, recurring, reactive, relationship, admin), (2) roughly how often it comes up or how much time it takes, and (3) whether it's likely visible to my manager or team. I'll share the list now: [paste your task dump]. Don't invent anything I haven't mentioned. Flag where you need more information." Prompt 2: Turn a messy task dump into workstreams You have 35 things on a list. You need four or five buckets, not 35 line items. "Here is a raw list of tasks and responsibilities I'm juggling right now: [paste list]. Please group these into logical workstreams (no more than six). For each workstream, give it a short name, list which tasks belong to it, and note any that seem duplicated or unclear. Don't assign owners or deadlines. Just organize what I've given you." Prompt 3: Find overload risks before they find you Most overload isn't a surprise in retrospect. This prompt helps you see the shape of it earlier. "Based on this task list [paste list], identify the three to five areas most likely to create overload, conflict, or dropped balls in the next two to four weeks. For each one, describe what the risk is and what kind of decision or conversation it might require. Don't invent deadlines or assumptions. Use only what I've given you and flag anything that's unclear." Prompt 4: Rank work by impact and urgency The classic Eisenhower matrix is fine. This version adds a business priority lens. "Here are my current tasks: [paste list]. My top business priority right now is [describe the one thing that matters most to your team or company this period]. Please sort my tasks into four buckets: high impact + urgent, high impact + not urgent, low impact + urgent, low impact + not urgent. Flag any tasks where you don't have enough context to sort them confidently. Don't assume urgency I haven't told you about." For a deeper version of this kind of prioritization work, the AI prioritization prompts article has templates specifically built for backlogs and roadmaps. Prompt 5: Prepare a manager capacity conversation This is the one most people avoid. This prompt helps you walk into that conversation with something concrete instead of vague stress. "I need to have a conversation with my manager about workload capacity. I currently have [X] active priorities. I have capacity for roughly [Y hours or days] of focused work per week after meetings and recurring tasks. I'm concerned about [specific area of overload or conflict]. Please help me draft: (1) a one-paragraph summary of my current load and the specific conflict or constraint, (2) two or three options I could propose, and (3) the key question I need them to answer. Use only what I've told you. Don't invent timelines or business context." This kind of prompt pairs well with the AI manager prompts collection if you want to prep the full 1:1 agenda, not just the capacity piece. Prompt 6: Draft a stakeholder tradeoff note Someone asked for something. You can't do all of it. This prompt helps you explain why without making it a complaining email. "I need to send a note to [role, not name] explaining that I can deliver [X] by [date], but not [Y] in the same timeframe. The business reason is [brief reason]. Please draft a short, professional note that: states what I can deliver, explains the constraint clearly, offers one or two alternative options or timelines, and asks for their input on how to proceed. Keep it under 150 words. Don't make up deadlines or commitments I haven't specified." Prompt 7: Build a weekly focus plan Monday morning chaos is optional. This prompt turns your messy list into a structured week. "It's the start of the work week. Here's my current task list: [paste list]. My most important deadline this week is [deadline + what it is]. I have [X] hours of unblocked focus time available. Please help me draft a simple day-by-day focus plan that protects time for the highest-priority work first, batches similar tasks where possible, and flags anything that probably shouldn't be on this week's list at all. Don't invent time estimates I haven't given you." Prompt 8: Spot dependencies and blockers Things that look like your problem are often someone else's problem that landed on you. This prompt finds those. "Here is my current task list: [paste list]. For each task, help me identify: (1) whether it has a dependency on another person, team, or system I haven't mentioned, (2) whether it's currently blocked by anything, and (3) whether I'm a blocker for someone else and might not realize it. Use only what I've given you. Flag where you're guessing." The workflow audit prompts collection goes deeper on this if you're doing a full process review, not just a weekly check-in. Prompt 9: Create a workload review agenda If you run a team, a regular workload review beats a fire drill. This prompt builds the agenda. "I want to run a 30-minute workload review with my team. Our current focus areas are [list two to three]. The main risks I'm aware of are [list]. Please draft a meeting agenda that covers: a check-in on current load and blockers, a review of priorities for the next two weeks, a discussion of any tradeoffs or resource conflicts, and clear decisions or action items we need to leave with. Suggest time allocation for each section. Don't assume specific names, deadlines, or team dynamics." Prompt 10: Turn progress notes into a capacity-aware status update Status updates usually describe what happened. This one also signals what's at risk. "Here are my notes on work completed and in progress this week: [paste notes]. Please help me draft a status update for [audience: team, manager, or stakeholders] that covers: what's done, what's in progress and where it stands, what's blocked or at risk, and what I need from others to keep things moving. Keep it under 200 words. Flag any commitment I should double-check before sending. Don't invent completion percentages or deadlines I haven't stated." For more status update templates across different audiences, the AI status report prompts article has the full set. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What AI should never do in workload management This is the part people skip. Rule #7 in Don't Replace Me is about taste being your moat: knowing what matters, what can slip, which promise was never realistic. AI has no taste. It'll produce a beautifully formatted workload plan that puts a low-stakes admin task above a relationship-critical deliverable, and do it with complete confidence. Here's the short list of things AI should not be doing when you're managing workload: Inventing deadlines you haven't confirmed Assigning owners it doesn't know Estimating team capacity without real data Making employee performance judgments Approving budget decisions or resource allocation Providing legal, security, or HR guidance Making commitments to customers or stakeholders Setting strategic priorities for your organization AI is useful for the inventory, the structure, the first draft of the tradeoff memo. The actual decisions still need a human who knows the full context. That's not a limitation to work around. It's the job. If you're not sure where AI helps and where it makes things worse, the plain-language breakdown of what AI can and can't do is worth five minutes before you hand anything important to a chatbot. One more thing before you start The biggest workload problem most people have isn't a tool problem. It's a visibility problem. The work is real, the capacity is real, but nobody has said out loud what has to give. These prompts help you see the shape of the problem faster. They don't negotiate your priorities with your manager, they don't push back on the stakeholder who keeps adding scope, and they definitely don't have the conversation with the person on your team who's running on empty. That's still on you. The prompts just make it easier to walk in prepared. If you want the full framework for staying useful when everything is changing fast, the book is the longer version of all of this, including the parts where tidy lists and real leadership are not the same thing. Frequently asked questions What are AI workload management prompts? AI workload management prompts are structured instructions you give to a tool like ChatGPT or Claude to help organize, prioritize, and communicate about your workload. They work best when you give the AI specific context: your actual tasks, real deadlines, capacity limits, and business priorities. Without that, the output is polished but generic. Can AI actually help me manage my workload, or is it just another distraction? It can help with specific, structured tasks: turning a messy list into categories, drafting a capacity note to your manager, spotting dependencies, or writing a status update. It won't replace your judgment about what matters most, and it shouldn't. The human value is knowing which priority is real and which is someone else's urgency. See what AI can and can't actually do for a clearer picture. What information should I never paste into an AI tool at work? Never paste customer PII, employee health information, private performance records, credentials, confidential financial data, legal disputes, security incidents, HR issues, board materials, or sensitive client information into unapproved tools. When in doubt, anonymize or generalize the details before using any AI tool for work. How do I stop AI from making things up when I ask for workload help? Be specific and constrained. Tell the AI what you know, what you don't know, and what it should not invent. Phrases like "use only what I've given you," "don't assign owners I haven't mentioned," and "flag where you're guessing" reduce fabrication significantly. Always review before you act on or share any output. What's the difference between AI workload management prompts and just making a to-do list? A to-do list captures tasks. These prompts help you structure them, spot the risks, identify tradeoffs, prepare conversations, and communicate about capacity. The goal isn't a cleaner list. It's a clearer picture of what's happening so you can make better decisions about what to protect, what to drop, and what to escalate. How often should I use these prompts? Weekly planning and status updates work well as a standing habit. Capacity conversations and tradeoff notes come up as needed. Dependency checks are worth doing any time a project feels stuck for reasons that aren't obvious. Don't turn it into a ritual for its own sake. Use it when the mess is real. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Action Item Prompts: 10 Templates That Work 10 copy-paste AI prompts to turn rambling meeting notes into owners, deadlines, and action lists. No invented tasks, no false consensus. ai action itemsmeeting promptsai at work AI Automation Prompts: Automate the Boring Stuff 10 copy-paste AI automation prompts for mapping workflows, writing SOPs, adding approval gates, and QA testing. Practical templates with safety rules included. ai automation promptsworkflow automationSOP templates AI Brainstorming Prompts That Actually Work 10 copy-paste AI brainstorming prompts for work. Generate better ideas, pressure-test them, and shortlist fast, without outsourcing your judgment. ai brainstorming promptsai at workcreative prompts --- ## ChatGPT Prompts at Work That Actually Work URL: https://dontreplace.me/articles/chatgpt-at-work-prompts-that-actually-help chatgpt at workpromptsAI toolswork productivityAI writing ChatGPT Prompts at Work That Actually Work Real ChatGPT prompt templates for emails, reports, and presentations. The role-context-task-format framework that gets you output worth using. May 13, 2026 Updated 13 May 2026 28% of employed US adults already use ChatGPT at work. Most of them are getting mediocre results from their ChatGPT prompts at work and don't know why. The problem isn't the tool. The problem is that people treat ChatGPT like a magic 8-ball. You shake it, type something vague, and wonder why it comes back with corporate word soup that sounds like it was written by a LinkedIn bot having a fever dream. It's not magic. It's more like a very fast, very confident intern who has read everything on the internet and has zero idea what you actually need unless you tell them. Specifically. With context. This guide gives you real prompts for real work tasks: emails, reports, presentations, meeting summaries. Copy them, tweak them, use them today. No $997 course required. Why your ChatGPT prompts at work aren't working The most common prompt people write is something like: "Write me an email about the project update." That's it. That's the whole thing. ChatGPT will write you something. It will be grammatically fine and completely generic. It won't know your tone, your audience, your relationship with the recipient, or what you actually need the email to accomplish. It'll sound like it was written by someone who has never met you, your boss, or your company. Because it was. The output matches the input. Garbage in, garbage out. This is Rule #13 in Don't Replace Me, and it's the one people skip because they assume the AI will just figure it out. It won't. You need to give it four things: a role, context, a task, and a format. Once you start doing that, the output gets dramatically better. Fast. The smart intern framework: the only prompting method you need for ChatGPT at work Think of ChatGPT as a smart intern. Not a genius. Not a mind reader. A fast, capable person who needs clear instructions and knows nothing about your specific situation unless you explain it. Here's the structure: Role: What kind of expert should it be right now? Context: What's the situation? Who's involved? What matters here? Task: What do you actually want it to produce? Format: How should the output look? Bullet points? A draft email? Three options? This isn't complicated. It takes 30 extra seconds to write a better prompt, and it saves you five minutes of editing garbage output afterward. Here's what the same email request looks like with this framework applied: "You're a professional business writer who understands corporate communication. I'm a project manager writing to my VP. Our software launch got delayed by two weeks due to a vendor issue. I need to update her without making the team look incompetent. Write a short email, two to three paragraphs, that's direct, takes ownership, and ends with the revised timeline and next steps." That prompt gets you something you might actually send. Prompt templates for emails you write every week Stop starting from scratch. Here are copy-paste templates for the emails that eat your time. The "bad news" email "You're a professional communicator. I'm a [your role] writing to [recipient and their relationship to you]. I need to share [the bad news] without being defensive. The context is [brief background]. Write a direct email that acknowledges the issue, gives the key facts, and ends with what happens next. Keep it under 200 words." The follow-up that doesn't sound desperate "I sent a proposal to [type of person] two weeks ago. No response. Write a follow-up email that assumes they're busy (not ignoring me), references the original proposal briefly, and asks for a simple yes or a good time to talk. Tone: friendly, professional, not needy. Under 100 words." The meeting recap nobody asked for but everyone needs "Here are my rough notes from today's meeting: [paste notes]. Turn these into a clean recap with three sections: decisions made, action items with owners, and open questions. Use bullet points. Be concise." You can adapt any of these. The structure does the work. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Prompt templates for reports and documents Reports are where most people waste the most time. The writing itself rarely takes long. It's the organizing, the formatting, the figuring out how to say something without sounding like you're hedging everything into oblivion. ChatGPT is genuinely useful here, because structure is exactly what it's good at. Turn bullet points into a structured report section "I have rough notes on [topic]. Turn them into a professional report section with a short intro paragraph, three to four key points with one sentence of context each, and a brief summary. Tone should be clear and confident, not overly formal. Here are the notes: [paste notes]." Executive summary from a long document "Read this document and write a four-sentence executive summary for a senior audience who won't read the full thing. Focus on: what the situation is, what we found, what we recommend, and what happens if we do nothing. [Paste document or key sections]." This works. The broader guide to using AI at work has more on how to build this into a regular workflow, not just one-off tasks. First draft from an outline "Here's an outline for a [type of document]. Write a first draft that I can edit. Write in a clear, direct style, not academic or formal. Assume the reader is a smart professional with limited time. [Paste outline]." You edit. ChatGPT drafts. That's the division of labor that actually saves time. Prompt templates for presentations Most presentations are too long, too vague, and built around what the presenter knows rather than what the audience needs to decide or understand. ChatGPT can help you fix that, if you tell it what you're trying to accomplish. Structure a presentation from scratch "I need to present [topic] to [audience]. The goal of this presentation is to [specific outcome: get approval, share findings, explain a decision, etc.]. I have [X] minutes. Give me a slide-by-slide outline with a one-sentence description of what each slide accomplishes and what goes on it. Don't just list topics. Each slide should do a specific job." Turn a data point into a story "I have this data point: [statistic or finding]. I need to explain why this matters to an audience of [who they are]. Write three to four sentences that turn this number into a clear, compelling point. Not dramatic. Just clear." Speaker notes for a slide "Here's what's on my slide: [paste slide content]. Write 60 to 90 seconds of speaker notes that explain what this slide means in plain language. Don't just read the bullet points. Assume I'll be talking to real people who need context, not a script." The double tap: why you can't skip verification Here's the thing nobody in the "AI will save you hours every day" camp leads with: ChatGPT lies. Confidently. Smoothly. With specific-sounding details that are just wrong. It will cite a statistic that doesn't exist. It will invent a quote. It will give you a legal interpretation that sounds plausible and is completely off. It once told a lawyer that six court cases were relevant to his case. None of them existed. The author behind this site calls this the Double Tap: before anything leaves your hands, you check the work. Every claim that can be verified should be verified. Every number should have a source you actually looked at. Every recommendation that matters to someone else should get a second look. This isn't about not trusting AI. It's about not trusting any single source without checking it. If an intern handed you a report, you'd read it before you sent it up the chain. Same rule applies here. The check doesn't need to be exhaustive. It needs to be proportional to the stakes. Quick email to a colleague? Skim it. Report going to your board? Read every sentence. If you're still figuring out where AI ends and your judgment begins, the article on what AI can and can't actually do is worth five minutes. What to do when the output is bad ChatGPT has an aggravating habit. If you push back, it agrees with you. Completely. Enthusiastically. And then gives you something only slightly different that's still wrong. Don't accept its apologies. Diagnose the actual problem. If the output is too generic, you didn't give it enough context. Add more. Specifics about your audience, your tone, your relationship with the reader. If the tone is off, tell it explicitly. "This sounds too formal. Rewrite it the way I'd talk to a colleague I've worked with for two years." Or: "This sounds too casual. Make it more professional without being stiff." If it's missing the point, your task instruction was probably unclear. Break it down. Tell it the one thing the output needs to accomplish. Not everything it could accomplish. The one thing. If you've edited it twice and it's still not right, just write the damn thing yourself. Sometimes the fastest path is the direct one. AI is a tool, not a replacement for your judgment. That's the whole point. For people wondering about building real AI fluency without going back to school, the guide to AI skills for non-technical people lays out exactly what's worth learning and what's filler. The only setup that actually matters You don't need a paid tier to start. The free version of ChatGPT handles everything in this guide. You don't need plug-ins, extensions, or a custom GPT built by someone selling a course. What you do need is a habit. Pick one task you do every week that involves writing. One. Apply the smart intern framework to that task for two weeks. See what happens. That's it. That's the full onboarding. Once you're consistently getting good output on that one thing, expand. Add another task. Your prompts will get better because you'll start noticing what information ChatGPT needs to do good work versus what you left out. AI is just your keyboard now. You don't take a masterclass on keyboards. You type, you figure out what works, you keep going. Frequently asked questions What's the best way to prompt ChatGPT for work emails? Use the role-context-task-format structure. Tell ChatGPT who it is (a professional business writer), what the situation is (the specific context and relationship), what you need (the actual email), and how you want it formatted (length, tone, structure). Generic prompts produce generic output. Can I use ChatGPT at work without telling my boss? 68% of employees who use AI at work don't tell their employer. Whether you should disclose depends on your company policy and what you're using it for. The full breakdown is in the guide on whether to tell your boss you use AI. How do I stop ChatGPT from writing in that stiff, corporate tone? Tell it explicitly. Add to your prompt: "Write this the way a confident professional would explain it to a colleague, not the way a consultant writes a white paper." Or give it a sample of your own writing and say "Match this tone." Specific tone instructions work better than vague ones like "make it natural." Is the free version of ChatGPT good enough for work tasks? For most work tasks, yes. Writing, editing, summarizing, structuring documents: the free tier handles all of it. The paid version gives you faster responses and access to more recent information, but it's not necessary to get started or to do the tasks in this guide. What should I always check before sending AI-written work? Any specific fact, statistic, or number. Any name, date, or reference to a real document. Any recommendation with legal, financial, or HR implications. The AI will sound confident even when it's wrong. Verification is proportional to stakes: quick email gets a skim, board presentation gets a full read. Do I need to learn prompt engineering to use ChatGPT effectively at work? No. The prompt engineering course industry is largely built on making something simple sound complicated. The role-context-task-format framework in this article is enough for 95% of work tasks. You don't need a certificate. You need to practice on real things. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading How to Use AI at Work: A No-BS Starter Guide Learn how to use AI at work without the hype, the courses, or the 50-tool stack. Start with the tasks you hate and go from there. how to use AI at workAI for beginnersChatGPT at work AI Anxiety at Work: How to Stop Spiraling 65% of employees feel AI anxiety at work. Here's how to convert that fear into something useful without buying a $997 course or losing more sleep. ai anxietyai job fearwill ai replace my job AI Job Replacement Statistics: What the Data Says Goldman Sachs, WEF, McKinsey: what AI job replacement statistics actually say vs. what gets quoted in headlines. The real numbers, with context. AI job replacementAI statisticsfuture of work --- ## How to Future-Proof Your Career Against AI URL: https://dontreplace.me/articles/how-to-future-proof-your-career-against-ai future-proof careerAI career strategycombination moatAI fluencycareer resilience How to Future-Proof Your Career Against AI The combination moat framework: build the career AI can't replace using domain expertise, AI fluency, and relationships. No coding required. May 11, 2026 Updated 12 May 2026 The most common career advice for the AI era is "learn to code." It's wrong. Not slightly off. Wrong. Future-proof your career the right way and coding has almost nothing to do with it. Here's the thing most people miss: AI doesn't care about your individual skills. It cares about what's hard to replicate. A single skill, no matter how polished, is just a target. What makes you hard to replace is the specific combination of things you know, who you know, and how you work. Nobody else has that exact combination. That's the point. This isn't a vibe. It's a framework. And it's what separates the people who come out of this period stronger from the ones who spent two years panicking on LinkedIn and buying $997 prompt engineering courses that are already obsolete. Let's get into it. Why "future-proof your career" advice usually fails Most career advice in the AI era falls into one of two categories. The first is pure doom: AI is coming, update your resume, good luck. The second is fake optimism: just learn some new skills and you'll be fine! Both are useless. The doom camp ignores the fact that AI replaces tasks, not jobs. The hype camp ignores the fact that "learn some skills" without a strategy is how you spend six months on YouTube tutorials and come out the other side with nothing you can actually use at work. The real problem is that most advice treats your career like a list of individual skills that need upgrading. Beef up skill A, add skill B, survive. But that's not how irreplaceability works. Irreplaceability is about combination. It's about the specific mix of what you know, how you think, and who trusts you that nobody else has in quite the same configuration. One chef can julienne carrots faster than any robot. Big deal. Another chef knows the owner, trained half the kitchen staff, and understands the clientele better than anyone. That second person is much harder to swap out, even if the first one has a sharper knife. What the combination moat actually means Forget the word "moat" for a second because it sounds like something a LinkedIn prophet invented at a mastermind retreat. Here's the practical version. Your combination moat is made of four things: Domain expertise. What you know deeply about your specific field. Not generic "project management." Your specific industry, your specific problems, the specific context that took years to accumulate. AI fluency. Not coding. Not machine learning. The ability to use AI tools well, know their limits, and apply them to your actual work. This is a learnable skill in weeks, not years. Human relationships. Clients who trust you. Colleagues who want to work with you. A manager who knows what you're capable of. Referrals. Reputation. None of this lives in a model. Institutional knowledge. The stuff you know because you've been somewhere long enough. Who actually makes decisions. Why the system works the way it does. What failed before you arrived. AI doesn't have access to this and probably never will. The World Economic Forum's 2025 Jobs Report found that the skills most valued by employers over the next five years include analytical thinking, creative thinking, and resilience. Analytical thinking ranked first. None of the top skills require a CS degree. No one skill makes a moat. The combination is the moat. And yours is different from everyone else's, which is exactly the point. The T-shape model: how to future-proof your career without starting over Here's the practical shape of a career that holds up in the AI era. You need depth in one area, and breadth across a few others, including AI fluency. The classic T-shape model gets thrown around a lot in tech recruiting, but it applies here too. Vertical bar: deep expertise in your core domain. Horizontal bar: working knowledge of adjacent areas, and specifically AI tools. Consider what Dee, the author of Don't Replace Me, describes as his own moat: designer by training, AI engineer by practice. "A designer thinks about how it's going to be used. An engineer thinks about how it's going to work." Most designers don't think like engineers. Most engineers don't think like designers. The combination is rare, and rare is valuable. You don't need to be an engineer. You need your version of that cross. A marketer who understands data. A lawyer who understands the business side. A nurse who knows the operational systems. An accountant who can communicate clearly to non-accountants. Add AI fluency to any of these combinations and you've got something that's genuinely hard to automate away. The key insight is this: a generalist who uses AI well beats a specialist who doesn't, at most tasks, most of the time. The specialist's advantage shrinks when AI can cover the gaps in the generalist's knowledge. But the generalist who also has deep expertise in one area? That's the combination that's hard to beat. