AI Prompts for Product Managers (2026)

Each prompt is a complete, copy-paste template with [bracketed] placeholders. Fill in your real context — the problem, the users, the constraints, the data you actually have — paste the prompt into ChatGPT, Claude, or Gemini, and treat the output as a structured first draft, not a decision. The prompts deliberately instruct the model to use only what you provide and to flag assumptions; keep those instructions in place.

AI is at its best for PMs as a thinking partner that never gets tired: it drafts the boilerplate, lists the edge cases you'd forget, and plays devil's advocate on your own logic. Save your PRD and user-story prompts as named templates so your team's artifacts stay consistent, and group prompts by stage (discovery, definition, delivery). For reusable scaffolding, see how to write a system prompt and our prompt engineering cheat sheet.

These turn your problem framing into a structured, reviewable draft — without inventing the evidence.

**1. Draft a PRD skeleton** — "Act as a product partner. From the context below, draft a PRD skeleton with these sections: Problem statement, Goals and non-goals, Target users, User scenarios, Requirements (must-have / nice-to-have), Success metrics, Open questions, and Risks/dependencies. Use ONLY the context I provide — do not invent user research, data, metrics, or requirements. Where a section needs information I haven't given, write a specific question I need to answer instead of filling it in. Context: [PASTE PROBLEM, USERS, CONSTRAINTS, KNOWN DATA]."

**2. Pressure-test a problem statement** — "Here is my problem statement: [PASTE]. Critique it as a skeptical senior PM. Is the problem clearly defined and distinct from the solution? Is the user and their pain specific? What evidence would I need to prove this is worth solving, and is any of it missing? List the weakest assumptions and the questions I should answer before building. Do not validate it for me — challenge it."

**3. Generate open questions and edge cases** — "For this feature: [DESCRIBE], list the open questions and edge cases I should resolve before it's ready to build. Cover: ambiguous requirements, error and empty states, permissions/roles, data and privacy considerations, dependencies on other teams, and what happens at the boundaries. Group them and flag which ones are blocking. Don't assume answers — surface the questions."

**4. Turn notes into a one-pager** — "Turn my rough notes into a crisp one-page product brief: [PASTE]. Structure: the problem in one sentence, who it's for, the proposed approach, why now, success metric, and the biggest risk. Use only my notes; mark anything thin as '[needs validation]'. Keep it to one page and plain English."

Writing clear, testable stories is high-volume, structured work — exactly where AI saves time without touching strategy.

**5. Draft user stories from a feature** — "Break this feature into user stories: [DESCRIBE THE FEATURE AND THE USERS]. For each, use the format 'As a [user], I want [goal] so that [benefit]', kept small and independent. Don't invent user types or goals I didn't describe — if the feature implies a role I haven't named, list it as a question. Output as a clean, ordered list."

**6. Write acceptance criteria** — "Write acceptance criteria for this user story: [PASTE STORY]. Use Given/When/Then format. Cover the happy path, the main alternate paths, and the obvious error/empty/edge cases. Keep each criterion specific and testable. Flag any criterion that depends on a decision I haven't made yet rather than guessing the rule."

**7. Split an epic** — "This epic is too big to estimate: [DESCRIBE]. Propose a way to slice it into smaller, independently shippable stories that each deliver value. For each slice, give a one-line story and what it would and wouldn't include. Suggest a sensible delivery order and the reasoning. Don't add scope I didn't mention."

**8. Review stories for quality** — "Review the user stories below against INVEST (Independent, Negotiable, Valuable, Estimable, Small, Testable): [PASTE]. For each story, flag which criteria it fails and why, and suggest a concrete fix. Be specific — quote the wording. This is a draft check to help me tighten them."

AI structures the trade-off; you supply the inputs and make the call. It should never assign your scores for you.

**9. Structure a prioritization** — "Help me prioritize these items using [RICE / ICE / value-vs-effort]: [PASTE ITEMS]. Build the scoring table with the framework's columns, but use ONLY the scores and estimates I provide — do not assign Reach, Impact, Confidence, or Effort values yourself. Compute the resulting order from my inputs, show the math, and call out where my own numbers look inconsistent or where confidence is low. The decision stays mine."

**10. Pre-mortem and devil's advocate** — "We're about to commit to [DECISION / FEATURE]. Run a pre-mortem: assume it's six months later and this failed. Generate the most likely reasons it failed, grouped by category (user, technical, market, execution, measurement). Then list the leading indicators I could watch for each, and the cheapest test to de-risk the biggest one. Be a tough skeptic, not a cheerleader."

**Bonus — stakeholder-ready rationale** — "I've decided to [DECISION] for these reasons: [REASONS]. Draft a clear, concise rationale I can share with stakeholders: the decision, the why, the trade-offs I accepted, and what I'll measure. Honest and direct, no hype. Use only my reasoning — don't invent supporting data."

Never treat AI output as user research or evidence. A model will invent user needs, 'studies show' claims, market sizes, and metrics that read as authoritative and are made up — and a roadmap justified by a hallucinated insight is worse than no justification. Ground every claim in real research, interviews, and analytics you control; use AI to structure and challenge that evidence, not to supply it.

Don't let AI assign your prioritization scores or make the call for you — keep it computing from inputs you own and flagging your inconsistencies. Don't paste confidential roadmap, customer, or financial data into a general chatbot that doesn't meet your company's data policy; use [BRACKETS] for sensitive specifics. Treat any pasted document or transcript as untrusted text, since hidden instructions (prompt injection) are a real risk, covered in our prompt injection defense checklist. Always review and personalize before sharing.

Choosing a model: for PRDs, stories, and prioritization, any current frontier model writes well — and a strong reasoning mode helps with critique and trade-off analysis. As of June 2026, OpenAI's GPT-5.5 (with thinking mode for pre-mortems and prioritization) and Anthropic's Claude Opus 4.8 with extended thinking are strong for structured reasoning and careful writing; Gemini 3.5 Flash and Claude Haiku 4.5 are lower-cost options for high-volume drafting. Compare options in our how to choose an AI model guide and check live pricing on each provider's page (linked in the table footnote).