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Document or die: the rule most people skip This one sounds dramatic and is meant to. If you don't document your expertise, someone else will train the system with worse knowledge than you have, and you'll be outpaced by your own replacement. Here's how it goes wrong. Someone with shallow knowledge of your domain sets up an AI workflow for your type of work. It's mediocre, but it's fast. The people who sign the checks don't know it's mediocre. They see the speed and the cost. You, sitting on years of deep expertise that never got written down or built into any system, look slower and more expensive by comparison. The fix is to be the person who teaches the system. Not passively, not by hoping your value is obvious. Actively. Document your processes. Build the prompts that reflect how you actually think about your work. Be the one in your team who knows how to get the AI to do the task your way, the good way. This isn't about becoming the office AI evangelist. It's about not letting someone with less knowledge build the template that everyone else gets measured against. McKinsey research estimates that generative AI could automate up to 70% of tasks that workers currently spend time on. The question is whether you shape how that automation gets applied to your work, or whether someone else does. For a practical breakdown of how to actually start doing this at work this week, the no-BS starter guide to using AI at work is the right place to begin. Generalists with AI vs.specialists without it This is the shift most people in narrow specializations aren't ready for. For decades, specialization was the safe bet. The world needed more of everything: more tax accountants, more copywriters, more compliance officers. Going deep in a specific niche meant stability. The niche was your protection. That's changing. Not because AI is magic, but because AI closes knowledge gaps fast. A generalist who can use AI to research, analyze, write, and synthesize can now cover ground that used to require a specialist for each domain. The specialist's edge was information asymmetry. AI erodes that asymmetry. This doesn't mean specialization is dead. It means specialization alone isn't enough anymore. The accountant who only knows accounting is more exposed than they were five years ago. The accountant who knows accounting deeply, uses AI fluently, and has strong relationships with clients who trust their judgment? Still very employed. If you're worried about where you sit on this spectrum, the breakdown of which jobs are genuinely hard for AI to replace is worth reading. It cuts through the noise about which skills actually hold their value. The short version: judgment, taste, trust, physical presence, and context are still human advantages. A narrow technical skill that doesn't involve any of those things is more exposed than it used to be. A practical monthly plan to build your moat Vague advice about "staying relevant" helps nobody. Here's something you can actually do. Month one: map your current combination Write down your domain expertise. Not job title, not years of experience. What do you know, specifically? What problems can you solve that most people in your field can't? What context do you have that took time to accumulate? Then write down your relationships: who trusts you, who you have access to that others don't. Then write down your institutional knowledge: what do you know about your organization, your clients, your market that isn't written down anywhere? This is your current moat. Look at it honestly. Where's the depth? Where are the gaps? Month two: add AI fluency to what you already know Don't start with AI in general. Start with the part of your work you hate most, the repetitive, slow, draining stuff. Get ChatGPT or Claude to help with that specific thing. Spend 30 minutes actually trying it. You'll be bad at it. Keep going. The goal isn't to become an AI expert. The goal is to become fluent enough in AI tools to do your actual job better and faster. That's it. The AI skills non-technical people actually need are a lot simpler than the course sellers want you to think. Month three: start documenting Write down how you approach your core work. Your process, your judgment calls, your shortcuts, the things that took you years to figure out. Start building prompts that capture how you think. Do one task the AI-assisted way and write down what worked and what didn't. This is how you become the person who trains the systems instead of the person the systems replace. Month four and beyond: extend the combination Find one adjacent skill that would make your combination rarer. If you're technical, learn the business side. If you're creative, learn the data side. If you're an individual contributor, learn how to communicate your work to non-specialists. Pick one. Don't chase every trend. One genuine extension is worth more than ten half-learned certifications. When to dig in and when to walk away Not every situation rewards staying. Sometimes the right future-proofing move is leaving. If you're in a role where 80% of the work is routine information processing and there's no path to more judgment-intensive work, that's worth thinking about honestly. Not with panic, but with clear eyes. Some jobs are going to get genuinely smaller. Some industries are moving faster than others. The question isn't "is AI threatening my field?" Almost every field has some exposure. The question is: "Is there a version of my work in this field that requires the combination I can build?" If yes, build toward that version. If the honest answer is no, better to figure that out now and move on your terms than to wait for the decision to be made for you. Rule #20 in Dee's framework for this exact situation is simple: know when to dig in and when to bail. Playing the long game doesn't mean staying in a losing position out of stubbornness. Frequently asked questions How do I future-proof my career against AI without learning to code? You don't need to code. The combination that makes you hard to replace is domain expertise plus AI fluency plus human relationships plus institutional knowledge. AI fluency means using tools like ChatGPT and Claude for your actual work, not building them. Most people can get genuinely useful at this in a few weeks of regular practice. What does "AI fluency" actually mean for non-technical workers? It means knowing how to get AI tools to do useful things for your specific job. Writing prompts that produce good output, knowing when to trust the results and when to check them, and applying AI to the parts of your work that are repetitive or time-consuming. For more on this, the no-BS guide to using AI at work covers the practical steps. Is specialization still a good career strategy in the AI era? Specialization alone is more exposed than it used to be. Deep expertise in one area is still valuable, but it works best when combined with AI fluency and strong relationships. A specialist who uses AI well and has domain knowledge that took years to accumulate is in a strong position. A specialist with narrow technical skills and no adjacent capabilities is more at risk. How long does it take to build a combination moat? The domain expertise you may already have. Relationships take time and can't be rushed. AI fluency is the fastest part: a few weeks of actual practice on your real work is enough to get functional. The combination starts providing protection faster than most people expect, because most people aren't building one at all. Should I stay in my current field or switch to something "more AI-proof"? Most fields have roles that are more exposed and roles that are less. The question is whether your specific role or trajectory has room to move toward more judgment-intensive, relationship-dependent work. If yes, build toward that. If the honest answer is that the trajectory is blocked, it's worth evaluating a move while you still have options. Waiting for the situation to force your hand is the worst version of this decision. What's the single most important thing I can do this month to protect my career? Map what you actually know. Write down your specific domain expertise, your relationships, and your institutional knowledge. Most people have more of a moat than they think, and they've never looked at it clearly. Once you can see it, you can build on it. That's the starting point everything else follows from. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Skills for Non-Technical People (2026) The AI skills non-technical people actually need in 2026. No coding, no bootcamp, no $997 course. Four real skills and a minimum viable stack by profession. ai skillsnon-technicalprompt engineering AI Anxiety at Work: How to Stop Spiraling 65% of employees feel AI anxiety at work. Here's how to convert that fear into something useful without buying a $997 course or losing more sleep. ai anxietyai job fearwill ai replace my job AI Job Replacement Statistics: What the Data Says Goldman Sachs, WEF, McKinsey: what AI job replacement statistics actually say vs. what gets quoted in headlines. The real numbers, with context. AI job replacementAI statisticsfuture of work --- ## How to Use AI at Work: A No-BS Starter Guide URL: https://dontreplace.me/articles/how-to-use-ai-at-work-a-no-bs-starter-guide how to use AI at workAI for beginnersChatGPT at workAI productivityworkplace AI How to Use AI at Work: A No-BS Starter Guide Learn how to use AI at work without the hype, the courses, or the 50-tool stack. Start with the tasks you hate and go from there. May 12, 2026 Updated 13 May 2026 28% of employed American adults already use ChatGPT for work. Most of them figured it out in an afternoon, not a bootcamp. Nobody gave them a certification. They just started typing. If you've been waiting for the right moment to start using AI at work, this is your sign that the right moment was six months ago. But today is fine too. Here's how to actually do it, without the $997 course, the 47-tool stack, or the LinkedIn posts about your "AI journey." How to use AI at work without overcomplicating it The mistake most people make is treating this like a new skill that requires study. It's not calculus. It's closer to learning to use Google, except the results come out in full sentences and you have to tell it what you actually want. Start here: open ChatGPT or Claude. Pick one. They're both free at the basic tier. Type something work-related that you'd normally spend 20 minutes on. See what comes back. That's it. That's the orientation. Everything after this is just getting better at describing what you need. The reason people don't start isn't that they lack tools or knowledge. It's that they're waiting to feel ready. You won't feel ready. You'll feel ready after you've used it a dozen times and it's saved you two hours on something tedious. Start with the shit you hate This is Rule #12 in Don't Replace Me, Dmitry Kargaev's field guide for normal people with real jobs. He calls it the 40% Rule: roughly 40% of most knowledge workers' tasks can be assisted by AI right now, today, without any special training. The trick is to start with the worst 40%, not the interesting parts of your job. Think about your week. What made you want to stare at the ceiling? The status update email nobody reads but everyone demands. The meeting notes you have to turn into action items. The job description HR asked you to write. The slide deck that's basically a bulleted list of things everyone already knows. The contract summary your client needs by Thursday. The performance review boilerplate. That's your starting list. Not "how can AI help me think more strategically." Not "how do I use AI to disrupt my industry." Just: what is the tedious, time-draining, low-creativity work that fills your calendar and makes you feel like a human copy-paste machine? Start there. It's unglamorous. It also works immediately. You'll get your first real win in 20 minutes, and that's what gets you hooked on the right thing: time back, not novelty. The smart intern framework: how to actually prompt it Bad prompts get bad results. Not because AI is dumb, but because you're talking to something that can't read your mind and doesn't know your context. Rule #13 from the book puts it plainly: talk to it like a smart intern. A smart intern is capable, eager, and completely missing your context. They haven't sat in your meetings. They don't know your clients. They don't know your boss hates bullet points or that "streamline" is a forbidden word at your company. You have to tell them. Here's a simple structure that works: Role: Tell it what it is ("You're a senior HR manager at a mid-size tech company") Context: Tell it what it needs to know ("We're doing annual reviews and I need to give feedback on someone who's technically strong but struggles with stakeholder communication") Task: Tell it what to do ("Write a draft performance review comment that's honest but constructive") Format: Tell it how you want the output ("Keep it under 150 words, no jargon, no bullet points") That's it. Four things. You don't need to learn "prompt engineering." Anyone selling you a course on prompt engineering in 2025 is selling you something that was barely useful in 2023 and is mostly irrelevant now. The models got smarter. Clear instructions are enough. If you don't love the first result, just say "make it shorter" or "make it less formal" or "that sounds corporate, rewrite it." It's a conversation. Go back and forth until you get something usable. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What this looks like by role Different jobs have different tedious work. Here's where to start depending on what you actually do: RoleStart here MarketingFirst drafts of emails, ads, social posts, briefs FinanceSummarizing reports, writing commentary on data, formatting HRJob descriptions, offer letter language, policy summaries OperationsProcess documentation, meeting recaps, vendor comms LegalContract summaries (not legal advice, summaries) SalesProspecting emails, call prep, CRM note cleanup Managers (all)Performance review drafts, agenda creation, status updates The pattern is the same in every row: the output that somebody needs but that you'd rather not spend two hours writing. AI is not replacing your judgment here. It's replacing the blank page. You review it. You edit it. You add the context it couldn't know. The first draft is just the part that was eating your afternoon. Stop collecting tools Here's where people lose the plot. They read a roundup article about "50 AI tools every professional needs" and spend three weeks exploring products instead of using AI for actual work. Most of those tools are wrappers. A wrapper is a product built on top of ChatGPT or Claude with a prettier interface and a monthly subscription. Sometimes they're useful for specific tasks. Usually you don't need them. The author of Don't Replace Me, who has built AI systems professionally, puts it plainly: "You don't need anything but Claude or ChatGPT. Everything else is mostly wrappers." Pick two, maybe three tools total. For most people that's: ChatGPT (GPT-4o) for general writing and research tasks Claude (claude.ai) for longer documents, analysis, and writing that needs to sound more human Optionally: whatever AI is baked into the software you already use (Microsoft Copilot in Office, Gemini in Google Workspace) That's a complete stack. You don't need Jasper, Copy.ai, Notion AI, and six browser extensions. Every hour you spend evaluating new tools is an hour you didn't spend using AI for actual work. Master the two you have. Get fast at them. Then, if there's a specific workflow where another tool genuinely helps, you'll know it because you'll hit a real limitation, not because someone wrote a sponsored post about it. The question nobody asks: should you tell anyone? This comes up sooner than you'd expect. You start using AI, you get good results, and then you wonder whether to mention it to your manager or just quietly keep doing it. The answer is more complicated than it sounds. About 68% of workers who use AI at work don't tell their employer, according to Microsoft's 2024 Work Trend Index. Some of that is caution about company policy. Some of it is fear of being judged. Some of it is just: nobody asked. The short version: check if your company has an AI policy. Many do now, and they often restrict what data you can paste into external tools. If your company has one, know it. Don't paste confidential client data into a free-tier AI tool if the policy says not to. That's not paranoia, it's professional hygiene. If your company doesn't have a policy, you're in the vast majority. Use common sense about what you share. Keep proprietary information vague or anonymized when you're using AI to help with it. "A manufacturing client in the midwest" instead of the actual company name. It's not complicated. As for disclosure: you don't have to announce it. You also shouldn't lie about it. If someone asks how you turned that report around so fast, "I used AI to get a first draft and then edited it" is a fine and true answer. If that makes your boss uncomfortable, that's a different conversation. Your first 72 hours: a simple plan Don't make this a project. Don't schedule a "learning sprint." Just do this: Day one: Pick one task from your actual to-do list today. Something you'd rather not do. Write a prompt using the smart intern structure above. See what comes back. Edit it. Use it. Day two: Do it again with something different. Notice what kind of output you had to fix a lot. That tells you where you need to give it more context. Day three: Find one more use case. Try a slightly more complex task. Maybe a short analysis or a document that requires structure. By the end of 72 hours, you'll have a sense of what it's good at for your specific job. You'll have a few prompt patterns that work. And you'll have made a dent in some work you'd normally hate doing. That's the whole thing. There's no phase two that requires a bootcamp. You learn by using it. The rest is refinement. For the full framework, including how to audit your own tasks, identify your combination moat, and build AI fluency without losing what makes you valuable, Dmitry Kargaev lays it all out in Don't Replace Me. What AI at work actually feels like once you're in the habit After a few weeks, it stops feeling like a special tool and starts feeling like, as Kargaev puts it, "just your keyboard." It's infrastructure. You reach for it the way you reach for a search engine, except you get a useful answer instead of ten ads and three Reddit threads. The people who end up ahead aren't the ones who learned the most about AI. They're the ones who got comfortable using it fast, while everyone else was still watching YouTube explainers about it. According to Goldman Sachs research on AI's labor market impact, the productivity gains from AI adoption are real, but they accrue to workers who actually adopt it, not workers who have opinions about it. The gap between "I know about AI" and "I use AI" is the gap that matters. Cross it. It takes one afternoon. And if you want to go deeper on the stuff that actually protects your career long-term, not just how to save time on emails but how to build the skills and positioning that genuinely make you harder to replace, there's more on that here. Frequently asked questions How do I start using AI at work if I've never used it before? Open ChatGPT or Claude (both free), pick one task you find tedious, and describe what you need using this structure: role, context, task, format. You don't need training. You need one real task to try it on. Most people get useful results in their first session. Do I need to learn prompt engineering to use AI effectively at work? No. Prompt engineering as a formal skill is largely obsolete with modern AI models. Clear, specific instructions work fine. Tell the AI what role to play, give it the relevant context, describe the task, and specify the output format. That's all you need. What are the best AI tools for work in 2025? For most professionals, ChatGPT and Claude cover everything. Both have free tiers. If your company uses Microsoft 365 or Google Workspace, there's AI built into those already. Avoid the impulse to collect tools. Pick two and get fast at them. Is it cheating to use AI at work? No more than it's cheating to use spell-check or a calculator. You're responsible for the output, which means you still need to review and edit what AI produces. If your company has a specific policy against it, follow the policy. Otherwise, you're just working smarter. More detail on the workplace politics of this is over here. What tasks should I use AI for first? Start with your most tedious, repetitive, low-creativity work: drafting routine emails, writing meeting summaries, creating first drafts of documents you'd rather not write, reformatting content, generating options for something you're stuck on. The goal is to get time back, not to immediately use AI for your most complex or judgment-heavy work. How much time can AI actually save me at work? It varies by role and how you use it. McKinsey's 2023 research on generative AI found productivity gains of 20-40% on specific knowledge work tasks. In practice: if AI cuts the time to produce a first draft from 90 minutes to 20, and you do that three times a week, you're reclaiming a few hours. That compounds fast. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Anxiety at Work: How to Stop Spiraling 65% of employees feel AI anxiety at work. Here's how to convert that fear into something useful without buying a $997 course or losing more sleep. ai anxietyai job fearwill ai replace my job ChatGPT Prompts at Work That Actually Work Real ChatGPT prompt templates for emails, reports, and presentations. The role-context-task-format framework that gets you output worth using. chatgpt at workpromptsAI tools AI Job Replacement Statistics: What the Data Says Goldman Sachs, WEF, McKinsey: what AI job replacement statistics actually say vs. what gets quoted in headlines. The real numbers, with context. AI job replacementAI statisticsfuture of work --- ## How to Use ChatGPT at Work (That Actually Works) URL: https://dontreplace.me/articles/how-to-use-chatgpt-at-work-prompts-that-help chatgpt at workprompt templatesAI at workChatGPT promptshow to use ChatGPT How to Use ChatGPT at Work (That Actually Works) Real prompt templates for emails, reports, and presentations. Learn how to use ChatGPT at work without courses or hype. 150 chars. March 25, 2026 Updated 12 May 2026 28% of employed US adults are already using ChatGPT at work. Most of them are getting mediocre results and wondering why. The answer is almost always the same: they're prompting it like a search engine. One sentence, vague request, disappointment. Then they conclude AI is overhyped and go back to doing everything manually. That's the worst possible outcome. Learning how to use ChatGPT at work doesn't require a course, a bootcamp, or a $997 "prompt engineering masterclass." It requires understanding one simple thing: how to give good instructions. That's it. The rest follows. Here's what actually works. Why your ChatGPT prompts aren't working The model isn't stupid. You're just not telling it enough. Think about it this way. If you hired a new intern on Monday and handed them a sticky note that said "write the client update," you'd get something generic and probably wrong. Not because the intern is useless. Because you gave them nothing to work with. ChatGPT works the same way. It'll fill in every gap you leave with its best guess. Sometimes that guess is fine. Usually it isn't. The model has no idea who you're writing for, what tone your company uses, what already happened with this client, or what you actually want the output to look like. Vague in, vague out. That's not a bug. That's cause and effect. The smart intern framework (role + context + task + format) Stop thinking of ChatGPT as a magic box. Start thinking of it as a smart intern who just started this week. The intern is capable. They're fast. They'll work through the night without complaining. But they need clear instructions. They don't know your business, your clients, your writing style, or your internal politics. You have to tell them. The framework is four parts: Role. Tell it who it is. "You are an experienced B2B copywriter" or "You are a senior HR manager." Context, Give it the situation. Who's involved? What's happened so far? What matters here? Task, Be specific about what you want. Not "write an email." Write a two-paragraph email that does X for Y. Format, Tell it what the output should look like. Bullet points? Three paragraphs? Under 150 words? That's it. Role, context, task, format. Every good prompt has all four. Most bad prompts have maybe one. Prompt templates for common work tasks These are real templates. Copy them. Adjust the brackets. Use them today. For emails Bad prompt: "Write an email to a client about a delayed project." Good prompt: You are a senior account manager at a B2B software company. A client has been waiting six weeks for a feature launch that's now delayed by another three weeks due to unexpected technical issues. They've already followed up twice. Write a professional but warm email to the client acknowledging the delay, taking responsibility without making excuses, offering a specific check-in call on [DATE], and keeping the tone reassuring rather than defensive. Keep it under 200 words. The good version tells it who you are, what happened, what tone to hit, and what length to stay under. The bad version gives it nothing. You'll see the difference immediately. For reports and summaries Bad prompt: "Summarize this document." Good prompt: You are a business analyst. I'm going to paste a 10-page quarterly performance report. Your job is to extract the three most important findings, note any red flags, and write an executive summary of no more than 250 words that a CEO could read in 90 seconds. Use plain language, no jargon, and end with one sentence that states the single most important action item. Then paste the document. For presentations Bad prompt: "Help me make a presentation about Q3 results." Good prompt: You are a management consultant helping a mid-size marketing agency. I need to present Q3 results to our leadership team. Key numbers: revenue up 12%, client churn up 4%, two major new logos signed. The mood in the room will be cautiously optimistic. Create an outline for a 15-minute presentation with 8-10 slides, each with a headline that tells the story (not just labels the data), and suggest one visual type per slide. Tone: direct, professional, no spin. Notice that last line. "No spin" is specific. Without it, you'll get corporate fluff. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → How to use ChatGPT at work without embarrassing yourself Here's the part most guides skip. Rule #15 in Don't Replace Me calls it the Double Tap: always verify before you ship. ChatGPT will lie to you. Not on purpose. It doesn't have purposes. But it will confidently produce statistics that don't exist, quotes that were never said, and facts that are just wrong. The model is trained to produce plausible text. Plausible isn't the same as accurate. Before you send anything generated by AI, run it through this quick check: Any statistics or data? Find the original source. Don't trust the citation it gives you, either. It invents those too. Any names, titles, or dates? Verify them. Especially if the document is going to someone external. Does the tone actually match your voice? Read it out loud. If it doesn't sound like you, fix it. Would you be comfortable if your boss saw the unedited version? If not, edit it. This takes two minutes. It's the difference between AI making you faster and AI making you look incompetent. How to use ChatGPT at work for harder tasks: negotiations, tricky conversations, and things you've been putting off Most people start with the easy stuff. Emails. Summaries. That's fine. But the real time savings kick in when you use it for the work you've been dreading. Think about the things that sit on your to-do list for weeks because you don't know how to start. The difficult feedback email to a colleague. The proposal for a budget you're not sure how to justify. The performance review you've been staring at for three days. These are exactly the tasks where ChatGPT earns its keep. Difficult feedback and sensitive emails Give it the full situation. What happened, what you need to say, what relationship you're trying to protect. Then ask it to draft something that's direct without being cold. You'll get a first version in ten seconds. It won't be perfect, but it'll be a starting point, and starting is usually the hardest part. Example prompt: You are a team lead at a 20-person marketing agency. One of your direct reports has been missing deadlines consistently for six weeks. They're talented but disorganised. I need to have a conversation with them this week and follow it up with a written summary email. Draft the email for me. Tone: direct, supportive, not punitive. The email should name the pattern, explain the impact, and propose a weekly check-in going forward. Under 200 words. You still decide whether to send it. You still decide what the check-in looks like. The judgment is yours. The first draft is the machine's. Proposals and documents you're stuck on Blank page paralysis is real. Research on workplace productivity consistently points to task-switching and starting costs as the biggest drains on knowledge worker time. Getting the shape of a document down in two minutes instead of two hours changes the whole day. Ask ChatGPT to build the skeleton. Then you fill in the bones. Ask it for an outline of a budget proposal for X, with the key arguments a skeptical CFO would want to see answered. Give it the numbers you have. Let it structure the argument. Then rewrite every section in your voice with your actual knowledge. That's not cheating. That's how every good writer uses editors. You're editing a draft, not transcribing one. What ChatGPT is genuinely good at (and what it isn't) This matters because people either expect too much or too little. ChatGPT is good at: First drafts of anything written (emails, reports, proposals, summaries) Reformatting content (turning bullet notes into prose, turning a long email into a short one) Brainstorming options when you're stuck Editing and improving text you've already written Explaining complex topics in plain language Generating structures and outlines Drafting difficult conversations you've been avoiding ChatGPT is not reliable for: Current events or recent data (its training has a cutoff) Anything that requires real-time information Making judgment calls that require context you haven't given it Reading your company's internal politics Knowing what your audience actually wants to hear Verifying facts, citations, or statistics The line is pretty clear. Use it for the written grunt work. Don't use it as a source of truth for facts you haven't verified. And check out what AI can and can't do if you want a fuller breakdown of where the model actually falls down. The tool stack you actually need Stop downloading AI apps. Seriously. Dee's exact quote: "You don't need anything but Claude or ChatGPT. Everything else is mostly wrappers." That's not contrarianism. That's the practical reality. Most AI tools you're seeing marketed right now are just ChatGPT or Claude with a different interface and a monthly fee. You need two, maybe three tools max: ToolWhat it's forCost ChatGPT (free or Plus)Writing, drafts, summaries, brainstormingFree / $20/month ClaudeLonger documents, more nuanced toneFree / $20/month PerplexityResearch with citations (for when you need current data)Free / $20/month That's it. Pick one or two and actually use them every day for 30 days. You'll get further than someone who has 15 tools and uses none of them well. If you want the broader guide on building confidence with these tools, how to use AI at work walks through the full approach, including the framework for figuring out which parts of your job to hand over first. The one habit that separates good ChatGPT users at work from everyone else It's iteration. Not the first prompt. The conversation. Most people write one prompt, read the output, decide AI doesn't understand them, and give up. The people who get good results treat it like a dialogue. The first output is a rough draft, not a final answer. Try this: Send your first prompt. Read the output. What's close but not right? Send a follow-up. "That's too formal. Make it sound more casual." Or "The second paragraph is good. Rewrite the first to match that tone." Keep going until it's close enough to edit into shape. You're not outsourcing your thinking. You're doing the thinking out loud, with a fast first-drafter who never gets tired. The judgment is still yours. The grunt work is theirs. OpenAI's own usage data shows that multi-turn conversations produce significantly better outputs than single-prompt queries. Which tracks. Every experienced user learns this within the first two weeks. The people who bounce off the tool early almost always gave up after one or two attempts without iterating. And if you're wondering whether any of this changes how transparent you need to be with your manager, that's a separate question worth reading: should you tell your boss you use AI covers the workplace politics without the hand-wringing. The goal isn't to get AI to do your job. The goal is to spend more of your time on the parts of your job that only you can do. That's the whole thing, honestly. Use the intern for the boring draft. Use your brain for the part that actually matters. Frequently asked questions How do I write a good ChatGPT prompt for work? Use the role + context + task + format structure. Tell it who it is, what the situation is, exactly what you want, and what the output should look like. The more specific you are, the better the output. Vague prompts produce vague results every time. Can I use ChatGPT to write work emails? Yes, and it's one of the best use cases. Give it the recipient, the relationship, what needs to be communicated, the tone, and the length. Then edit the result to sound like you. Don't send the raw output without reading it first. Is ChatGPT free to use for work? The free version of ChatGPT works for most basic tasks. ChatGPT Plus at $20/month gives you access to GPT-4o and faster responses. For most office workers, the free version is a fine starting point. Try it before you pay. Will ChatGPT make up facts in my work documents? Yes. It will. Confidently. Always verify any statistics, quotes, dates, or names before the document leaves your hands. ChatGPT generates plausible text, not verified facts. Treat it like a fast researcher who needs fact-checking, not a database. Do I need to learn prompt engineering to use ChatGPT well? No. Prompt engineering as a formal discipline is mostly hype at this point. You need to learn how to give clear instructions, which you already know how to do. Role, context, task, format. That's the whole skill. The AI skills non-technical people actually need article covers what's worth learning and what you can skip. How is ChatGPT different from Claude for work tasks? Both are good general-purpose tools. Claude tends to handle longer documents better and produces more nuanced prose. ChatGPT has a larger ecosystem of integrations. Most people pick one, use it for two weeks, and develop a preference. Start with whichever one you already have access to. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Skills for Non-Technical People (2026) The AI skills non-technical people actually need in 2026. No coding, no bootcamps. Four practical skills you probably already half-have. ai skillsprompt engineeringAI at work Should You Tell Your Boss You Use AI? 68% of workers hide their AI use from employers. Here's the honest answer on when to disclose, when to stay quiet, and how to handle it. AI at workworkplace AIAI disclosure AI Anxiety Is Real — Here's How to Stop Spiraling 65% of workers feel AI anxiety. Here's why it makes sense, what history tells us, and a practical first-72-hours plan to stop spiraling. ai anxietyai job fearhow to use ai at work --- ## Jobs AI Can't Replace: Human Skills That Hold Up URL: https://dontreplace.me/articles/jobs-ai-can-t-replace-human-skills-that-matter jobs AI can't replaceAI-proof careershuman skillsfuture of workAI job displacement Jobs AI Can't Replace: Human Skills That Hold Up Jobs AI can't replace aren't myths. Empathy, taste, physical judgment, trust: here's what humans still own and why it matters for your career. May 14, 2026 Updated 15 May 2026 The nurse who held your grandmother's hand when she died wasn't replaceable. Neither was the lawyer who read your boss's face during the deposition and knew exactly when to stop talking. Or the designer who looked at 40 AI-generated logos and said "none of these, they all feel like they were made by someone who's never been embarrassed before." Jobs AI can't replace aren't mythical. They're right in front of you, and they're yours if you stop pretending they don't count. The panic about AI and jobs is real, but it's pointed at the wrong thing. What's actually happening is that AI is eating the repetitive, the predictable, and the easily documented. What it's leaving behind, what it genuinely cannot do, is more valuable than most people give themselves credit for. Here's the honest version. What does it actually mean for a job to be "AI-proof"? Nothing is completely immune. If your job is 100% documentation, 100% pattern-matching, and 100% done over email, you've got a problem. But almost no job is that. Most jobs are a mix of things: some of it boring and automatable, some of it deeply human and not. The World Economic Forum's 2025 Future of Jobs Report makes this clear. It projects 85 million jobs displaced by automation and 97 million new ones created, and the roles growing fastest are the ones requiring human judgment, empathy, and physical dexterity. Not coding. Not "prompt engineering." Judgment. The stuff you're already doing. So when we say "jobs AI can't replace," we don't mean jobs that are 100% robot-proof forever. We mean jobs where the core value, the thing clients pay for, managers depend on, and customers remember, is something AI consistently fails at. And right now, that list is longer than the LinkedIn doomers want you to believe. Jobs AI can't replace: the categories that hold up Let's be specific. There are five categories where human workers consistently beat AI, not because of some sentimental attachment to humans, but because the tasks actually require something the technology can't do. Empathy-driven roles. Social work, nursing, therapy, counseling, hospice care, special education. The entire value proposition here is human presence. You can't automate grief. You can't automate the thing that happens when someone who has lost a child sits across from another person who actually sees them. Research from Harvard Medical School has documented the measurable clinical outcomes of human presence in care settings. An AI can generate a technically correct care plan. It cannot hold a hand. Physical and unpredictable environments. Electricians, plumbers, surgeons, carpenters, emergency responders. Robots are genuinely impressive in controlled factory settings. They are terrible in your leaking basement at 11pm. The physical world is messy, contextual, and full of exceptions. The plumber who shows up and says "this isn't what I expected, we need to change the plan" is exercising judgment AI cannot replicate. Creative direction and taste. Not "can AI make art" (it can, sort of). The question is whether AI can tell the difference between the good version and the bad version. It can't, not reliably. Someone still has to call that. Rule #8 in Don't Replace Me puts it plainly: "AI can generate ten thousand versions of the wrong thing. You still need someone who knows which one is worth keeping." Moral reasoning under pressure. Judges, ethics boards, doctors making triage decisions, executives deciding whether to blow the whistle. These aren't just hard problems. They're problems where accountability, lived experience, and the weight of consequences matter to the outcome. AI can surface options. It cannot own a decision. Trust-based relationships. The financial advisor your family has used for 20 years. The lawyer who knows your company's history. The account manager who's navigated three crises with your client. These aren't just relationships. They're institutional knowledge plus trust plus history, and they're extremely hard to replicate. Why taste is the most underrated human advantage Everyone talks about empathy and physical work. Fewer people talk about taste, probably because it sounds soft and hard to put on a resume. But taste is a professional superpower. It's the accumulated judgment of everything you've ever seen, read, experienced, and cringed at. It's what makes a senior editor immediately know that a headline is wrong before they can explain why. It's what makes a good architect know that the technically correct solution is going to feel wrong to live in. AI generates. It does not evaluate. It can produce a thousand options and has no idea which one is actually right for your audience, your brand, your moment, your client who said they wanted "modern" but actually meant "safe." You know that. You've worked with that client for two years. This is why creative professionals aren't as replaceable as everyone thought. The threat isn't "AI makes the thing." The threat is "AI makes the thing and you don't add judgment on top of it." The people who stay employed are the ones who sit in what the book calls the director's chair: letting the machine generate, and then actually deciding. Think about what taste looks like in non-obvious professions. A finance director reviews an AI-generated quarterly report and immediately flags that the tone is too confident given what the team knows about Q4. A senior HR manager looks at an AI-drafted job description and spots that the language will put off exactly the candidates they want. A product manager reads an AI summary of user research and notices it's technically accurate but missing the thing three different users kept circling back to in slightly different words. None of those catches are in the data. All of them required someone who knew what "right" looks like for that specific context. That's taste. And it compounds with experience in a way that AI can't replicate, because AI doesn't accumulate judgment the way a person does. If you want to think harder about where your own role falls on the automation risk spectrum, this breakdown of what jobs AI will actually replace is worth reading before you spiral. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → The "it knows everything, it understands nothing" problem Here's a thing AI genuinely cannot do: understand context that wasn't written down. Your company has a difficult stakeholder. Everyone who's been there more than six months knows it. Nobody wrote it in any document. It lives in institutional memory, in the raised eyebrow at a certain name in a meeting, in the way the team automatically reframes proposals before sending them to that person. AI has access to your documents. It doesn't have access to that. It can't read the room. It can't notice that the client seemed distracted today and maybe this isn't the moment to push. It can't feel the political temperature in a reorganization. A 2024 study from Anthropic examining Claude's reasoning capabilities found that large language models still struggle significantly with novel contextual reasoning, the kind that requires understanding implicit social dynamics rather than stated facts. This isn't a bug that gets patched. It's a fundamental gap between pattern-matching on text and actually understanding how humans operate. That gap is your job security. And it shows up in places people don't always think about. A consultant who's worked with a client for three years knows which recommendations will die in committee before they're even presented. A sales rep knows that a particular procurement manager always asks about implementation timelines because a previous vendor burned them, even though it's not in any CRM note. A project manager knows which team member goes quiet when they're overwhelmed rather than when they're fine. None of that lives in a database. All of it changes how you do your job. Physical reality is still a human domain Automation has made incredible progress in controlled environments: manufacturing lines, warehouse robotics, sorting systems. But the physical world outside of those controlled environments is still largely a human domain. The Bureau of Labor Statistics projects that roles like home health aides, nurse practitioners, physical therapists, and electricians are among the fastest-growing occupations through 2032. These roles share a common trait: they require real-time physical problem-solving in environments that are never exactly the same twice. An AI can diagnose from an image. It cannot thread a catheter through a nervous patient who just flinched. A robot can map a floor plan. It cannot figure out that the walls in this specific 1960s house aren't where they're supposed to be and adapt on the fly. The tactile, the spatial, the improvised, all of that is still deeply human. And before anyone says "but robots are getting better at physical tasks," yes they are. But the cost, the reliability in uncontrolled environments, and the regulatory and liability frameworks around deploying them for complex physical care work? Those barriers are significant, and they don't disappear by 2030. You've got runway. Consider what it actually takes to fully automate an electrician. You need a robot that can fit into a crawl space, identify which of three dozen unlabeled wires is live, make a judgment call about whether the existing panel can handle the load, explain to the homeowner why the job is going to cost more than estimated, and then deal with the fact that the crawl space is smaller than the blueprints said. That's before lunch. The gap between "AI is impressive" and "AI replaces the trades" is wide and full of liability, physics, and humans who flinch. More examples of jobs AI can't replace in practice It helps to get concrete. So here's a table of roles that consistently hold up, and specifically why. RoleWhy AI can't replace it TherapistPresence, genuine attunement, the therapeutic relationship itself Emergency room nurseReal-time physical assessment in chaotic, unpredictable conditions Trial lawyerReading the room, improvising under pressure, jury psychology Art directorTaste, knowing what's right for a specific brand voice and moment Elementary school teacherEmotional attunement to 25 different kids at once, behavioral reading ElectricianPhysical improvisation in uncontrolled environments Hospice workerHuman presence during dying, family support, grief Senior account managerRelationship history, institutional knowledge, trust SurgeonFine motor judgment in live situations, adaptive decision-making WhistleblowerMoral courage. AI literally cannot own consequences None of these are "low-skill" jobs. Most of them are underpaid relative to their actual complexity. And all of them require something that doesn't exist in training data. What happens when you combine human skills with AI tools Here's the uncomfortable nuance. The jobs that are safest aren't just the ones with irreplaceable human skills. They're the ones held by people who have irreplaceable human skills and also use AI well. A therapist who uses AI to handle their notes, scheduling, and intake paperwork has more time for actual sessions. A creative director who uses AI to generate rough concepts can evaluate 40 options in the time it used to take to generate four. A surgeon who's trained on AI-assisted imaging catches things earlier. The "taste moat" gets deeper when you can produce more to evaluate. The empathy advantage gets sharper when the administrative noise disappears. Your human edge compounds when you stop spending it on the parts of the job that were always mechanical. This is the thing most people miss in the "AI replaces jobs" conversation. It's not a binary. It's not you or the machine. It's what you can do when the machine handles your least interesting 40%. If you want a framework for actually starting that process, this guide on using AI at work is the practical starting point. The combination nobody can copy Your actual job security isn't one skill. It's a combination: your domain expertise, your specific relationships, your institutional knowledge, and now your AI fluency. That combination is yours. It's not on any training dataset. Think about it this way. An AI can write a legal brief. It cannot walk into your client's office with 12 years of their company's history in its head and say "I know this feels like a big decision but here's why I think we push, and I also know you have a board meeting Thursday so let's make sure you're comfortable before then." That sentence required domain expertise, relationship history, political awareness, and judgment about timing. Five things at once. All human. The people who get replaced are the ones who only do one thing, and that one thing is now done by a $20/month subscription. The people who stay are the ones who do several things at once, and who've started using the subscription to do it better. Your combination is your moat. Domain expertise plus relationships plus institutional knowledge plus AI fluency. None of those individually is enough. All four together is hard to replicate. The version of this that catches people off guard is how fast the combination becomes specific to you. Two years of using AI to assist with your particular client base, in your particular industry, with your particular institutional history, produces a workflow and a judgment that nobody else has. It's not transferable. It's not replicable. And it gets more valuable the longer you build it. If you want to build that combination deliberately, the career future-proofing guide lays out the framework. Which chapters in the book cover this? If this article resonated, the book goes much deeper on all of it: Chapter 7 (You Have Taste) / Rule #8 covers why creative judgment is irreplaceable and how to stop underselling yours. Chapter 8 (Nobody Sends a Robot to a Funeral) / Rule #9 is about empathy-driven work and why presence matters in ways that can't be automated. Chapter 9 (It Forgets Everything by Morning) / Rule #10 covers the gap between AI "knowing" and actually understanding, with specific implications for your role. Chapter 10 (Sit in the Director's Chair) / Rule #11 is the practical one: how to position yourself as the person making calls, not the person executing them. This page is one chapter's worth of insight. The book has 22 chapters, and they get into the specifics of your actual job in ways a single article can't. Get the full field guide here. Frequently asked questions What jobs can AI not replace? Jobs requiring empathy, physical dexterity in unpredictable environments, creative judgment, moral reasoning, and trust-based relationships are hardest for AI to replace. Roles like therapists, nurses, electricians, surgeons, creative directors, and senior advisors consistently show resilience. The WEF's 2025 Future of Jobs Report projects 97 million new roles emerging that prioritize exactly these human skills. Will AI ever be able to replace human empathy? Not in any meaningful professional sense, at least not within the timelines most people are worried about. AI can simulate empathetic language, but it can't provide presence, can't actually understand lived experience, and can't carry the weight of genuine accountability. Research from Harvard Medical School shows measurable clinical differences when human care providers are present versus absent, outcomes AI-generated responses can't replicate. Are creative jobs safe from AI? The honest answer is: it depends on which part of the creative job you're talking about. AI is good at generating volume. It's bad at taste, judgment, and knowing which version is right for a specific context. Creative professionals who act as directors rather than executors are in a strong position. The ones at risk are those doing pure production work with no evaluative role. What human skills does AI struggle with most? Reading implicit social dynamics, exercising taste and aesthetic judgment, building and maintaining trust over time, reasoning morally under pressure with real accountability, and operating in physical environments that are unpredictable. A 2024 Anthropic study found that even advanced language models struggle significantly with contextual reasoning that requires understanding unspoken social dynamics. Should I try to become an "AI expert" to stay employed? You don't need to become an expert. You need to become a user. There's a big difference. Starting with ChatGPT or Claude for 30 minutes a day on the tasks you find most tedious is enough to build real fluency over a few months. The goal isn't to understand how the technology works. It's to know what it can do well enough to hand it the right tasks and evaluate what comes back. This guide on the real AI skills non-technical people need breaks it down by profession. Is it possible that AI replaces my whole job, not just tasks? Possible in theory, rare in practice. Most jobs are bundles of tasks, and the tasks that AI replaces are usually not the core value the job provides. The jobs most at risk are narrow, highly repetitive, well-documented roles with minimal human interaction. If your job requires judgment, relationships, or physical presence in any significant way, you're looking at partial automation, not elimination. This breakdown of what jobs AI will actually replace gives a more specific read by industry. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Job Replacement Statistics: What the Data Says Goldman Sachs, WEF, McKinsey: what AI job replacement statistics actually say vs. what gets quoted in headlines. The real numbers, with context. AI job replacementAI statisticsfuture of work New Jobs AI Is Creating (With Evidence) AI is displacing 85M jobs but creating 97M new ones. Here's what the new jobs actually look like and why the pattern always repeats. new jobs created by AIAI job creationfuture of work AI Anxiety at Work: How to Stop Spiraling 65% of employees feel AI anxiety at work. Here's how to convert that fear into something useful without buying a $997 course or losing more sleep. ai anxietyai job fearwill ai replace my job --- ## New Jobs Created by AI: The Real Numbers URL: https://dontreplace.me/articles/new-jobs-ai-is-creating-that-don-t-have-names-yet new jobs created by AIAI job creationfuture of workAI careertechnology and employment New Jobs Created by AI: The Real Numbers The WEF says AI creates 97 million new jobs while displacing 85 million. Here's what those jobs actually are and how to position for them. March 25, 2026 Updated 12 May 2026 340,000 Americans worked as telephone switchboard operators at the peak of that profession. Then direct dialing arrived, and the job vanished almost completely within a generation. If you'd told those operators that their industry would eventually employ millions of people in call centers, telecom engineering, phone sales, and then an entire mobile economy, they'd have laughed in your face. They didn't have a word for "app developer" yet. The job didn't exist. That's the pattern with new jobs created by AI. We're standing at the same moment those switchboard operators stood at, except now the internet exists and we can watch the transition happen in real time, which mostly means watching LinkedIn posts about it instead of doing anything useful. The evidence is clear. The pattern is consistent. And the jobs being created right now, plus the ones coming that don't have names yet, are worth taking seriously. What history actually tells us about new technology and jobs Every major technology that people panicked about followed the same arc. Destruction of specific job titles. Creation of far more new ones. A gap in between where everyone was uncomfortable. The printing press killed professional scribes. There were maybe a few thousand people in Europe whose entire job was copying manuscripts by hand. The press eliminated that work within decades. What it created: publishers, editors, printers, journalists, librarians, booksellers, and eventually a literate public large enough to need teachers at scale. Net result: vastly more jobs than it destroyed. The automobile killed the horse-related economy. Farriers, stable hands, carriage makers, horse breeders supplying urban transport. The car replaced all of it. What it created: auto mechanics, gas station workers, road construction crews, traffic police, car salespeople, insurance adjusters, parking garage attendants, and eventually an entire suburban economy that wouldn't exist without the car. Again, more jobs than it destroyed. The internet killed travel agents, video store clerks, and classified ad departments. It created web designers, SEO specialists, social media managers, e-commerce logistics workers, content creators, and data analysts. The Bureau of Labor Statistics tracks these shifts, and the pattern holds every time. This is not optimistic spin. It's the actual record. The WEF number most people ignore The World Economic Forum's Future of Jobs report gets cited constantly for its estimate that AI and automation will displace 85 million jobs by 2025. That's the scary number. That's the one that shows up in headlines. The other number, the one that doesn't make for a good panic post, is 97 million. That's the WEF's estimate for new jobs created in the same period. Net positive: 12 million jobs. You can argue with the precision of those projections. Any economist would. But the direction of the estimate, net positive job creation, matches every historical precedent we have. The WEF isn't a cheerleading organization. They're tracking this because they're worried about it, and even they came out net positive. The Goldman Sachs research on AI and employment points the same direction. Disruption, yes. Displacement in certain categories, yes. Net destruction of employment at the macro level, no. This is the part worth sitting with if you're anxious about your job. The fear is real and valid. The specific outcome most people are fearing, mass permanent unemployment, doesn't match what happens historically or what the data currently suggests. New jobs created by AI: what's already here Some of these exist as job titles. Some exist as responsibilities that got added to existing roles. All of them are real, paid work that didn't exist before 2020. AI trainer / data annotator. Someone has to teach AI systems what good looks like. Medical AI needs doctors or radiologists reviewing outputs. Legal AI needs lawyers checking whether the generated contract clauses make sense. Creative AI needs humans deciding which outputs are actually good. Scale AI and similar companies employ large numbers of people doing exactly this, and it's just the visible surface. Prompt engineer. Yes, this one gets mocked. Yes, some of the "prompt engineering courses" selling for $997 are scams (the AI grifter economy is real). But inside companies deploying AI at scale, someone has to figure out how to reliably get useful outputs from these systems. That's a real job function whether or not you put "prompt engineer" on the business card. AI product manager. Building products on top of AI requires a different skill set than traditional software product management. Understanding what these models can and can't do, designing for their failure modes, explaining the outputs to non-technical stakeholders. This is a hybrid role that barely existed three years ago. AI ethics and compliance specialist. The EU AI Act is real. The regulatory environment around AI is expanding. Companies building and deploying AI systems need people who understand both the technology and the legal, ethical, and reputational implications. Legal departments are already hiring for this. Synthetic data specialist. AI models need training data. Getting high-quality real-world data is hard, expensive, and often legally complicated. Creating synthetic data that faithfully represents real-world patterns is an emerging specialty. That's just the first wave. The named roles. The unnamed ones are more interesting. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → New jobs created by AI that don't have names yet In 1993, "social media manager" didn't exist as a job title because social media didn't exist. By 2010, it was a real role. By 2015, there were dedicated degree programs. The entire career path appeared in less than 20 years. The same thing is happening now, just faster. Think about what happens when AI handles most of the pattern-matching, data processing, and routine communication tasks in a field. What's left? Judgment calls. Relationship management. Creative direction. Ethical review. Novel problem-solving. All the things that were previously squeezed out by the volume of routine work. That creates new specializations. Healthcare will need people who specialize in explaining AI-generated diagnostic information to patients. Not doctors, not nurses exactly, but a new kind of clinical communicator. Finance will need AI output reviewers who understand both the models and the regulatory requirements well enough to sign off on AI-generated analysis. Education will need people who design AI-augmented learning experiences. Nobody is advertising for those jobs yet. That doesn't mean they won't exist. It means we haven't named them. Dee writes about this directly in Don't Replace Me, the book this site is built around. The framing he uses is useful: you're not waiting to see if AI creates jobs. It already is. You're deciding whether to position yourself for the ones coming or spend the next few years worrying about the ones leaving. The fruit duty argument There's a specific pattern in how new technology frees up human time. Call it the fruit duty pattern. Imagine a small farming village where a significant portion of everyone's time goes to picking, sorting, and preserving fruit. Tedious, repetitive, necessary. Then someone invents a machine that handles most of it. What happens to that time? It doesn't disappear. It gets redirected. Some people maintain the machines. Some people use the freed time to do things the village previously didn't have bandwidth for, planning ahead, teaching children, building better systems. The total amount of work in the village might stay the same or increase. But the nature of the work changes. AI is doing this to knowledge work. Drafting routine emails, summarizing documents, generating first-pass analysis, creating initial designs. All of that was human "fruit duty." Time-consuming, not especially creative, necessary but not the point of the job. When AI handles it, where does the time go? Into the parts of the job that actually required a human to begin with. The judgment call about whether the strategy makes sense. The relationship with the client who's on the fence. The creative direction that makes the work worth doing. That reallocation creates new roles. Not always obviously. But it does. You can see this happening already in law firms where junior associates spend less time on document review and more time on actual legal reasoning. In marketing agencies where the junior team isn't writing 40 variations of the same email but is instead doing the strategic analysis that used to be too expensive to staff properly. What this means for your actual career If you're reading this because you're worried about your job specifically, the relevant question isn't "will AI destroy jobs in general." It won't. The question is whether your specific role is in the category being displaced or the category being created. The honest answer is that some specific jobs will go away. Not because of some abstract AI apocalypse but because the specific tasks those jobs were built around get automated. If your entire job is pattern-matching tasks that AI handles well, that's worth taking seriously. But "pattern-matching tasks" is a smaller category than it feels like from the inside. And most jobs are a mix, which means they transform rather than disappear. The question is whether you're the person who navigates that transformation or the one who waits for someone else to tell them what to do. The historical pattern, from scribes to switchboard operators to travel agents, isn't that the people doing the displaced work got destroyed. Most of them adapted. The ones who didn't adapt mostly chose not to. The machines were coming regardless. If you want to think more seriously about building actual resilience here rather than just absorbing reassurance, the career future-proofing guide is a good next step. The actual optimistic argument The optimistic case for AI and jobs isn't that nothing changes. Things are changing. The optimistic case is that the historical pattern of technology and employment is extremely consistent, and that pattern says net job creation, not net destruction. 85 million jobs displaced. 97 million new jobs created, per the WEF estimates. The switchboard operators gone, replaced by an entire mobile economy nobody predicted. What's coming in the AI wave that we can't name yet will be bigger than what's leaving. That's not wishful thinking. It's what every comparable transition has produced. The gap in between, right now, where we know the old jobs are changing and can't fully see the new ones yet, that's uncomfortable. But uncomfortable isn't the same as catastrophic. The printing press looked like the end of the world to the scribes. It was actually the beginning of the information age. Frequently asked questions Are new jobs being created because of AI? Yes, and this is already measurable. The World Economic Forum estimated that AI and automation will create 97 million new jobs while displacing 85 million, a net gain of 12 million. Current real-world examples include AI trainers, AI product managers, ethics and compliance specialists, and synthetic data specialists, with many more emerging roles not yet named. What kinds of jobs is AI creating right now? The most visible new roles include AI trainer (humans who evaluate and correct AI outputs), prompt engineer (people who design reliable AI workflows at scale), AI ethics specialist, and AI-focused product managers. Beyond these, almost every professional field is seeing hybrid roles emerge that combine domain expertise with AI fluency. Didn't AI take 85 million jobs according to the WEF? That's the displacement estimate, yes. The part that usually doesn't make the headline is the other WEF number: 97 million new jobs created in the same period, for a net positive of 12 million. The full AI job replacement statistics break this down in more detail. Has any technology ever created more jobs than it destroyed? Every major technology in the historical record. The printing press eliminated scribes and created publishing. The telephone displaced messenger services and created telecom and call centers. The internet killed travel agents and created the entire web economy. The mobile phone created the app economy. The pattern is consistent enough that it should be the default assumption. How long does it take for new jobs to replace displaced ones? The transition period is uncomfortable and varies by industry, typically 10 to 30 years for a full labor market shift following a major technology introduction. The switchboard operator job peaked around 1970 and was largely gone by the mid-1980s, but the jobs built on top of phone networks far exceeded that employment by the same period. What jobs will AI create that don't exist yet? Nobody can name them precisely, because they don't have names yet. But based on historical patterns, they'll cluster around maintaining and directing AI systems, roles that use the time freed by AI automation for higher-value human work, and entirely new categories built on top of AI infrastructure the way social media manager was built on top of the internet. Positioning for these roles is the core of any serious AI career strategy. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Job Replacement Statistics: What the Data Says Goldman Sachs, WEF, McKinsey: what AI job replacement statistics actually say vs. what the headlines tell you. Real numbers, real context. ai job replacement statisticsai job lossfuture of work Jobs AI Can't Replace (And Why) Which jobs are genuinely hard for AI to replace? From care work to skilled trades, here's what the data says about human skills that still matter. jobs AI can't replaceAI-proof careershuman skills What Jobs Will AI Replace? A Role-by-Role Guide Finance, legal, marketing, tech, creative: an honest breakdown of which jobs AI will replace, which survive, and what to do about it. AI job replacementjobs AI will replacecareer risk --- ## New Jobs AI Is Creating (With Evidence) URL: https://dontreplace.me/articles/new-jobs-ai-is-creating-with-evidence new jobs created by AIAI job creationfuture of workAI careersjobs AI is creating New Jobs AI Is Creating (With Evidence) AI is displacing 85M jobs but creating 97M new ones. Here's what the new jobs actually look like and why the pattern always repeats. May 15, 2026 Updated 15 May 2026 340,000 Americans worked as telephone switchboard operators at the job's peak. Then direct dialing arrived, and the job vanished. Not slowly. Gone. Here's what didn't vanish: all the work that phone networks made possible. Telemarketing. Customer support. Telecommunications engineering. And eventually, everything built on voice communication that became the internet's backbone. The new jobs created by AI follow the same pattern, and the data backs it up. The World Economic Forum's Future of Jobs Report puts a number on it: 85 million jobs displaced by AI and automation by 2025, but 97 million new roles created in the same window. Net positive: 12 million jobs. Nobody's leading with that headline. This isn't cheerleading. It's pattern recognition. The same thing happened with the printing press, the automobile, and the internet. The technology always wins. The people who learn the new rules keep working. The ones who don't, don't. The historical pattern nobody wants to admit Every major technology in history has pulled off the same trick. It kills a job category, then quietly creates three more. The printing press put professional scribes out of work. A scribe who copied manuscripts by hand wasn't going to survive when a press could do it in an afternoon. But the press also created publishers, editors, journalists, proofreaders, typesetters, and eventually librarians. It didn't shrink the world of written work. It exploded it. The automobile killed horse-related trades: blacksmiths, farriers, stable hands, carriage makers. In exchange, it created mechanics, gas station workers, highway engineers, traffic police, auto insurance adjusters, and eventually the entire suburban economy that reshaped where and how people lived and worked. The internet killed travel agents, video rental clerks, and print classifieds. It created web developers, SEO specialists, social media managers, UX designers, e-commerce logistics coordinators, and cloud infrastructure engineers. Most of those job titles didn't exist in 1995. By 2010 they were normal careers with career ladders and annual reviews. Mobile added another layer. The entire app economy. Gig work. Influencer marketing. The job "app developer" barely existed in 2006. By 2015 it was a lucrative career path with its own bootcamps. The pattern is not subtle. Destruction is visible and immediate. Creation is invisible until it isn't. New jobs AI is already creating right now Some of these roles have names. Some don't yet. Prompt engineer was a joke title two years ago. Now it's a real position at serious companies, with salaries between $100,000 and $300,000 at the top end, according to data from job boards like LinkedIn and Indeed. Whether that specific title survives is debatable. That the underlying skill persists is not. AI trainers are the humans who review AI outputs, flag errors, and teach the model what good looks like. It's tedious, granular work that requires genuine subject-matter expertise. A doctor reviewing medical AI outputs. A lawyer checking contract summaries. A writer grading creative quality. These jobs are being created at scale at companies like Scale AI and Appen. AI ethicists and policy specialists are now employed at every major tech company and in government agencies. The EU's AI Act alone created demand for compliance roles that didn't exist 18 months ago. That's a category, not a niche. Then there are the roles that don't have clean titles yet. The person who manages a company's relationship with its AI vendors. The specialist who audits AI outputs for a law firm before anything goes to a client. The human who handles the customer escalations that the AI chatbot couldn't resolve. These are real jobs happening right now, at real companies, carried out by people who didn't know that was their job description when they finished school. Why "new jobs created by AI" always sounds unbelievable Here's the problem with optimism about technology and employment. The jobs that get destroyed are concrete and visible. You can name them, count them, interview the people who held them. The jobs that get created are abstract and invisible until they already exist. In 1994, if you told a laid-off travel agent that millions of jobs would be created around something called "search engine optimization," they'd have stared at you. If you told a video store clerk that entire studios would be built to create content exclusively for an internet streaming service, they'd have needed the term "streaming service" explained first. We're in that exact same moment right now with AI. The new roles are forming in real time, and most of them don't have consensus titles yet. That makes them easy to dismiss. It also makes them easy to miss entirely. Dmitry Kargaev addresses this directly in Don't Replace Me, calling it Rule #21: New Tech = New Jobs. Always. The historical record is consistent enough that you can treat it as close to a rule as economics offers. Not because technology is kind, but because human demand for things doesn't disappear when efficiency improves. It expands. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What the WEF numbers actually mean for your career The 85 million displaced, 97 million created figure from the WEF isn't uniformly distributed, and pretending it is would be dishonest. Some industries face more displacement than others. If you want the realistic breakdown by sector, the ai job replacement statistics page goes through that in more detail. The more useful question is: where do the 97 million new roles cluster? According to the WEF, the growth is concentrated in: care economy roles (human-facing work that's hard to automate), green economy jobs (AI is being deployed in climate tech at scale), data and technology roles (obviously), and what the report calls "human-machine interface" roles: the jobs that exist precisely because AI exists and needs managing, auditing, directing, and interpreting. That last category is the interesting one, because it's the least predictable and the fastest growing. Every AI deployment at scale creates overhead. Legal overhead, compliance overhead, quality-control overhead, customer-trust overhead. Humans fill that overhead. The people best positioned for those roles aren't necessarily technical. They're domain experts who understand their field well enough to supervise AI in it. An accountant who understands what the AI got wrong. A nurse who can tell when the diagnostic model made an error. A lawyer who can spot when the contract summary missed a clause. If you're worried about where you fit into this, the how to future-proof your career against AI breakdown is the practical version of this. The "free from fruit duty" argument There's a version of this that's less about job titles and more about what work actually feels like. Imagine you spend 40% of your working week on tasks you find tedious. Formatting reports. Pulling data for presentations. Drafting first versions of documents you'll rewrite anyway. Scheduling and rescheduling. Researching things you already half-know. AI does a lot of that. Not perfectly, not always, but well enough to save you hours a week. That time has to go somewhere. Historically, it goes toward the work that actually required you in the first place. The judgment calls. The relationship maintenance. The creative problem-solving that nobody can automate because it depends on context that lives in your head and your history with the organization. This is what Kargaev calls being "freed from fruit duty" in the book: the same way agricultural machinery freed humans from manually picking every individual piece of fruit, AI is freeing office workers from the manual equivalent. The work doesn't disappear. It gets redistributed toward humans doing things only humans can do. Whether that's utopian or annoying depends on your tolerance for the tedious parts of your job. But the mechanism is real. McKinsey's research on automation consistently shows that even in roles facing significant automation, the displaced tasks represent a fraction of the overall job, not the whole thing. What kinds of roles will exist in 10 years Naming specific jobs is a fool's game, and anyone selling you a list of "the 20 AI jobs of the future" is guessing. But the categories are readable. AI supervision roles will exist in every industry. Someone has to be responsible when AI outputs go wrong, and "the AI did it" is already not a legal defense in most jurisdictions. That liability creates human roles. Human-facing work will grow relative to automated work. When every company's customer-facing AI sounds identical, the companies that offer a real human as a differentiator win a customer segment. That's not idealism, that's a market opportunity. Domain experts with AI fluency will be the most valuable category in most professional fields. Not the person who understands AI best. The person who understands their field best AND can work with AI tools without friction. The gap between those people and domain experts who refuse to adapt will get wide fast. Creative and taste roles won't disappear. AI generates. It doesn't discriminate well. Knowing what's good, what fits the audience, what will land, what to throw out: that's still human territory. If you want a longer argument for why, the breakdown of jobs AI can't replace makes the case. The roles that don't have names yet are the ones that emerge when two things collide: a new capability and a real-world problem nobody fully anticipated. Every technology produces these. AI is producing them faster than any previous technology. Fast enough that some of them are showing up on job boards before the industry has agreed on what to call them. The jobs AI is creating don't look like anything that existed before. Neither did "webmaster" in 1994, and that turned into a $100 billion industry. Frequently asked questions How many new jobs is AI creating? The World Economic Forum's Future of Jobs Report estimates AI and automation will create 97 million new roles by 2025, while displacing 85 million, for a net gain of 12 million jobs. The distribution isn't even across all industries, but the aggregate number is positive. You can see the full breakdown in the AI job replacement statistics breakdown. What new jobs is AI already creating right now? Current AI-created roles include prompt engineers, AI trainers (human reviewers who grade AI outputs), AI compliance and ethics specialists, and AI-vendor relationship managers. Less formally, most professional roles are expanding to include AI supervision tasks that didn't exist three years ago. These often don't have clean job titles yet. Will AI create more jobs than it destroys? Historical evidence and current WEF data both point to yes, but with a caveat: the timeline doesn't line up neatly. Destruction happens faster than creation, and the new jobs aren't in the same place as the old ones. The net number is positive, but the transition involves real disruption for real people. If you're concerned about your specific situation, the will AI replace my job article gives a more personal framing. What jobs will AI create that don't exist yet? Nobody knows the full list, and anyone who claims otherwise is guessing. The identifiable categories are: AI supervision and audit roles, human-machine interface specialists, domain experts who manage AI in their field, and expanded human-facing roles in industries where automated alternatives feel impersonal. The switchboard operator's job disappeared, but the telephone created more jobs than anyone anticipated. Do I need to be technical to get one of these new AI jobs? No. The most in-demand category is domain experts who understand their field and can supervise or audit AI in it. That's an accountant, a nurse, a lawyer, a teacher. Technical AI skills help but aren't the entry requirement for most of these roles. The AI skills you actually need article covers what "AI fluency" looks like without a computer science degree. Is optimism about AI and jobs naive? It's evidence-based, which is different from naive. Every major technology shift in recorded history has produced more jobs than it eliminated. That doesn't make the transition painless or the disruption evenly distributed. But dismissing the creation side of the equation because the destruction is more visible is just bad pattern recognition. The printing press didn't shrink the world of written work. Neither will this. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Job Replacement Statistics: What the Data Says Goldman Sachs, WEF, McKinsey: what AI job replacement statistics actually say vs. what gets quoted in headlines. The real numbers, with context. AI job replacementAI statisticsfuture of work Jobs AI Can't Replace: Human Skills That Hold Up Jobs AI can't replace aren't myths. Empathy, taste, physical judgment, trust: here's what humans still own and why it matters for your career. jobs AI can't replaceAI-proof careershuman skills AI Anxiety at Work: How to Stop Spiraling 65% of employees feel AI anxiety at work. Here's how to convert that fear into something useful without buying a $997 course or losing more sleep. ai anxietyai job fearwill ai replace my job --- ## Should I Tell My Boss I Use AI? URL: https://dontreplace.me/articles/should-you-tell-your-boss-you-use-ai AI at workworkplace AIChatGPT at workAI disclosureAI job anxiety Should I Tell My Boss I Use AI? 68% of workers hide their AI use from employers. Here's the honest answer on when to disclose, when to stay quiet, and how to handle it either way. May 17, 2026 Updated May 17, 2026 Sixty-eight percent of workers who use AI at work don't tell their boss. So should I tell my boss I use AI? That question sounds simple until you actually sit with it. Then it gets complicated fast. You found something that works. You're not sure if it's allowed. You don't want to look lazy, or like you're cheating, or like you got your job under false pretenses. So you just... don't mention it. You do the work. You send the email. You move on. Here's the framework that actually helps. The three types of AI users at work Before you decide whether to say anything, it helps to know which category you're currently in. Most people fall into one of three. The Hider keeps AI use completely off the radar. Deletes browser history, pastes into Word docs instead of leaving ChatGPT tabs open, never mentions it. The fear is getting "caught," which implies they already feel like they're doing something wrong. This is the most common posture, and it creates a low-grade anxiety that doesn't go away. The Evangelist won't stop talking about it. Every meeting becomes an AI briefing. Every workflow becomes a prompt engineering opportunity. They've replaced their personality with a subscription to every AI newsletter. Their coworkers have started avoiding them. The Quiet User does the work, shares the results, and doesn't make it anyone else's problem. They use AI for the boring parts, they sanity-check the output, and they show up with better work and fewer complaints. Nobody asks how the sausage is made. The Quiet User wins. Almost always. Should you tell your boss you use AI? The honest answer The honest answer is: probably yes, eventually, in some form. But the timing, framing, and level of detail matter more than the disclosure itself. Here's what doesn't work: either extreme. Hiding it completely creates risk. Announcing it constantly creates resentment. The goal is to land somewhere in the middle, where your boss knows you're using the tools without feeling like you've been deceiving them, and without getting an unsolicited lecture about "responsible AI use" every time you submit a deliverable. According to a 2023 Fishbowl/SHRM survey, 68% of professionals using AI at work hadn't told their manager. Around 28% of employed US adults were already using ChatGPT for work tasks. Those numbers have gone up since then, not down. So if you're sitting on this question, you're in very good company. The real question isn't whether to tell them. It's how to tell them in a way that reads as "I'm good at my job and I use every available tool" rather than "please don't fire me, I wasn't sure if this was allowed." What your company culture is actually telling you Most companies haven't figured out their AI policy yet. Which means the absence of a policy isn't the same as a ban. But it also means you can't always just look it up and check. A few things worth reading before you decide what to say: Your industry's compliance norms. Legal, healthcare, and finance have real data privacy concerns. If you're pasting client information into a public-facing AI tool, that's a problem regardless of what your boss thinks. Know your regulatory environment before you know your disclosure strategy. Your boss's personal tech comfort level. Some managers have already integrated AI into their own workflow. Others still print emails. If your boss is in the second camp, a detailed explanation of how you use Claude to synthesize meeting notes is going to confuse and unsettle them. Adjust accordingly. Whether anyone at your level is already talking about it openly. If colleagues are mentioning AI use in stand-ups and nobody's getting fired, you have social proof. If you'd be the first, that's a different calculation. The actual language in your employment contract or company policy. Some organizations have issued AI use policies in the last year. Some of those policies ban certain tools. Some require disclosure. Worth checking before you assume it's all fine. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → The quiet weapon approach: results, not methods There's a concept in Don't Replace Me called the quiet weapon. The idea is that AI is most useful when it makes your work better without making your AI use the story. You're not trying to hide anything unethical. You're just not narrating your tool usage any more than you'd explain which keyboard shortcut you used to format a spreadsheet. Most of the time, your boss doesn't need to know your methodology. They need to see the output. If the output is good, questions about how you got there are mostly academic. This changes if you're in a role where the process is auditable. If you're a lawyer, a doctor, a researcher, or anyone whose professional judgment is explicitly what the client is paying for, the provenance of your work matters. That's a different conversation, and it should probably happen with your compliance team before it happens with your boss. For most office workers in most roles, the practical answer is this: use AI for the tasks it genuinely helps with, check the output before it leaves your hands, and let the quality of your work do the explaining. If someone asks, tell them. If nobody asks, you don't owe anyone a press release. How to actually tell your boss you use AI (if and when you do) If you decide to bring it up, or if you get asked directly, here's how to not make it weird. Don't confess. You didn't do anything wrong. The framing matters. "I've been using AI to help draft first versions of client reports" lands differently than "I have to tell you something." Don't oversell it. Your boss doesn't need a presentation on how AI is going to transform the business. They need to know that you have a new tool and that it's producing good work. That's it. Do be specific. "I use ChatGPT to help me structure my weekly updates before I edit them" is clear and concrete. "I've been exploring AI solutions to optimize our workflow processes" sounds like you're hiding something behind jargon. Do be honest about what you check. Bosses and clients care about accuracy and reliability. If you explain that you use AI as a starting point and then review and edit everything yourself, you're not admitting to laziness. You're describing a sensible process. If your company announces a formal AI policy and asks employees to disclose, disclose. Don't treat it like a loyalty test you need to game. If you're worried about AI anxiety at work making this feel bigger than it is, that's worth separating out. Sometimes the fear of telling your boss is less about AI and more about general workplace anxiety. Those are related but not the same problem. When hiding it actually is a problem There are situations where staying quiet is the wrong call. If your company has an explicit policy requiring disclosure, and you're ignoring it, that's a compliance issue, not a judgment call. If you're in a client-facing role and the client's data is going into an AI system that the client hasn't consented to, that's a legal and ethical problem. It doesn't matter how good the output is. If AI is producing work that you're presenting as original creative or analytical thought in a context where that distinction matters (academic research, expert testimony, original creative work under contract), the line gets complicated fast. Know where your industry draws it. And if your entire job output has quietly become AI-generated with almost no human input or review, and you're billing or being paid for expert judgment, you've crossed from "using a tool" into "misrepresenting your work." That's not a disclosure problem. That's a bigger problem. For the vast majority of people using AI to draft faster, summarize better, and think through problems more efficiently? None of that applies. You're not doing anything that requires confession. You're doing your job with better tools. Should I tell my boss I use AI? The decision matrix Here's a quick way to think through your specific situation: SituationWhat to do No company AI policy existsUse it, share results, mention it casually if asked Company has a disclosure policyDisclose as required. That's it. Client data is involvedCheck data privacy rules before using AI at all Boss is tech-forward and already mentions AIBring it up naturally. They'll probably be interested. Boss is tech-averse and won't get itLet the results speak. Tell them if directly asked. You're in a regulated industryConsult compliance before anything else You're already getting asked "how did you do this so fast?"That's an invitation. Tell them. The most important column is the last one. If your work has gotten noticeably better or faster, people will notice. When they ask, being able to say "I've been using AI to handle the first draft and then editing it heavily" is a lot better than getting flustered and changing the subject. If you want practical guidance on what to actually use and how to start, the starter guide to using AI at work covers the specifics without the theory spiral. And if you want to know which prompts actually do something useful in a real job context, this breakdown of ChatGPT at work is more concrete than most. The short version of all of this: don't hide it indefinitely, don't broadcast it constantly, and make sure the work is actually good before it leaves your hands. That combination covers most situations. Don't be weird about it. That's Rule #16 for a reason. Frequently asked questions Should I tell my boss I use AI, or just keep quiet? For most people in most roles, quiet disclosure on a need-to-know basis is the right move. You don't have to announce it, but you shouldn't actively lie if asked. A 2023 Fishbowl survey found 68% of AI-using employees hadn't told their manager, and widespread firings haven't followed. Check if your company has a formal policy, and let that be your floor. Is it cheating to use AI at work without telling your boss? No, using AI at work isn't cheating any more than using Google or Grammarly is cheating. The exception is if your company has an explicit policy requiring disclosure, or if you're in a regulated field where the source of your work matters legally or ethically. For most office workers doing most tasks, using AI tools is just using the tools available to you. What happens if my boss finds out I've been using AI? In most cases, not much. A 2023 Fishbowl survey found that 68% of AI-using employees hadn't told their manager, and widespread mass firings haven't followed. The more likely outcome is a conversation about your company's AI policy, if one exists. The bigger risk is if you've been pasting client or confidential data into public AI tools without checking your data privacy obligations. Should I ask my company if AI use is allowed before I start? It depends. If you're in a large, compliance-heavy organization in finance, law, or healthcare, checking the policy before you start is smart. If you're in a smaller company or a less regulated field and there's no policy, asking unprompted might create scrutiny that didn't exist before. Use common sense, check any existing written policies, and don't input confidential client data without knowing where it goes. What if my boss asks directly if I used AI? Tell the truth. Say yes, explain briefly how you used it (as a drafting tool, for summarization, for research), and emphasize that you reviewed and edited everything. Lying when directly asked is a much bigger problem than the disclosure itself. Most managers who ask are curious, not trying to build a case against you. What's the difference between using AI and letting AI do your job? Using AI means you're using it as a tool in your workflow. You're reviewing output, catching errors, adding judgment, editing for accuracy and tone. Letting AI do your job means you're passing through unreviewed outputs and billing your expertise when you haven't actually applied it. The first is fine. The second is a professional integrity issue that has nothing to do with AI specifically. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Customer Service Prompts: 10 Templates That Work 10 copy-paste AI customer service prompts for ChatGPT and Claude. Draft replies, de-escalate complaints, and summarize tickets without sounding like a robot. AI promptscustomer serviceChatGPT at work AI Decision-Making Prompts: 10 Templates That Work 10 copy-paste AI decision-making prompts for ChatGPT and Claude. Structure messy decisions, run pre-mortems, and draft memos without outsourcing your judgment. ai decision makingdecision promptsai at work AI Prompts for Work: 12 Templates You Can Use Today 12 copy-paste AI prompt templates for emails, reports, and decisions. Real examples for normal people, with what to check before you send anything. ai prompts for workchatgpt promptsworkplace ai --- ## The $997 AI Masterclass Is a Scam URL: https://dontreplace.me/articles/the-997-ai-masterclass-is-a-scam AI griftersAI coursesprompt engineeringAI scamsAI learning The $997 AI Masterclass Is a Scam The AI masterclass scam is a $5B industry built on fear. Here's how to spot the grifters and actually learn AI for free. May 08, 2026 Updated 12 May 2026 The AI course market hit $5 billion in 2023. A significant chunk of that was people paying to learn how to use software that costs $20 a month and has a free tier. Let that sit for a second. The AI masterclass scam is not a fringe thing. It's a $5 billion industry built on a very simple insight: scared people buy things. And right now, a lot of professionals are scared. So someone built a funnel around that fear, slapped a $997 price tag on it, and started posting on LinkedIn about the robots coming for your job. This is about how to spot those grifters, ignore their noise, and actually learn to use AI without paying someone four figures to paste prompts into a Google Doc. The AI gold rush has its own pickaxe sellers In every gold rush, the people who get rich are the ones selling pickaxes. Not digging. Selling. The '49ers mostly went home broke. The people selling boots and shovels did fine. AI is the same. There's real money being made right now, and most of it isn't from AI itself. It's from selling the idea of AI to people who are scared they're falling behind. The formula is simple: find a large audience of anxious professionals, tell them the robots are coming, then offer to save them for $997 (or $497 if they act in the next 20 minutes). Sprinkle in some screenshots of ChatGPT conversations as "proof." Done. Dee writes about this exact dynamic in Chapter 2 of Don't Replace Me, which he calls "The Pickaxe Salesmen." The observation that opens this site still holds: "Every third post says AI is coming for your job. Usually right before someone sells you a course. That panic is profitable for them. It is not useful for you." The grifter taxonomy: five types you'll recognize immediately Not everyone selling AI content is a fraud. But the grifters have tells. Here's how to sort them. The Course Seller Charges between $297 and $2,997 to teach you to use ChatGPT. The curriculum is: open ChatGPT, type a prompt, look at the output. Sometimes there's a module on "prompt engineering." There are worksheets. There's a Facebook group full of people who also paid too much. The teacher's most impressive credential is that they took a similar course six months ago. The tell: their course teaches you to use a free tool. The tool has documentation. The documentation is free. The LinkedIn Prophet Posts every single day about "the future of work." Never shows their actual work. The posts are structured like parables: short line, short line, dramatic pause, short line, broader point about humanity. Everything is a "trend to watch." Nothing is a thing they built or tested themselves. The tell: high engagement, zero specifics. Ask them what AI tool they used last Tuesday for a real task. Watch them pivot to philosophy. The Conference Speaker Keynotes for $15,000 about AI disruption. Has never trained a model, fine-tuned anything, or shipped a product that uses AI. Their slide deck is Getty Images of robots shaking hands with humans. They quote the McKinsey report. They say "paradigm" without irony. The tell: their "AI expertise" started about four months after ChatGPT launched. Before that, they were speaking about blockchain. Or NFTs. Or "the gig economy." Same slides, new acronym. The Prompt Engineering Guru This one is specific and almost sad. They built an entire personal brand around writing better prompts. They sell template packs. "The 47 Best Prompts for Marketing Professionals." Then GPT-4o shipped, and the model got better at understanding bad prompts. Then Anthropic built prompt generation into Claude's interface. The prompts that "worked" in 2023 are obsolete. The tell: the job title "Prompt Engineer" has basically evaporated. Indeed data from 2024 showed prompt engineering job listings fell sharply after peaking in early 2023. The guru is still selling the course. The Actual Builder Uses AI in their real job. Occasionally shares what worked and what didn't, usually in a blog post or a thread that feels a little rough around the edges because they wrote it between actual tasks. Doesn't have a course. Might have a newsletter they update when they have something worth saying. Their results are described in specifics, not screenshots of ChatGPT saying "Great question!" The tell: you can find evidence of their actual work. They talk about failure. They update their thinking when something changes. What the red flags actually look like You don't need a taxonomy if you know the red flags. Here's a quick checklist. They sell fear first. The pitch starts with some version of "AI is replacing people like you right now." Then the course appears. Fear that leads directly to a purchase button is a sales funnel, not a warning. Their results are screenshots. Showing ChatGPT output is not a result. Anyone can get ChatGPT to write something. The question is whether it was useful, accurate, and worth the time it took. Screenshots prove nothing except that the tool works, which is not in dispute. They never show their process failing. Real AI use is messy. You get hallucinations. You have to re-prompt. The output needs editing. Anyone who only shows clean wins either cherry-picks aggressively or has never used the tool on a real problem. They have a pricing ladder. The $97 intro course leads to the $497 advanced course leads to the $2,997 mastermind. Each tier promises the "real" secrets. This is the structure of a cult, not an education. They invented a framework with a name. "The CRAFT Prompt Method." "The AI Success Flywheel." "The 5-Step Neural Workflow." Real tools don't need proprietary names. ChatGPT is just ChatGPT. It doesn't need a methodology to be useful. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Why the AI masterclass scam works so well on smart people Here's the frustrating part. The people buying these courses are not dumb. They're often experienced professionals who are genuinely good at their jobs. That's exactly why the pitch works on them. Smart people got where they are by investing in skills. Courses, certifications, degrees. The idea that learning something new requires structured instruction is not unreasonable. It's usually correct. So when someone packages "AI fluency" as a learnable skill and sells it as a course, it maps neatly onto a pattern that has worked before. The problem is that AI tools aren't like most professional skills. You don't get better at them by studying curriculum. You get better by using them on your actual work, watching them fail, and adjusting. There's no body of knowledge to absorb. There's just practice. The grifters know this. The good ones even admit it, somewhere in the fine print, while still charging you for the course. Because what they're actually selling isn't knowledge. It's the feeling of doing something about the anxiety. That feeling is worth $997 to a lot of people. For a day or two, anyway. The WEF's 2023 Future of Jobs report found that 60% of workers will require significant reskilling by 2027. That's a real finding, and it's legitimately concerning. But "significant reskilling" does not mean buying a ChatGPT course. It means adapting how you work, which you can do by using the tools, not studying them. What actually works instead (and costs almost nothing) Here's the thing. Learning to use AI effectively does not require a course. It requires using AI. Thirty minutes with Claude or ChatGPT on an actual work task will teach you more than any $997 curriculum. Not because the tools are magic, but because the only way to get good at prompting is to do it badly a few times and see what happens. The free resources that beat any paid course: Anthropic's documentation at docs. anthropic.com is written by people who built Claude. It's thorough, it's accurate, and it costs nothing. Same with OpenAI's documentation for ChatGPT. YouTube is genuinely excellent. Not the "react to AI news" channels. The ones where someone screen-shares their actual workflow. Search specifically for "[tool name] + [your job] + workflow" and filter for recent. Ignore anything about "prompt engineering" from before 2024. Your colleagues who actually use the tools. Ask around. Someone in your office has figured out a few things that work for your specific type of work. Buy them coffee. That's a $7 investment and it beats any online curriculum, because they already know your job, your context, and the dumb constraints of your specific industry. If you want a structured framework without the upsell, the practical guide on how to use AI at work covers the core of it for free: start with the tasks you hate, see if AI speeds them up, iterate from there. No worksheets required. The "learn AI" advice is mostly wrong "You need to learn AI" sounds responsible. It's actually not very useful advice. You don't learn a hammer. You use it to hit nails. AI is the same. The "learning" that matters is figuring out which tasks in your actual job the tool helps with, and building a lightweight habit around using it for those tasks. The grifters have turned "learn AI" into a category of product to sell you. A course. A certification. A bootcamp. A mastermind. None of it helps as much as opening Claude and trying to do something you were going to do anyway. The anxiety driving the course purchases is real. Surveys from Pew Research show most American workers are more worried about AI's impact on jobs than they are excited about it. That worry is legitimate. But the solution isn't spending money on a course. It's using the tools at work until they stop feeling foreign. If you're worried about your skills being stale, the breakdown of AI skills non-technical people actually need is worth reading. The short version: you don't need to code, you don't need prompt engineering certification, you need to use the tools regularly on real problems. One more thing about the $997 price point That number isn't arbitrary. It's designed to clear a psychological threshold. Below $1,000 feels like a "reasonable investment in yourself" rather than a major purchase. It's below most corporate approval thresholds, so you might pay out of pocket without thinking too hard. It's enough to make the seller real money at scale, but not so much that you feel obligated to do serious due diligence before clicking buy. The grifter economy runs on a very specific anxiety, which is: "What if everyone else learns this and I'm left behind?" That's a real fear. But a $997 course on ChatGPT doesn't resolve it. Using the tools at work does. The actual skill gap in AI isn't between people who took a course and people who didn't. It's between people who tried the tools on real work and people who didn't. You can close that gap right now, today, for $0. Open ChatGPT. Type something. See what happens. Iterate. That's the course. Frequently asked questions Are AI courses worth paying for? Almost never, especially if they're charging $500 or more. The free tools (Claude, ChatGPT), their official documentation, and YouTube tutorials from people who actually use the tools will teach you more. The rare exception: a very specific technical course on building with AI (APIs, fine-tuning) from a credible source, if that's genuinely what you need for your job. What makes an AI masterclass a scam? The clearest sign is that the fear comes before the offer. If the pitch starts by telling you AI is replacing people like you, then immediately presents a course as the solution, that's a sales funnel, not education. Other tells: pricing ladders, proprietary framework names, results shown only as screenshots, and credentials that trace back to taking a similar course 6 months earlier. What's wrong with prompt engineering courses? Prompt engineering as a distinct, teachable skill has largely been overtaken by better models. Indeed data showed prompt engineering job listings declining sharply through 2024 after a brief spike. Modern LLMs understand poorly worded prompts much better than older ones did. You don't need a system to write good prompts; you need to practice writing them on real tasks. How do I tell if an AI influencer is legit? Look for specifics. Do they describe what tool they used, for what task, and what the actual output looked like (including failures)? Do they update their thinking when things change? Do they have a product to sell you immediately after making you worried about AI? Legit people share messy, specific experiences. Grifters share clean screenshots and sell courses. Is there actually a good way to learn AI without spending money? Yes. Open Claude or ChatGPT, pick something you'd normally do at work this week, and try to get the AI to help you do it faster. Do that for a few weeks. You'll learn more than any course teaches, because you'll be learning for your specific job, your specific tasks, and your specific failure modes. The beginner guide to using AI at work is a useful starting point if you want some structure. Why do so many people buy AI courses if they don't work? Because the fear is real, even if the solution is fake. If your boss just mentioned "AI strategy" in a meeting and you have no idea what's going on, paying $997 for something that promises to fix that feels rational. It's not. But the anxiety that drives the purchase is legitimate. The better move is 30 minutes with the actual tools, not a curriculum designed to make you feel productive while you avoid using them. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI Skills for Non-Technical People (2026) The AI skills non-technical people actually need in 2026. No coding, no bootcamp, no $997 course. Four real skills and a minimum viable stack by profession. ai skillsnon-technicalprompt engineering AI Anxiety at Work: How to Stop Spiraling 65% of employees feel AI anxiety at work. Here's how to convert that fear into something useful without buying a $997 course or losing more sleep. ai anxietyai job fearwill ai replace my job AI Job Replacement Statistics: What the Data Says Goldman Sachs, WEF, McKinsey: what AI job replacement statistics actually say vs. what gets quoted in headlines. The real numbers, with context. AI job replacementAI statisticsfuture of work --- ## The AI Skills Gap: Power Users Pull Away Fast URL: https://dontreplace.me/articles/the-ai-skills-gap-is-real-and-power-users-are-winning ai skills gapai power usersai productivityai upskillingfuture of work The AI Skills Gap: Power Users Pull Away Fast The AI skills gap is real and measurable. Power users are 4x more productive than casual users. Here's what they do differently and how to close the gap. March 26, 2026 Updated 12 May 2026 There are two kinds of people using AI at work right now. One of them is four times more productive than the other. That’s not a metaphor. That’s the AI skills gap, and it’s already showing up in performance reviews, promotions, and headcount decisions. The gap isn’t between people who use AI and people who don’t. It’s between people who use it well and people who use it badly. And “badly” doesn’t mean wrong. It means shallow. One prompt, one answer, copy and paste, done. That’s the pattern for most people. It’s fine. It’s also leaving most of the value on the table. If you’ve been wondering why some colleagues seem to be doing twice the work in half the time, this is usually why. They figured something out that nobody taught them. And you can figure it out too, but only if you understand what they’re actually doing differently. The AI skills gap is real and it’s already measurable In 2025, OpenAI published research showing a 4x productivity gap between power users and typical employees using the same tools. Not different tools. Same tools. The difference was in how they used them. Anthropic followed up in March 2026 with research showing that power users tackle complex, iterative, multi-step tasks while beginners mostly do simple one-shot queries. They’re not just asking better questions. They’re having different kinds of conversations with the AI entirely. Gartner estimates that 80% of software engineers will need AI upskilling by 2027. That’s not a skills shortage in the traditional sense. It’s a skills split. The top tier is pulling away fast, and the middle is realizing it too late. This is the part nobody tells you when they hand you a ChatGPT account and call it “digital transformation.” Access to the tool isn’t the same as knowing how to use it. A typewriter doesn’t make you a writer. A scalpel doesn’t make you a surgeon. And a ChatGPT subscription doesn’t make you a power user. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What power users actually do differently It’s not about which tools they use. Power users and casual users often use the same tools. The difference is in how they approach the problem before they open the tool. A casual user has a task. They describe the task to the AI. They take whatever comes out and submit it with minor edits. They think they’re using AI. They are. Badly. A power user has a task. They break it into pieces. They prompt for one piece, evaluate the output, iterate, redirect, combine the result with something else, run it through a different tool, check it against their own judgment, and ship something that’s genuinely better than what they could do alone. They’re not prompting. They’re cooking. That’s the exact framing from Chapter 8 of Don’t Replace Me: “A chef doesn’t follow one recipe. A chef understands ingredients, techniques, heat, timing. They improvise. They taste as they go.” Power users treat AI like a kitchen, not a microwave. Here’s what that looks like in practice: Casual user Power user One prompt, one output Multi-step, iterative workflow Accepts first response Critiques, redirects, refines Uses AI for whole tasks Uses AI for specific subtasks Never checks the output Cross-checks against own expertise One tool for everything Right tool for each job Outputs look like AI wrote them Outputs sound like themselves The last row is the one that matters most for your career. If your AI output is indistinguishable from the AI output of your colleagues, you have no edge. You’re all using the same accelerant and going the same speed. Power users bring their own judgment, taste, and domain expertise to the output. That’s what makes the difference visible. For a deeper look at which human skills hold their value when AI does the heavy lifting, the breakdown at jobs AI can’t replace is worth reading. Why most AI training fails to close the gap DataCamp’s 2026 workforce survey found that 82% of organizations offer some form of AI training. The majority of those programs use passive video formats. Watch a module. Take a quiz. Get a certificate. Move on. Those programs don’t produce power users. They produce people who can explain what an LLM is at a dinner party. That’s not nothing. It’s also not enough. The gap between knowing about AI and being good at using AI is the same gap that exists in every skill. You can watch YouTube videos about swimming for a month and still drown. At some point, you have to get in the water. With AI, most people never get in the water. They watch the videos, feel like they’ve done something, and go back to their regular workflow. There’s also a confidence problem. People try one or two things, the output is mediocre, and they conclude the tool isn’t that useful for their work. What actually happened is they didn’t push past the shallow end. The tool is capable of significantly more. They just didn’t know how to ask for it. The research from Anthropic makes this concrete. Power users don’t just ask better questions. They use AI for genuinely hard things: reasoning through ambiguous problems, stress-testing their own thinking, generating options they wouldn’t have considered, catching errors in their own logic. Casual users ask AI to summarize things and write their out-of-office emails. Both are fine. One of them makes you harder to replace. The minimum viable AI stack You don’t need to learn everything. You need to learn enough to be dangerous in your specific context. That means different things for different jobs. The framework from the book is simple: identify the tasks in your work that are repetitive, time-consuming, or miserable, and figure out which AI tool handles each one best. You don’t need 50 tools. You need three to five tools you actually use, not tools you have accounts for. For most office workers, the minimum viable stack looks something like this: A general-purpose LLM (ChatGPT, Claude, or Gemini) for writing, thinking, summarizing, drafting, analysis A document or search tool that connects to your actual files and data One domain-specific tool relevant to your industry or role That’s it. Three tools, used well, beats a browser full of AI bookmarks you open twice. The discipline is in using them consistently and intentionally. Every time you have a task you hate doing, that’s the trigger. Open the tool. Try to do it with AI. Evaluate the result. Iterate. Over time, you get faster, your prompts get sharper, and the output quality improves because your judgment about what “good” looks like gets better too. This is the honest guide to using AI at work without the hype: start with what you already hate doing, and build from there. The 40-hour gap and how to close it Here’s a number worth sitting with. The difference between someone who uses AI and someone who’s genuinely good at it is about 40 hours of deliberate practice. One focused week. That’s the gap. Most people never put in that week. They dabble. They use AI occasionally, for easy things, without much intention. They stay in the shallow end indefinitely. The power users put in the 40 hours because they had a specific project, or a deadline, or just decided to figure it out. After that week, everything changes. The tool starts feeling different. More like an extension of how you think than a search engine you’re querying. Software developers who build that proficiency see a 40% productivity boost according to GitHub’s research on Copilot. That’s not a marginal improvement. That’s the difference between being the person who ships and the person who’s still in the meeting. For non-technical roles, the gains are less documented but visible to anyone paying attention. The marketing manager who uses AI for research, drafting, and iteration is producing more and better work than the one who doesn’t. The HR lead who’s figured out how to use it for policy drafts and job descriptions is getting through a week of work in three days. The consultant who iterates strategy documents with AI is billing more hours and doing less of the drudge work. The concrete skills you need to get there, without a coding bootcamp or a $997 course, are laid out at AI skills for non-technical people. The real risk isn’t that AI replaces you The scary prediction is always that robots take your job. The actual risk is quieter and more immediate: someone at your level, with your experience, figures out the power user workflow before you do. Then they’re doing your job better and faster. That’s when the math changes. This isn’t doom. It’s arithmetic. And you can change the math. The companies with the best AI adoption rates right now aren’t the ones with the most sophisticated tools. They’re the ones where a meaningful percentage of employees put in real time learning to use the tools well. The gap is a skills gap, which means it’s closable. You’re not competing with a machine. You’re competing with your colleagues who are willing to spend 40 focused hours getting good at something. If you want to build the longer-term career strategy around this, the combination moat framework is worth reading: how to future-proof your career against AI covers the whole picture. The tools aren’t hard. The discipline is. Pick one task you hate. Do it with AI this week. Do it again next week, but try to do it better. That’s the 40 hours. Nobody sells that as a course because it doesn’t sound impressive enough. It’s also exactly how every power user got there. Frequently asked questions What is the AI skills gap and why does it matter? The AI skills gap is the measurable difference in productivity between employees who use AI well and those who use it superficially. OpenAI’s 2025 research found a 4x productivity difference between power users and typical users of the same tools. It matters because organizations are starting to notice this gap in performance reviews and headcount decisions. How long does it take to become an AI power user? About 40 hours of deliberate practice, according to people who’ve tracked their own learning. That’s roughly one focused week where you use AI for real tasks, iterate on the outputs, and push past the shallow single-prompt approach. Most people never put in this week, which is why the gap exists. What do AI power users do that casual users don't? Power users break tasks into pieces, iterate on outputs, combine multiple tools, and apply their own domain judgment to the results. Casual users run one prompt and accept whatever comes out. The behavioral difference is significant: Anthropic’s March 2026 research found power users tackle complex multi-step reasoning tasks while beginners stick to simple queries. Does AI training at work actually help? Most of it doesn’t, unfortunately. DataCamp’s 2026 survey found 82% of organizations offer AI training, but the majority use passive video formats that build awareness without building competency. Hands-on practice with real tasks is the only thing that actually closes the gap. Do I need coding skills to become an AI power user? No. The productivity gains are documented across non-technical roles including marketing, HR, consulting, and operations. You need to understand how to structure requests, evaluate outputs critically, and iterate. None of that requires code. See the guide to AI skills for non-technical people for specifics. Which AI tools should I actually be using at work? For most office workers, a general-purpose LLM like ChatGPT or Claude handles the majority of use cases. Add one document or search tool connected to your actual files, and one domain-specific tool for your industry. Three tools used consistently beats a collection of tools you open twice. For specific prompts and workflows, the ChatGPT at work guide is a good starting point. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading How to Use AI at Work: A No-BS Starter Guide The practical guide to using AI at work without the hype. Start with the tasks you hate, brief it like a smart intern, and pick 2 tools. That's the whole plan. how to use AI at workChatGPT for workAI productivity AI Skills You Actually Need in 2026 (No Coding Required) The AI skills gap is widening. Here are the non-technical skills that actually increase your leverage and make you harder to replace. AI skillsnon-technicalcareer growth How to Future-Proof Your Career Against AI A practical framework for building your combination moat so AI amplifies your work instead of replacing it. future-proof careerAI strategycombination moat --- ## What AI Can and Can't Do: The 5-Minute Explanation URL: https://dontreplace.me/articles/what-ai-can-and-can-t-do-the-5-minute-explanation what ai can dowhat ai cant doai explained simplyai limitationsai capabilities What AI Can and Can't Do: The 5-Minute Explanation Plain-language breakdown of what AI actually is, what it's genuinely good at, and where it falls completely flat. No jargon, no hype, five minutes. May 18, 2026 Updated May 18, 2026 In 1770, a chess-playing robot toured Europe and beat everyone it faced. Royalty. Grandmasters. Benjamin Franklin. The machine was called The Turk, and it caused a sensation for 84 years before anyone admitted the truth: there was a human chess master folded up inside the cabinet, moving the pieces. The lesson isn't that people were gullible. The lesson is that "looks like intelligence" and "is intelligence" are two completely different things. We've been confusing them ever since. What AI can and can't do is the question everyone's actually asking when they panic about their job. The honest answer isn't scary or reassuring. It's just useful. So here it is, in plain English, no jargon, five minutes. What AI actually is (and no, it's not a brain) The most accurate description anyone has come up with is "spicy autocomplete." That's it. That's the whole thing. When you type a message and your phone suggests the next word, that's autocomplete. AI does the same thing, but at a scale that makes it look like magic. It's been trained on an incomprehensible amount of text, images, and data. So when you ask it a question, it doesn't "think" about the answer. It predicts, at a statistical level, what words should come next given everything it's seen before. No reasoning. No understanding. No awareness that it exists. Just very, very fast pattern matching. Amazon even named a product after this illusion. Amazon Mechanical Turk, their crowdsourcing platform, is named after the 18th-century chess robot. Humans do tasks that are labeled as "AI." The name is basically a confession: sometimes what looks like a machine is just a person hiding inside. Modern AI isn't hiding a person. But it is hiding the fact that there's no understanding happening either. What AI can and can't do, part one: what it's genuinely good at This is where people undersell the tools, because they're too busy either worshipping them or dismissing them. AI is fast. Faster than any human at the tasks it can handle. Give it a 50-page report and ask for a summary. It'll do it in seconds. Give a human the same job, they'll do it in 45 minutes and still miss two things on page 31. Here's what AI does well: Speed and volume. Processing more text, data, or requests in a minute than a team could handle in a day. First drafts. A blank page is the enemy. AI kills the blank page. The draft is bad. Edit it. You're still ahead. Pattern recognition. Spotting trends in data, flagging anomalies, finding the thing that doesn't fit. Summarization. Long document, short summary. Every time. Translation. Not perfect, but good enough for most professional contexts. Repetitive generation. Writing 200 product descriptions, 50 subject line variants, 30 meeting agendas. The stuff that was boring and slow before. The key word in all of these is "fast." AI isn't better than you at creative judgment, nuanced communication, or genuine problem-solving. It's just faster at the mechanical parts of those tasks. It's also worth being specific about the gap. A McKinsey report on AI and the future of work found that generative AI could automate tasks accounting for 60 to 70% of employee time across industries. That's not 60 to 70% of jobs. That's the fraction of time within most jobs that's spent on mechanical, repeatable work. The rest stays human. If you want to understand which specific tasks to hand off first, the no-BS starter guide to using AI at work walks through exactly that. What AI can and can't do, part two: the hard limits Here's Rule #10 from Don't Replace Me: "It knows everything. It understands nothing." That sentence is doing a lot of work. Here's what it means in practice. It can't tell when it's wrong. This is the big one. AI doesn't know what it doesn't know. If you ask it a question it doesn't have good data for, it will still answer confidently, in complete sentences, with the same tone it uses when it's absolutely right. The word for this is "hallucination." The problem is you can't always tell which answers are real and which are confabulated. You have to check. That's a human job. It doesn't understand meaning. Dee spent a session telling an AI "you are absolutely right" six times in a row. It kept agreeing and kept missing the point. That's not stupidity. It's the absence of comprehension. The system processes tokens. It doesn't know what the words mean in any real sense. It doesn't know that your email to your CEO needs a different tone than your email to your intern. It doesn't know that the joke you're about to send will land badly with this particular client. You know those things. It doesn't. It can't exercise judgment. Judgment means deciding between options when there's no clear right answer. When your boss asks you to prioritize two projects and both matter, you read the room, you think about politics, you weigh relationships. AI can give you a framework. It cannot make the call. It doesn't care about outcomes. This sounds philosophical, but it's practical. You care whether your client is satisfied. You care whether your team is burning out. You care whether the presentation lands. AI produces output. What happens next is not its problem. It can't read a room. Physical presence, group dynamics, the moment when the meeting energy shifts, the colleague who's technically agreeing but whose face says something else. None of this is accessible to a language model. And a lot of actual work depends on exactly this. It has no taste. Taste is knowing what's good before you can explain why. It's the thing that separates a good editor from a grammar checker. AI can follow rules. It can't develop taste, because taste comes from years of experience, failure, and caring about the outcome. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → A quick reference: what AI handles vs.what it doesn't The split isn't random. There's a pattern. Tasks that are well-defined, text-based, and repetitive tend to go to AI. Tasks that require judgment, context, or human relationships tend to stay human. AI handles wellHumans still own Summarizing long documentsDeciding what the summary means Drafting routine emailsKnowing when the email shouldn't be sent Generating data reportsJudging which data actually matters Producing first draftsKnowing when a draft is wrong in ways a spell-check won't catch Spotting patterns in dataUnderstanding why the pattern exists Translation and localizationReading the cultural subtext Repetitive content at scaleMaking anything that requires genuine taste The dividing line is basically: can you write a complete instruction manual for this task? If yes, AI can probably do it. If the task requires improvisation, relationship awareness, or accumulated judgment, it can't. Not yet, and not in any timeline that's clear. Why this distinction actually matters for your specific job Here's the thing nobody says in the LinkedIn posts: AI being bad at these things is good news for you. If your job is purely mechanical, repetitive, and output-based, yes, worry. Some of those jobs are going. But most jobs aren't that. Most jobs involve a mix: some mechanical parts (the stuff AI is fast at), and some human parts (the stuff AI can't touch). The people who are actually at risk are the ones who refuse to use AI for the mechanical parts, because they end up slower than the person next to them who does use it. The threat isn't robots replacing humans. It's humans using AI replacing humans who don't. Think about what a mid-level marketing manager actually does in a week. They write briefs. They sit in strategy meetings. They review agency work. They manage up. They manage down. They send a lot of emails. They occasionally fight for budget. AI can help with the briefs and the emails. It cannot sit in the strategy meeting and notice that the CMO's energy completely changed when the Q3 numbers came up. It cannot fight for budget. It cannot manage the relationship with the agency contact who's difficult but brilliant. The mechanical tasks shrink. The human tasks remain, and they become more visible because you have more time for them. The practical move is simple: figure out which parts of your job are mechanical and which parts are human. Hand the mechanical parts to AI. Put more of your time into the human parts. You get faster at the boring stuff, you get better at the valuable stuff, and you become harder to replace. If you want to know which specific skills are genuinely hard for AI to replicate, the breakdown of jobs AI can't replace gets specific by role and industry. The visibility problem (why this feels new when it isn't) You've been using AI for years. Spam filters are AI. Autocorrect is AI. Netflix recommendations are AI. Google search ranking is AI. Your bank's fraud detection is AI. None of those things caused a mass panic. Why is this different? Because ChatGPT put a text box in front of you and started generating complete sentences. That made the AI visible in a way it wasn't before. You weren't scared of the spam filter because you never had to talk to it. When the AI talks back in paragraphs, it feels like a person. It isn't. But it feels like one. The panic isn't about what changed. It's about what became visible. The tools got better, yes, and some jobs really are at risk. But the hysteria is bigger than the actual threat, because the actual threat is nuanced and nuance doesn't trend on LinkedIn. The World Economic Forum's Future of Jobs Report projects that AI will displace 85 million jobs by 2030 while creating 97 million new ones. That net positive number gets buried under every headline about what's being lost. Understanding what the job replacement statistics actually say versus what the headlines say is a decent place to ground yourself if the panic is getting loud. The confidence problem: why AI sounds so sure when it's wrong This deserves its own section because it's the thing that catches smart people out. AI doesn't have a confidence meter. It can't tell you "I'm 90% sure about this" versus "I made this up." It produces output in the same tone regardless of whether the underlying information is solid or completely fabricated. A hallucinated statistic sounds exactly like a real one. A made-up legal citation reads identically to an actual court case. A plausible-sounding product recommendation might be based on nothing at all. The Stanford HAI research group has documented this pattern extensively: models produce confident, fluent text even when their outputs are factually wrong, internally inconsistent, or simply invented. The problem isn't that AI is wrong sometimes. Every source is wrong sometimes. The problem is that it's wrong with complete confidence and no warning label. This matters practically. If you're using AI to draft customer communications, legal summaries, financial analyses, or anything where accuracy is load-bearing, you need a human in the loop checking the output. That's not a temporary workaround until AI gets better. It's the correct workflow, because AI's role is generating, and your role is verifying. The people who get burned are the ones who outsource both steps. The one-sentence summary AI is a tool that's fast and tireless at mechanical tasks, completely blind to meaning and judgment, and incapable of caring what happens next. Use it for speed. Keep the parts that require you to be human. That's it. That's the five minutes. Frequently asked questions What does "AI is just spicy autocomplete" mean? It means AI generates text by predicting what words should come next, based on patterns in its training data. It's not reasoning or understanding. It's a very fast, very large pattern-matching system that produces confident-sounding output even when it's wrong. Can AI actually think or understand things? No. AI processes statistical patterns in data. It doesn't understand what words mean, can't form intentions, and has no awareness of its own outputs. It can produce text that sounds thoughtful without any thinking happening. This is why checking AI outputs matters, because it can't tell you when it's wrong. What can AI do better than humans? Speed and volume, mostly. AI can summarize documents, generate first drafts, spot patterns in data, translate text, and produce repetitive content much faster than any human. It doesn't get tired, distracted, or bored. For the mechanical parts of most jobs, it's genuinely useful. What can't AI do? AI can't exercise judgment, read a room, develop taste, understand meaning, care about outcomes, or know when it's making something up. Any task that requires empathy, political awareness, context sensitivity, or genuine creativity still belongs to humans. See the full breakdown of jobs AI can't replace for role-specific examples. Will AI replace my job? Probably not the whole job. Likely some tasks within it. The risk goes up if you refuse to use AI for the mechanical parts of your work, because someone who does use it will be faster than you at those parts. The honest answer, with data, is at will AI replace my job. Do I need to understand how AI works to use it? No. You don't know how an internal combustion engine works and you can still drive to the grocery store. The basics help you set appropriate expectations (it can be wrong, it doesn't understand you, check important outputs). Beyond that, just use it and see what it does. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management AI Accessibility Testing Prompts: Catch Barriers 10 AI accessibility testing prompts for keyboard, screen reader, forms, contrast, zoom, motion, media, cognitive checks, and human-verified evidence. AI promptsaccessibility testingquality assurance --- ## What Jobs Will AI Replace? A Role-by-Role Guide URL: https://dontreplace.me/articles/what-jobs-will-ai-actually-replace-honest-assessment AI job replacementjobs AI will replacecareer riskAI automation by industryfuture of work What Jobs Will AI Replace? A Role-by-Role Guide Finance, legal, marketing, tech, creative: an honest breakdown of which jobs AI will replace, which survive, and what to do about it. March 25, 2026 Updated 12 May 2026 The question everyone's actually asking isn't "will AI replace jobs." It's "will AI replace my job." And the honest answer varies wildly depending on what you do all day. Some roles are genuinely cooked. Others are barely touched. Most are somewhere in the middle, with certain tasks getting automated while the core work stays stubbornly human. This is a role-by-role breakdown of what jobs will AI actually replace, what survives, and what you should do about it. No cheerleading, no doom. Just the assessment. The Kodak test: how to know if you're in real trouble Before we go role by role, here's the framework that matters most. Ask yourself one question: is AI automating tasks in my field, or is it making my entire field obsolete? Kodak didn't fail because cameras got better. They failed because the product itself, physical film, stopped being a thing people needed. That's a different problem than "some of our workflow is getting faster." Most workers are facing the workflow version. A few are facing the Kodak version. The difference matters enormously, because the right response is completely different. One calls for adaptation. The other calls for a pivot. The WEF's 2025 Future of Jobs Report estimates that 92 million jobs will be displaced globally by 2030, but 170 million new roles will emerge. Net positive, on paper. But that math doesn't help you if you're in one of the 92 million. What jobs will AI actually replace in finance? The honest verdict: highly exposed at the junior level, surprisingly resilient at the top. A junior financial analyst spends a meaningful chunk of their week pulling data, formatting reports, building basic models, and summarizing earnings calls. All of that is automatable right now, today, with tools that exist. Goldman Sachs has been deploying AI for code and document generation internally for years. The entry-level work is shrinking. But here's what doesn't go away: the client call where someone needs to hear a trusted voice tell them their portfolio is fine. The M&A negotiation where the human relationship is the whole point. The judgment call on a deal that doesn't fit any model cleanly. The senior analyst who's watched three market cycles and has a feel for when the numbers are technically right but the story is wrong. That's not in the model. What to do if you're in finance: Stop competing on speed and accuracy. Machines already win there. Compete on trust, interpretation, and client relationships. If your entire value proposition is "I can build a clean Excel model fast," that's a problem worth addressing now, not later. What jobs will AI actually replace in legal? The honest verdict: document-heavy roles are in trouble. Courtrooms are not. Legal research, contract review, due diligence, first-draft boilerplate. McKinsey estimates that about 22% of a lawyer's work is highly automatable. That number is probably conservative for junior associates at large firms, where 60-70% of billable hours go toward exactly those tasks. The Kodak test here: law itself isn't going away. Litigation, negotiation, strategy, ethics opinions, the actual job of advocating for a client in a room with consequences. That's not automatable. But if your role is "junior associate who reviews 400-page contracts for defined terms," you should be concerned. Not panicking, but honestly concerned. Paralegals and legal assistants face a sharper version of this. The routine document work that defines many of those roles is already being compressed. A single lawyer using AI can handle research that used to require two or three support staff. The math isn't pretty. What to do if you're in legal: Push toward client-facing work as fast as you can. Learn the AI tools so you can do in two hours what used to take eight. Then use those six hours to do the work that can't be automated. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Marketing and content: what jobs will AI actually replace here? The honest verdict: commodity content is gone. Strategic and brand work is fine. This one stings because it's already happening, not in five years. If your job is to produce a certain volume of SEO articles, product descriptions, or social captions per week, AI does that now. Companies are already using it. The demand for pure volume content production has cratered. The work that survives: brand strategy, campaign thinking, knowing why something will resonate with a specific audience, creative direction, client and stakeholder management. These require taste and context that AI genuinely lacks. One honest test: could a smart intern follow a checklist to produce what you produce? If yes, you're in the automatable zone. If your work requires years of context about a brand, an audience, or a client relationship, you're much safer. More on what humans still own in this space. What to do if you're in marketing: Build the taste. Build the relationships. Position yourself as the person who directs AI, not the person who competes with it. The marketers who are thriving right now are the ones who use AI to produce the commodity stuff in 20 minutes and spend the rest of their day on the thinking. Customer service: this one's not subtle The honest verdict: Tier 1 support is being automated at scale, right now. This isn't a prediction. It's a description of what's already happening. Chatbots handle enormous volumes of basic customer queries. Voice AI is getting good enough to handle calls that used to require humans. The Gartner forecast from a few years back predicted chatbots would be a primary channel by 2027. We're on track. The customer service roles that survive are the ones handling genuinely complex, high-stakes, or emotionally loaded situations. The customer who is furious and threatening to leave. The situation that doesn't fit any script. The enterprise account that needs a real human relationship manager. These are the situations where a bot getting it wrong has real consequences, and companies know it. If your job is "answer the same 12 questions in slightly different orders all day," that job is going away. Not eventually. Now. What to do: Move toward complex case resolution, account management, or team leadership. The human judgment tier of customer service isn't automated. The scripted tier is. Tech roles: developers aren't safe either The honest verdict: writing code is getting automated. Knowing what to build is not. GitHub Copilot, Claude, and ChatGPT write decent code for standard tasks. Junior developers who spend most of their time writing boilerplate are facing real compression. A senior developer using AI tools can produce significantly more than they could before, which means teams need fewer juniors. But here's what's not going anywhere: understanding the problem, making architectural decisions, debugging weird edge cases, managing the humans on a project. Software engineering is still fundamentally a thinking job. The thinking part has AI assistance now. The people who can think clearly about complex systems are more valuable than ever. The junior pipeline is the actual concern. If AI tools let senior engineers do 30% more, hiring for entry-level shrinks. This is already showing up in tech hiring data, with entry-level software roles taking longer to fill and some companies announcing they're shrinking graduate intake. What to do if you're in tech: Invest in the thinking skills. System design, architecture, product sense. Use AI aggressively so you build a track record of shipping things, not just writing code. Creative work: the nuanced one The honest verdict: output-for-hire is at risk. Genuine creative direction is not. Stock illustration, generic graphic design, templated video work, standard copywriting. These categories face real pressure. Not because AI makes perfect art, but because it makes "good enough for the brief" art at near-zero marginal cost. Original creative work, the kind with a perspective and an author behind it, is holding up better than the panic suggests. The photographers, illustrators, and designers who have a genuine point of view and an audience are mostly fine. The ones who competed purely on speed and technical execution are in a harder spot. Dee Kargaev covers this tension in Don't Replace Me, specifically the idea that taste is the one thing AI can't replicate. You can prompt it, but you can't inject 15 years of knowing why something works. A creative director who knows what's wrong with a piece without being able to fully articulate it is doing something AI genuinely can't do. What to do in creative fields: Build the perspective, not just the skill. The more your work looks like anyone could have made it with the right prompt, the more exposed you are. Healthcare and trades: safer than most The honest verdict: physical presence and human contact are significant moats. A nurse, a plumber, an electrician, a physical therapist. These jobs require being in a room with a person or a system and responding to what you find there. AI isn't replacing a nurse who's managing a patient's emotional state while monitoring vitals. It isn't replacing the electrician diagnosing a weird fault in a 40-year-old building. Diagnostic support AI is real and useful in medicine. It helps radiologists read scans faster, flags things that might be missed. But it's assisting, not replacing, and the liability and trust dynamics of healthcare create enormous friction for full automation. Trades are genuinely among the most AI-resistant jobs that exist. You can't Zoom a leaking pipe. You can't prompt your way through retrofitting an old commercial kitchen. The physical world is still stubbornly physical, and the people who know how to work in it have a durable edge. What to do: Use the AI tools for the administrative and documentation parts of your work. Free yourself up for the parts that require presence and judgment. That's the formula. HR and management: the awkward middle The honest verdict: administrative HR is exposed. People management is not. HR is an interesting case because it splits so cleanly. Recruiting workflows, job description generation, benefits administration, onboarding documentation, policy drafts. All of that is being compressed. AI can screen resumes faster than any team, generate compliant documentation from a template, and answer employee questions through a chatbot at 11pm. What it can't do: the conversation with an employee who's underperforming and doesn't know why. The judgment call on a misconduct situation that's genuinely ambiguous. The manager who reads a team's energy and knows something is wrong before it becomes a problem. Organizational politics, frankly. AI has no idea what's happening between two department heads, and it can't fix it. If your HR role is mostly transactional, that's the part under pressure. If you're doing the actual human work of managing people through hard situations, you're in better shape than the LinkedIn panic implies. What to do if you're in HR or management: Use AI for the paperwork. Protect your time for the conversations. The managers who are hardest to replace are the ones who are genuinely good with people, not the ones who are good at processes. The honest overall picture on AI job replacement Most people won't lose their job to AI directly. They'll face a more uncomfortable version: their role will evolve, some of what they do will get automated away, and the people who adapt fastest will get more of the available work. The real statistics on AI job replacement don't support the apocalyptic version. But they don't support "nothing to see here" either. The honest read is this: if you're in a role where most of your value comes from executing repetitive, codifiable tasks, you need to move. Not necessarily out of your field, but up the value chain within it. The parts of work that require trust, taste, judgment, and human presence are holding up. The parts that are essentially information processing are getting compressed fast. Check the perimeter. Run the Kodak test on your own role. And if the answer is uncomfortable, treat that as useful information rather than something to argue with. If you want the full framework for figuring out when to dig in versus when to bail, how to future-proof your career against AI is the practical next step. Frequently asked questions What jobs will AI replace first? Roles with high volumes of repetitive, codifiable tasks are getting automated fastest. Junior financial analysts, basic customer service agents, document review paralegals, and generic content producers are already seeing compression. These aren't predictions for 2030, they're descriptions of what's happening now. Will AI replace lawyers? Not the practice of law, but significant portions of legal work are already automatable. Contract review, due diligence, and legal research are being compressed. Litigation, client strategy, negotiation, and courtroom work are far less exposed. Junior associate roles at document-heavy firms face the most near-term disruption. Is creative work safe from AI? It depends on the type. Original work with a genuine perspective and a recognizable voice is holding up. Generic, templated, output-for-hire creative work (stock illustration, standard copywriting, templated design) faces real pressure from AI tools that produce "good enough" output cheaply. Build a point of view, not just a skill set. Will AI replace software developers? Not senior developers with strong systems-thinking skills. But the junior pipeline is facing compression. Senior developers using AI tools can produce significantly more than before, which reduces demand for entry-level hires. If you're in tech, invest in architectural thinking and product sense, not just code output. How do I know if my specific job is at risk from AI? Run the Kodak test. Ask whether AI is automating tasks within your role, or making your entire field obsolete. Then ask what percentage of your week involves codifiable, repeatable tasks versus judgment, trust, relationships, and context. The higher the first number, the more exposed you are. More detail at /will-ai-replace-my-job/. What's the timeline for AI job replacement? The WEF's 2025 Future of Jobs Report estimates 92 million jobs displaced and 170 million created by 2030. But displacement is already happening in specific task categories now, not in five years. Don't wait for 2030 to start adapting. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading Jobs AI Can't Replace (And Why) Which jobs are genuinely hard for AI to replace? From care work to skilled trades, here's what the data says about human skills that still matter. jobs AI can't replaceAI-proof careershuman skills Will AI Replace My Job? The Honest Answer 65% of workers fear AI will replace them. Here's what the data actually says about AI job replacement and what to do about it. AI job replacementwill AI replace my jobAI anxiety AI Job Replacement Statistics: What the Data Says Goldman Sachs, WEF, McKinsey: what AI job replacement statistics actually say vs. what the headlines tell you. Real numbers, real context. ai job replacement statisticsai job lossfuture of work --- ## What jobs will AI actually replace? Honest by role URL: https://dontreplace.me/articles/what-jobs-will-ai-actually-replace-honest-by-role AI job replacementwhat jobs will AI replaceAI career riskfuture of work What jobs will AI actually replace? Honest by role Role-by-role honest assessment of which jobs AI will replace, from finance and legal to trades and healthcare. Includes the Kodak test framework. May 19, 2026 Updated May 19, 2026 The question everyone's actually asking isn't "will AI replace jobs?" It's "will AI replace my job?" And the honest answer is: it depends on which job, and which parts of it. Some roles are already feeling it. Others are mostly absorbing panic from LinkedIn. The useful thing isn't a generic reassurance. It's a role-by-role look at what's genuinely automatable, what stays human, and what the data says about timelines. So here's the honest assessment nobody on your feed wants to give you, because it requires saying some uncomfortable things. The Kodak test: are your tasks changing, or is your whole industry dying? Before we get into specific roles, there's a useful frame from Don't Replace Me called the Kodak test. Kodak didn't lose because photography got slightly harder. They lost because the reason people hired them disappeared. Digital photos don't need film, full stop. The question for your career isn't "can AI do some of what I do?" Almost everyone's answer is yes. The real question is: is AI making the core reason you're hired irrelevant, or is it just shifting how you do your work? Those are completely different situations. One requires adapting. The other might require leaving. Keep that distinction in mind as we go through each sector. Some of these are "tasks are shifting" situations. A few are starting to look like Kodak. What jobs will AI actually replace in finance? Finance is where automation has been running longest, and the results are more nuanced than the headlines suggest. McKinsey estimates that banking and insurance have some of the highest potential for gen AI impact across all industries. What's automatable: routine data entry, transaction processing, basic financial report generation, fraud detection alerts, loan pre-screening, and a lot of the junior analyst work that used to involve pulling numbers from spreadsheets for 12 hours straight. What stays human: advising clients through emotionally charged decisions (the divorce, the inheritance, the retirement fear), detecting the fraud that doesn't look like fraud yet, building the relationships that make someone call you first when they have money to move, and the judgment calls that don't fit a template. The roles most at risk are entry-level data processing jobs and anything that's essentially "run the same report in a slightly different format." Mid-to-senior finance professionals who develop judgment, client relationships, and contextual knowledge are in a different conversation. What to do: if you're in finance and most of your day is data manipulation, learn the tools doing that automation. You want to be the person who understands what the AI is producing, not the person competing with it. Legal: the research revolution that isn't quite what you think Law is a great example of how AI disruption gets distorted in both directions. Legal AI tools like Harvey and Casetext are genuinely impressive at document review, contract analysis, and legal research. Goldman Sachs research identified legal as one of the sectors with the highest share of tasks potentially exposed to AI. But law has a harder wall than most people realize. Courts require licensed humans. Clients need representation by a person. The liability structures of legal work make full AI replacement structurally difficult in the near term. What's automatable: first-pass document review, contract drafting templates, legal research aggregation, billing time tracking, basic compliance checklists, and some of the paralegal work that's essentially structured data extraction. What stays human: advocacy, judgment, negotiation, ethics calls, client counseling, courtroom presence, and all the work where someone needs to trust that a human being is accountable for the outcome. The real concern is at the paralegal and junior associate level, where a lot of the work is the routine research and document review that AI handles well. If you're in one of those roles, you need to be fluent in these tools and able to do the higher-judgment work around them. If you're not, someone who is will do your job plus theirs. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → Marketing: the over-automated mess nobody's talking about honestly Marketing is where AI adoption has been fastest and the results have been most mixed. Every agency in the world is using AI to generate first drafts, SEO copy, ad variations, and social content. A lot of it is visibly worse than what came before. What's automatable: templated copy, basic SEO content at scale, A/B test variations, email sequences, performance report summaries, social media scheduling, and anything that's essentially "write 50 versions of this headline." What stays human: strategy, brand voice, the judgment call about what customers will actually respond to, creative direction, the meeting where someone needs to understand your business before they can say anything useful, and taste. The ability to tell good from average isn't something a model reliably has. The World Economic Forum's Future of Jobs Report puts marketing and communications roles in a nuanced position: high exposure to automation, but also high potential for human-AI collaboration rather than replacement. Content marketers who can't tell the difference between good and AI-slop are already losing work. Content marketers who use AI as a production tool while applying genuine strategic judgment are doing fine. The commodity end of the market is being squeezed hard. That's not the same as the whole field dying. For more on which human skills hold up specifically in creative fields, the jobs AI genuinely can't replace breakdown covers that territory in detail. Customer service: the one where the concern is real Customer service is probably the clearest case where "tasks changing" is shading into "Kodak territory" for some roles. Voice AI and chat AI have gotten good enough that a large slice of customer service interactions, the ones that are essentially "I have a problem, can you fix it according to this script," can be handled without a human. Call center volume for routine inquiries is going to decline. That's not a prediction, it's already happening. What's automatable: first-tier support, FAQ responses, order status, account changes, appointment booking, basic troubleshooting, and any interaction that follows a structured decision tree. What stays human: de-escalation of genuinely angry customers, complex complaints with no clean resolution, situations where someone needs to feel heard rather than processed, enterprise relationship management, and anything where the interaction is what the customer is paying for. If your entire role is handling tier-one support tickets for a software company, that's a vulnerable position. If you're handling complex escalations, managing enterprise accounts, or doing anything that requires genuine human judgment under emotional pressure, you're safer. Tech and software development: the weird case Here's the one that surprises people. Software development was supposed to be one of the safer fields, because surely AI can't code, right? Turns out AI can code reasonably well, and GitHub Copilot and similar tools have already changed how developers work. But the picture is complicated. AI generates code. It doesn't always generate correct code. It doesn't know your system, your constraints, your technical debt, or why you made the architectural choices you made two years ago. What's automatable: boilerplate code generation, unit test writing, documentation, code reviews for obvious issues, bug hunting in known patterns, and simple feature additions. What stays human: architecture decisions, debugging novel problems, understanding the business context well enough to build the right thing, security judgment, and the conversations between engineering and product that require translating between technical and human. Entry-level developers whose entire value was "I can write this function" are in a harder spot. Developers who understand systems, can communicate with stakeholders, and can critically evaluate AI-generated code are in a better one. This is a role that's transforming faster than most. Healthcare: more protected than you think, with exceptions Healthcare has a structural protection that most industries don't: licensing, liability, physical presence, and the fact that the thing being managed is a human body. AI tools are genuinely impressive in diagnostics. Studies have shown AI matching or beating radiologists at specific image-reading tasks. But a radiologist's job isn't just reading images. It's consulting on cases, communicating with other physicians, and making judgment calls in contexts with incomplete information. What's automatable: medical imaging analysis assistance, administrative documentation, prior authorization paperwork, appointment scheduling, and clinical decision support that flags potential issues. What stays human: diagnosis in complex cases, patient communication, physical examination, surgical procedures, mental health treatment, the judgment calls where the data points one way and the experienced clinician points another. Administrative healthcare roles, especially anything that's essentially data entry or paperwork routing, are more exposed than clinical roles. If you're a nurse, physician, or therapist, the tools change your workflow. They don't take your job. Trades: probably the safest sector you're not hearing about Electricians. Plumbers. HVAC technicians. Welders. Anyone who does physical skilled labor in unpredictable environments. These jobs are not being replaced by AI in any meaningful timeframe. AI requires a physical robot body to do physical work, and physical robots doing complex skilled trades work in variable environments (a 1920s house with weird wiring, a basement that floods) are not coming for your job in the next decade. The Bureau of Labor Statistics projects consistent demand growth for most skilled trades through 2030. While white-collar workers are panicking, the electricians are booked out three weeks. What stays human: all of it, mostly. Tools get better. Diagnostics assist. Scheduling software exists. But the work itself requires hands, judgment, and physical presence that isn't going anywhere soon. If you're considering careers for younger people you know, trades deserve a serious look in the current environment. The risk profile is genuinely different from most office work. For a bigger picture on which careers hold up against AI, that's worth reading alongside this. Creative fields: the honest version Writers, designers, illustrators, musicians. The creative fields are absorbing a lot of AI disruption right now, and pretending otherwise would be the kind of dishonesty this site doesn't deal in. Stock illustration, generic blog content, background music for videos, basic logo variations: these markets are being compressed by AI tools. If your creative work is at the commodity end of any of those categories, you're already feeling it. What stays human: distinctive voice, cultural taste, the ability to understand a client's actual problem (not the brief they wrote), work that requires genuine novelty rather than skilled recombination, and the kind of creative direction that requires taste that can be argued for and defended. The creative professionals who are struggling are the ones doing work that was already somewhat interchangeable. The ones doing work with a genuine point of view, who understand their clients deeply, or who are using AI tools to produce more distinctive work faster, are navigating a different situation. If you want the broader picture on what the data says about AI job replacement across industries, the numbers tell a more nuanced story than either the doomers or the dismissers admit. What to do with all of this Apply the Kodak test to your own role. Is the core reason someone hires you being made obsolete, or is the way you do that work shifting? For most people reading this, it's the second thing. Tasks are shifting. Tools are changing. The judgment, relationships, and domain expertise underneath the tasks remain yours. But for some roles, especially entry-level data processing work, commodity content creation, and routine customer service, the Kodak test is looking uncomfortable. And the honest answer there, per Rule #20 in the book, is sometimes the right move is to pivot rather than dig in. Either way, the answer isn't panic. The answer is a clear-eyed look at which category you're in. Then acting accordingly. Frequently asked questions What jobs will AI replace first? Roles built around repetitive, structured tasks are moving fastest. Data entry, basic customer service, routine financial processing, and templated content creation are already seeing significant AI automation. These aren't full job replacements in most cases, but the headcount needed to do those tasks is shrinking. Will AI replace software developers? Probably not in the near term, but the role is changing fast. AI tools like GitHub Copilot handle boilerplate code generation well. Developers who can critically evaluate AI-generated code, make architecture decisions, and communicate with business stakeholders are in a stronger position than those whose value was writing simple functions. Entry-level roles face more pressure than senior ones. Which jobs are completely safe from AI? No role is completely immune to change, but skilled trades (electricians, plumbers, HVAC), complex healthcare delivery, and work requiring real-time physical judgment in variable environments are structurally protected for the foreseeable future. The jobs AI genuinely struggles to replace have physical presence, emotional intelligence, or irreducible judgment at their core. Will AI replace lawyers? AI is already handling significant portions of legal research and document review. But licensed representation, courtroom advocacy, and client counseling have structural protections AI can't easily override. Paralegal and junior associate work is more exposed than partner-level work. The question for lawyers is whether their day-to-day work is in the automatable category or the judgment category. How long until AI replaces most jobs? The WEF's 2025 Future of Jobs Report projects 85 million jobs displaced by AI by 2030, but also 97 million new roles created. That's net positive, but it involves real disruption for specific roles in specific industries. "Most jobs" being replaced is not a realistic near-term scenario. Specific tasks in most jobs being automated is already happening. Is marketing a dying field because of AI? The commodity end of marketing, templated copy, generic SEO content, basic ad variations, is being compressed. The strategic, judgment-intensive end is not dying. Marketers who can think, direct, and evaluate AI output critically are fine. Marketers whose entire value was generating routine content are in a harder spot. The field is splitting, not disappearing. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading Will AI Replace My Job? The Honest Answer The Goldman Sachs "300 million jobs" report explained without panic. What AI job replacement data actually says vs. what LinkedIn tells you. AI job replacementAI anxietyfuture of work AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management --- ## Will AI Replace My Job? The Honest Answer URL: https://dontreplace.me/articles/will-ai-replace-my-job-what-s-actually-happening AI job replacementAI anxietyfuture of workAI statisticscareer advice Will AI Replace My Job? The Honest Answer The Goldman Sachs "300 million jobs" report explained without panic. What AI job replacement data actually says vs. what LinkedIn tells you. May 20, 2026 Updated May 20, 2026 The Goldman Sachs report said 300 million jobs could be "exposed" to AI. That number hit LinkedIn like a grenade. Three hundred million. By the time it reached your feed, "exposed" had become "replaced," and "could be" had become "will be," and suddenly everyone was refreshing their resume at midnight. Will AI replace your job? Here's the honest answer: probably not in the way you're imagining. But the question itself is the wrong one to be asking. What "will AI replace my job" actually means The Goldman Sachs figure everyone quotes comes from a 2023 report on generative AI and labor markets. The word they used was "exposed." As in, some tasks within those jobs could be automated. Not the whole job. Not the person. The tasks. That's a meaningful distinction, and almost nobody made it. Your job is a bundle of tasks. Some of those tasks are repetitive, rule-based, and mind-numbing. Those are the ones AI is genuinely good at. The rest, the judgment calls, the relationship management, the reading-the-room moments, the things that require knowing the history of a client or the personality of a boss, those are still yours. "Exposure" means some of your job got easier, or cheaper, or faster. It doesn't mean you're out. The headline math never adds up Let's look at the other numbers people throw around. The World Economic Forum's Future of Jobs Report 2023 predicts 83 million jobs displaced by 2027, but also 69 million new jobs created. Net loss: 14 million, in a global workforce of over 3.3 billion people. That's 0.4%. McKinsey's task-level research suggests that less than 5% of jobs can be fully automated with current technology. Fully. As in, every task in the role done by a machine, end to end. The 300 million number. The "80% of jobs by 2030" you saw on that podcast. The stats feel massive because they're presented without context, without timelines, without any explanation of what "exposure" or "disruption" actually means in practice. And yes, some jobs will go. Entirely. Some already have. Data entry clerks, certain paralegal functions, image editing tasks that used to take days. This is real. But it's not the apocalypse the headlines sell. Why the panic is profitable (and not for you) Every third LinkedIn post says AI is coming for your job. Usually right before someone sells you a course. That's not a coincidence. Fear is a business model. Scared people buy things. They buy $997 prompt engineering masterclasses. They buy "AI readiness audits." They pay for certifications in tools that will be obsolete in eight months. The people selling you the panic are the same people selling you the solution. That's a conflict of interest so obvious it's almost funny. EY research found that 65% of employees report anxiety about AI and job replacement. Sixty-five percent. That's a lot of nervous systems being rented out to headlines that benefit from the click, not from your career outcomes. Rule #1 in Don't Replace Me names this directly: "Fear Is a Subscription Model." The headlines make money keeping you scared. Your career gets better when you stop paying that subscription. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → What AI is actually doing to jobs right now Here's what's actually happening, on the ground, in real companies. AI is making some people faster. Those people are handling more work, in less time, with less effort. In some cases, that means companies hire fewer people for certain roles. In most cases, it means the people using AI look a lot better than the people who aren't. Most companies are nowhere near replacing whole departments with robots. Most of them still can't update a CRM without starting a small fire. The gap between "AI can do this in a demo" and "AI does this reliably in our actual tech stack" is enormous, and anyone who works inside a real organization knows it. What is changing is the baseline. If you're a writer who takes three hours to draft a client report, and your colleague uses AI to draft one in 20 minutes that you then spend an hour editing, the math on headcount starts to shift. Not because the machine replaced anyone. Because one person got faster and the other didn't. That's the real threat. Not a robot with your nameplate. A coworker with a faster workflow. What "faster" looks like by role This plays out differently depending on what you do. A junior analyst who used to spend two days pulling market research can now get a decent first pass in two hours. A recruiter who used to screen 200 resumes manually can process them in 20 minutes. An HR coordinator who drafted the same onboarding email 40 times a month has automated it entirely. None of those people got fired. But their manager noticed. And the next time a budget conversation happens, the headcount math looks different. The people at risk aren't the ones whose entire jobs are automated. They're the ones whose output looks identical to their AI-assisted colleagues, but takes three times as long. That gap is where the pressure actually lives. The "find your appendix" test for your own job There's a useful exercise from Chapter 1 of the book: find your appendix. Your appendix is the part of your job that's vestigial. You do it. It takes time. But it produces very little that couldn't be produced faster and cheaper by a machine. Think: formatting reports, pulling data into spreadsheets, writing first drafts of standard emails, summarizing meeting notes, researching background information. That's not your whole job. That's a piece of it. And it's the piece that's genuinely vulnerable. The rest of your job, the parts that involve trust, judgment, client relationships, institutional knowledge, navigating office politics, making the call when the data is ambiguous, that's the part AI hasn't touched. Check out what jobs AI genuinely can't replace for a longer breakdown on where those human edges actually live. The practical question isn't "will AI replace me?" It's: "what percentage of my current tasks could be done faster by AI, and am I using AI to do them?" If you're not, someone else in your field is. And they're getting faster while you're staying still. How to actually run the audit Grab a piece of paper. List every task you did last week, roughly. Don't filter, just list them. Then put each one in one of two columns: "requires me specifically" or "requires someone with my general skills." The second column is your exposure. Not your doom, your exposure. Those are the tasks worth experimenting on first. If 80% of your week lands in column two, that's important information. It doesn't mean you're getting fired next Tuesday. It means your job is more vulnerable to the baseline shift than you might have assumed, and you should probably start getting faster at the stuff in column one while using AI to compress column two. Most people who do this exercise are surprised. Not because AI can do everything, but because they've never really looked at how much of their week is genuinely rote. Will AI replace my job? The honest self-assessment Here's a rough way to think about your own vulnerability. Job characteristicLower riskHigher risk Requires physical presenceYesNo Involves high-stakes judgmentYesNo Requires reading peopleYesNo Output is mainly text or dataNoYes Tasks follow predictable rulesNoYes Client relationships are centralYesNo Work is largely solo and digitalNoYes No single factor determines anything. This is a texture check, not a verdict. A lawyer who does nothing but document review is in a different position than a litigator who spends 80% of her time in rooms persuading people. A journalist who summarizes press releases is in a different position than an investigative reporter with sources built over 20 years. Most jobs land somewhere in the middle. Which means most people have some exposure and some protection. The goal is to know which is which for your specific role, in your specific organization, in your specific industry. For a deeper look at which roles face the most pressure, what jobs AI will actually replace breaks it down by sector with a realistic timeline. Where the real dividing line is The honest dividing line isn't "knowledge worker vs.manual worker" or "creative vs. analytical." It's simpler than that. If your job fundamentally requires other people to trust you specifically, as a person, with a history and a relationship, you're in a much better position than someone whose job just requires that the output meets a spec. AI can meet specs all day. It can't be the specific person your client has worked with for six years who knows their business, their preferences, and their boss's personality. That's the thing to build toward. Not "AI-proof skills" in the abstract. Irreplaceability through specificity. The pattern has happened before In 1811, a group of skilled textile workers in Nottingham started breaking machines. They signed their letters "General Ludd." History turned them into a punchline for anyone afraid of technology. But they weren't stupid. They were scared, and they were right to be scared. The machines did change their work. Permanently. What history also shows: most people adapted. The machines stayed. The work changed shape. New jobs appeared that didn't exist before. The telephone killed telegraph operators and created hundreds of thousands of telephone operator jobs. The internet killed travel agents and created social media managers, UX designers, and SEO specialists. The spreadsheet was supposed to eliminate accountants. Instead the number of accountants in the US grew for three decades after Excel shipped. This pattern doesn't mean you personally will be fine automatically. It means the category of "human worker" has survived every previous technology wave, and it will survive this one too. The individuals who got hurt were mostly the ones who didn't adapt, or who adapted too slowly. That's the actual risk. Not that AI arrives one day and takes your job like a repo man. But that the field shifts underneath you while you're waiting to see what happens. Why this wave feels different (and why it isn't, quite) The reason people feel like this time is different is that AI is hitting knowledge work. Previous automation hit physical or highly repetitive tasks. The loom threatened weavers. The calculator threatened human "computers." But your doctor, your lawyer, your writer, your analyst, those felt safe. Now they don't. And that's genuinely new, psychologically if not structurally. But the underlying dynamic is the same. New technology compresses the time required for certain tasks. The people doing only those tasks face pressure. The people whose work goes beyond those tasks adapt and absorb the productivity gain. The economy eventually generates new categories of work, often ones that require the new technology as a baseline. The difference is that the cycle is moving faster now. Adaptation that used to take a generation might take five years. That's uncomfortable. It's not the end. For a more thorough look at how to build a position that holds up, how to future-proof your career against AI gets into the specific moves that actually work. The middle scenario nobody talks about Here's what the discourse skips. The doomers say robots take everything. The optimists say new jobs will appear and everyone will be fine. Both groups are selling a story. The actual middle scenario: your job changes significantly, probably within the next three to five years. Some tasks you currently do disappear or become trivially fast. New expectations appear around speed and output. Your value shifts from "can produce X" to "can produce X and Y and also know which AI output is good and which is garbage." That's a real adjustment. It's not a death sentence. But it does require you to pay attention and move, rather than waiting for clarity that isn't coming. The people who will struggle most are the ones who treat this period like a waiting room. They're watching the news, reading the think pieces, attending the all-hands where the leadership says "we're monitoring the situation." And they're not changing anything about their actual workflow. The people doing well right now are the ones who just started using the tools, found out what they're good for, built a slightly faster version of their job, and moved on. No $997 course. No certification. Thirty minutes with ChatGPT and a specific problem. The real threat isn't a robot. It's a coworker who's been using the tools for eight months while you were still deciding whether to try them. What to do right now that isn't a $997 course You don't need a certification. You need 30 minutes and a ChatGPT or Claude account, which you can get for free. Here's the only thing that matters this week: identify the one task in your job you hate most, the one that's repetitive and draining and you procrastinate on. Then spend 20 minutes asking AI to help you do it. Not because AI will transform your career in a day. Because you'll immediately understand what the tool can and can't do, without anyone explaining it to you. And that understanding is worth more than any course you could buy. The book's core framework for this is Rule #12: Start With the Shit You Hate. The idea is that the tasks you hate most are usually the ones most worth automating, and starting there gives you a quick win without disrupting the parts of your job that actually require you. Once you've done that once, you'll know more about your own vulnerability and your own opportunity than any report or LinkedIn prophet could tell you. Because you'll have specific, concrete information about what AI can do with your actual work, not someone else's demo. That's the whole move. Not a transformation. A test. Frequently asked questions Will AI replace my job completely? Almost certainly not completely, if your job involves judgment, relationships, or physical presence. The McKinsey research suggests less than 5% of jobs can be fully automated with current technology. What's more likely is that AI replaces specific tasks within your job, changing what the work looks like rather than eliminating the role entirely. What does "AI exposure" actually mean in the Goldman Sachs report? It means some tasks within a job could theoretically be done by AI, not that the job disappears. The original report used "exposure" to describe task-level overlap, which headlines quickly inflated to "replacement." A surgeon is "exposed" because AI can analyze scans, not because robots are performing surgeries. How worried should I actually be about AI taking my job? Worried enough to start using the tools, not worried enough to buy a $997 course. The real risk isn't mass replacement; it's that people using AI become visibly faster and more productive than those who don't. EY found 65% of workers feel AI anxiety, but the action that actually helps is experimentation, not stress. What jobs are safest from AI replacement? Jobs requiring physical presence, high-stakes human judgment, emotional attunement, deep client trust, and navigating complex social or political situations. Care workers, skilled tradespeople, therapists, surgeons, and people who build and maintain relationships are all in structurally safer positions. See jobs AI can't replace for more detail. Is there a way to assess my own job's vulnerability to AI? Yes. Make a list of every task you do in a given week, then ask: could a rule-following machine do this if given the right inputs? The tasks that answer "yes" are your appendix. The tasks that answer "no" are your protection. Most jobs have both, which means most people aren't cooked, they just need to know which parts of their work to lean into. Is the WEF really saying AI will destroy 83 million jobs? The WEF Future of Jobs Report 2023 projects 83 million job displacements by 2027 alongside 69 million new jobs created, for a net loss of about 14 million in a workforce of billions. It's a real and significant disruption, not a hiring freeze of one. The new jobs angle rarely makes the headline. How fast is this actually happening? Faster than previous technology waves, slower than the headlines suggest. Most organizations are still in early experimentation. The WEF projects meaningful disruption by 2027, not tomorrow morning. But "not tomorrow" doesn't mean "not soon." If your field has obvious AI applications and you're not paying attention, three years goes fast. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading What jobs will AI actually replace? Honest by role Role-by-role honest assessment of which jobs AI will replace, from finance and legal to trades and healthcare. Includes the Kodak test framework. AI job replacementwhat jobs will AI replaceAI career risk AI A/B Testing Prompts: Run Better Experiments 10 AI A/B testing prompts for hypotheses, metrics, variants, QA, analysis, and rollout decisions—without inventing certainty or outsourcing judgment. AI promptsA/B testingexperimentation AI Acceptance Criteria Prompts: 10 Templates 10 copy-paste AI prompts to turn vague tickets into testable acceptance criteria. Includes edge cases, QA scenarios, and stakeholder sign-off templates. acceptance criteriaAI promptsproduct management --- ## Your Company's AI Policy Is a Mess — What to Do URL: https://dontreplace.me/articles/your-companys-ai-policy-is-a-mess-what-that-means-for-you company ai policyai workplace rulesai at workai disclosureai anxiety Your Company's AI Policy Is a Mess — What to Do Most companies lack a real AI policy in 2026. Here's what the White House framework means for you, and how to protect yourself at work. March 30, 2026 Updated 12 May 2026 Your company’s AI policy is probably a Word document that nobody’s read, a Slack message from eight months ago, or nothing at all. And you’re the one sitting in the middle of it, wondering if using ChatGPT to draft that report last Tuesday was fine or a firing offense. Here’s where things stand with company AI policy in 2026: most organizations are genuinely winging it. A 2024 SHRM survey found that fewer than half of organizations have a formal AI use policy in place. The other half have vibes. And those vibes are inconsistent, unwritten, and occasionally reversed without notice. That’s the mess. And if you’re anxious about it, you’re not being paranoid. You’re being accurate. Let’s figure out what you actually do about it. The company AI policy situation: nobody knows what the rules are The number that matters most here isn’t about AI capabilities or job losses. It’s this: 68% of employees using AI at work don’t tell their employer. Sixty-eight percent. That’s not a handful of rogue workers sneaking tools past IT. That’s most of the people using AI, staying quiet, because the rules are either absent or confusing enough that disclosure feels risky. This isn’t an employee problem. It’s a policy problem. When people can’t figure out whether something is allowed, they default to not asking. That’s rational. If you ask and the answer is “no,” you’ve lost something. If you don’t ask and it’s technically fine, you’re ahead. So people keep their heads down and their Claude tabs minimized. The result is a workplace where AI use is widespread, invisible, and completely undocumented. Which is genuinely bad for everyone. What the White House framework actually says (and what it doesn’t) On March 20, 2026, the White House released its National AI Policy Framework. It sounds important. And in some ways it is. But here’s what it actually does for your workplace situation: not much, immediately. The framework is a set of recommendations, not binding rules. It doesn’t tell your employer what to do. It doesn’t create employee rights around AI use. What it does is signal a federal direction, one that pushes responsibility for AI governance down to the company level rather than up to regulators. The federal government is also actively trying to preempt state-level AI rules, which means the patchwork of state laws some workers were counting on might not survive contact with Congress. The practical translation: if your company was waiting for clear federal guidance before writing an AI policy, that guidance isn’t coming. The government has essentially said “figure it out yourselves.” Which is the policy version of your company’s existing non-policy. HR departments right now are either scrambling to write something, or continuing to ignore the problem while insisting it’s on the roadmap. SHRM has been pushing for “practical, flexible” AI policies since 2024. That advice sounds reasonable. What it means in practice is that the people closest to the work, people like you, are still operating in a gray zone. The three types of company AI stances (and why all of them are a problem) Here’s the honest taxonomy from Don’t Replace Me, and it holds up: your company’s AI position is one of three things. Type 1: No policy at all. Anything goes until it doesn’t. You’re fine right up until someone in leadership has a bad day, sees a news story about an AI data breach, and suddenly there’s an emergency meeting and a ban on “unauthorized AI tools.” Everything you built your workflow around is now a liability. Type 2: A “no AI” policy that everyone ignores. This is the most dangerous one. You have a written prohibition, but the culture treats it as performative. Your manager uses ChatGPT for their weekly updates. Your colleague runs client emails through Claude. You do the same. Then something goes wrong, a bad output, a data incident, a lawsuit, and suddenly the written policy matters very much, and all of you are exposed. Type 3: A reasonable policy that nobody’s read. This one is almost worse because of the false comfort. You think you’re covered. You’re not sure what “covered” means. You definitely haven’t checked whether the tools you’re using are on the approved list. All three of these are genuinely bad situations. They just have different failure modes. What this means for you as an employee right now The policy chaos at the top doesn’t change what you’re doing Monday morning. You still have work. AI tools still exist. The question is how to use them without creating a problem for yourself. First: find out what your company actually has. Search the intranet. Ask HR. Check the employee handbook. If there’s a policy, read it. Not a summary of it. The actual thing. Pay attention to whether it specifies which tools are allowed, whether there are data handling restrictions (there almost certainly should be for client data, financials, anything confidential), and whether there’s an approval process for new tools. If there’s no policy, that’s important information too. It means you need to make your own decisions about disclosure. The 68% who stay quiet aren’t all making the wrong call, but there are scenarios where transparency is the smarter play, especially if you’re in a client-facing role, a regulated industry, or working for someone who’d feel ambushed finding out later. Second: stop treating AI as invisible infrastructure. If you’re using it regularly, that usage is part of your work. Treat it that way. If you’d mention that you used a particular research database or software tool to get something done, apply the same logic to AI. Third: be thoughtful about what goes in. If your company doesn’t have a clear data classification policy for AI, assume that anything client-specific, proprietary, or confidential stays out of external AI tools. This isn’t paranoia. It’s the kind of judgment that keeps you out of trouble when policy finally catches up to practice. And if the anxiety about all this is genuinely getting to you, that’s worth addressing separately. AI anxiety at work is real, and it doesn’t go away by pretending the uncertainty doesn’t exist. But it also doesn’t go away by stopping. The tools aren’t going anywhere. This came from a book. Don't Replace Me 200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff. Get the Book → The personal AI policy: document, verify, demonstrate value Here’s the most practical thing you can do right now, and almost nobody does it. Write your own AI policy. Not a manifesto. Not a memo. Just a simple record of how you use AI in your work. Three things worth tracking: What you used it for. Not in detail, just enough to show the pattern. “Used Claude to draft first version of the Q2 marketing report. Used ChatGPT to summarize competitor research. Used AI to generate initial formatting for client proposal.” What you verified. This is the important one. Every output you acted on, note that you checked it. Confirmed the statistics. Validated the figures. Read the draft before sending. This is your receipt. If someone ever asks whether you just blindly published AI output, you have an answer. What value it created. This one’s optional but useful. “Reduced time on first draft from 3 hours to 45 minutes.” “Generated five concept directions instead of two.” “Caught formatting errors in the report before the deadline.” This turns your AI use from a liability story into a productivity story. Dee Kargaev calls this the CYA Protocol in Part V of Don’t Replace Me, and it’s the kind of thing that sounds tedious until the moment it matters. When policy uncertainty turns into policy enforcement, you want receipts. This is how you get them. The documentation doesn’t have to be formal. A running note in a personal doc is fine. A brief Notion page. Anything you could pull up in ten seconds if asked. The point is having it. How to actually use AI at work while all this gets sorted out The policy situation isn’t going to resolve itself next quarter. Companies are slow, regulation is slower, and the tools keep moving faster than either. So the practical question isn’t “wait until the rules are clear.” It’s “how do I operate sensibly in the meantime?” Using AI at work effectively right now means starting with the low-stakes stuff. Internal documents, research summaries, first drafts that humans review before anyone external sees them. This is both good AI practice and sensible risk management. You build fluency with the tools while staying in territory where an error is correctable. It also means building the skills that hold up regardless of what the policy says. AI skills for non-technical people aren’t about knowing which tool is best. They’re about understanding what AI is actually good at, what it consistently gets wrong, and how to use it in a way that adds value instead of just adding steps. That knowledge doesn’t expire when your company updates its policy. The people who will be in the worst position when policy eventually clarifies are the ones who either refused to use AI at all (now behind on skills and output speed) or used it carelessly without thinking about documentation, data, or disclosure. The people who will be fine are the ones who treated it seriously, documented what they did, and can show they were thoughtful about it. That’s not a heroic stance. It’s just basic professionalism applied to a new kind of tool. Frequently asked questions Does my company need an AI policy in 2026? Yes, and most don’t have a good one. The White House’s March 2026 framework pushed AI governance responsibility to organizations rather than creating binding federal rules, which means companies are largely on their own. If yours doesn’t have a policy, that’s a real gap, and in the meantime, employees are operating on guesswork. What should I do if my company has no AI policy? Find out whether one exists first by checking your employee handbook or asking HR directly. If there’s nothing, use your judgment: avoid putting confidential or client data into external AI tools, document what you’re using AI for and how you verify outputs, and consider having a low-stakes conversation with your manager about how they view AI tool use. See our guide on whether to tell your boss you use AI for a fuller breakdown. Is it against the rules to use ChatGPT at work? It depends entirely on your employer. Some companies explicitly ban external AI tools. Some have approved lists. Many have nothing written down either way. The 68% stat from Microsoft’s Work Trend Index suggests most employees using AI at work aren’t disclosing it, often because the rules are unclear. Read whatever policy exists, and if it’s silent on AI, ask. What did the White House AI policy framework say in 2026? The National AI Policy Framework released March 20, 2026 laid out guidelines for responsible AI use but didn’t create binding regulations for private employers. It emphasized pushing governance responsibilities to organizations and signaled federal interest in preempting state-level AI laws. For most workers, it changed nothing immediately. Companies still have to write their own rules. How do I protect myself if my company's AI policy is vague? Document your AI use. Keep a running note of what tools you used, what you used them for, what you verified before acting on outputs, and what the result was. This gives you a clear record if questions ever arise. Avoid putting sensitive, confidential, or client-identifiable data into external AI tools regardless of what the policy says. What's the difference between a company "no AI" policy and how people actually behave? Often, a significant one. Research consistently shows that employees use AI even when policies technically prohibit it, partly because enforcement is unclear and partly because managers use it too. The risk is that when something goes wrong, the written policy is what gets enforced, not the informal norm. If your company has a written “no AI” policy, take it seriously even if the culture seems to ignore it. You read the whole thing. Don't Replace Me A Survival Guide to the AI Apocalypse. 200+ pages covering what's actually changing, what isn't, and what to do about it. Get the Book → Dmitry Kargaev AI engineer, product designer, and author of Don't Replace Me — A Survival Guide to the AI Apocalypse. He writes about working alongside AI without being replaced by it. More at deeflect.com. Keep reading Should You Tell Your Boss You Use AI? 68% of workers hide their AI use from employers. Here's the honest answer on when to disclose, when to stay quiet, and how to handle it. AI at workworkplace AIAI disclosure How to Use AI at Work: A No-BS Starter Guide Use AI productively without creating unnecessary risk. Practical workflows, guardrails, and no hype. AI workflowproductivityworkplace AI Anxiety Is Real — Here's How to Stop Spiraling If policy uncertainty is fueling stress, this gives you a practical plan to move from rumination to action. ai anxietycareer anxietyfuture of work