How to Write a PRD with AI: A Prompt Framework and Review Loop
AI turns the PM from writer into editor. Here is the prompt framework, the copy-paste PRD prompt and the review loop that keep an AI-drafted PRD honest.
Share
To write a PRD with AI, write the problem, evidence and success metric yourself, then give the model that context with a clear task and hard constraints, and let it draft the structure, user stories, edge cases and acceptance criteria. Your job shifts from writer to editor: you check every claim, cut what is unsourced and decide the trade-offs. The draft is version 0.5, not version 1.
Key takeaways
PRD writing is one of the most common AI uses in product work. In Productboard's 2025 survey of 379 product professionals, 94% used AI daily or often, and writing PRDs was among the top time-savers (Productboard, 2025).
Models still invent. Even when only asked to summarise a supplied document, leading models hallucinate at rates from 1.8% to 24.2% on Vectara's leaderboard (Vectara, 2026). A PRD prompt with thin context invites far more.
Researchers trace hallucination to training and evaluation that "reward guessing over acknowledging uncertainty" (Kalai et al., 2025). So ask the model to mark gaps, not fill them.
The Context-Task-Constraint framework below, plus your evidence, does more for output quality than any choice of model.
Review is the job now: every claim in the problem and success sections needs a source you can open.
What changes when you write a PRD with AI?
The PRD's purpose does not change. Marty Cagan defined it as articulating the product's purpose, features, functionality and behaviour clearly enough that the team can build and test it (SVPG, 2005). What changes is where your time goes.
Product professionals spent about 18 hours a month on articulating and prioritising requirements in Pragmatic Institute's 2019 survey of 2,474 people (Pragmatic Institute, 2019). Much of that was drafting and formatting. With AI, drafting takes minutes; the remaining hours should go to the three things a model cannot do for you:
Validation. Is the problem real, and is the evidence strong enough to act on?
Strategic alignment. Does this feature serve the goal the company actually set?
Trade-off analysis. What are you choosing not to build, and what risk are you accepting?
Atlassian's State of Product 2026 survey of more than 1,000 product professionals found AI saving teams 10 to 60 minutes a day, mostly on documentation and research, while respondents said it did not help with complex work such as planning (Mind the Product, 2025). That is the right split. Hand the model the paperwork; keep the judgement.
Can AI write a good PRD on its own?
No. AI is a synthesis engine, not a discovery engine. It can turn your notes, transcripts and constraints into a clean, consistent structure. It cannot know which customer problem matters most, what your baseline is, or which trade-off your team accepted last week.
The failure mode is not an obviously bad document. It is a fluent one. Stack Overflow's 2025 survey found 66% of developers frustrated by AI output that is "almost right, but not quite", and more developers distrusted AI accuracy (46%) than trusted it (33%) (Stack Overflow, 2025). PRDs have the same problem, with a twist: a code bug fails a test, while an invented customer insight passes review unless someone checks the source.
Treat the AI draft as version 0.5. It gives you structure and coverage. You supply the evidence, the decisions and the voice of the team.
How do you write a PRD with AI step by step?
Gather the evidence first. Collect the 3-10 sources that bear on this decision: interview excerpts, support tickets, a usage query, the relevant OKR. Paste or attach them; do not ask the model to "imagine" users.
Write the problem and success metric yourself. Two or three sentences each. This is the part that decides whether the PRD is right, so it should be in your words.
Prompt with Context, Task and Constraints. Use the framework below. State the format, the scope limit and the rule that every claim must cite a supplied source.
Generate section by section. Ask for the modular parts one at a time: objective, personas, success metrics, user stories, acceptance criteria, non-functional requirements, risks. Smaller asks produce tighter output and are easier to review.
Ask for gaps, not just answers. Prompt: "List the questions an engineering lead would ask that this PRD cannot answer." Turn the answers into open questions with owners.
Trace every claim. Mark each factual sentence as sourced, inferred, assumed or invented, and delete the invented ones. The 15-minute claim trace walks through it.
Review with engineering and design, then version it. Store one canonical copy and link it from tickets. For the full manual method behind these steps, see how to write a PRD.
What is the Context-Task-Constraint prompt framework?
Context-Task-Constraint (CTC) is a three-part prompt structure that makes the model work inside your product's reality rather than produce a generic PRD.
Context: who the users are, what the product does, the problem, and the evidence. Anthropic's prompting guide suggests thinking of the model as "a brilliant but new employee" with no context on your norms, and adding the motivation behind instructions so it can generalise (Anthropic).
Task: the specific output you want, one section or one document, in a named format.
Constraints: scope limits, syntax (for example Given/When/Then), length, and grounding rules such as "if a claim is not supported by the sources, write OPEN QUESTION instead".
Three habits make CTC prompts work better:
Separate the parts. Anthropic recommends wrapping instructions, context and input in their own XML tags, such as <context> and <sources>, to reduce misinterpretation.
Show an example. The same guide recommends 3-5 well-chosen examples. One good past PRD from your team teaches tone and depth faster than any description.
Run the colleague test. If a teammate with no background would be confused by your prompt, the model will be too.
Here is the difference in practice:
Weak prompt
CTC prompt
"Write a PRD for automated leave approvals."
Context: B2B HR platform; admins at 50-500-person customers approve leave manually; 3 linked interview excerpts and 1 ticket query. Task: draft user stories and acceptance criteria. Constraints: MVP scope only, Given/When/Then, cite a source for every claim, list open questions.
Produces a plausible, generic PRD with invented personas and metrics
Produces a scoped draft you can verify line by line
Copy-paste AI PRD prompt
Use this with any capable model, whether ChatGPT, Claude or Gemini. Fill the brackets, paste your sources, and run it one section at a time if the output is long.
<role>
Act as a senior product manager writing for an engineering team.
</role>
<context>
Product: [what it does, for whom]
Problem (written by me): [2-3 sentences]
Success metric (written by me): [metric, baseline -> target, check date]
Strategic goal / OKR: [goal]
Known constraints: [tech, legal, time budget]
</context>
<sources>
[Paste interview excerpts, ticket summaries, data. Label each S1, S2, S3...]
</sources>
<task>
Draft a PRD for [feature name] with these sections:
1. Executive summary (5 lines)
2. Target persona(s), based only on the sources
3. Success metrics: use mine; suggest at most 2 guardrails
4. User stories, ranked, with Must / High-want / Nice priority
5. Acceptance criteria in Given/When/Then form
6. Non-functional requirements: latency, security, privacy, accessibility
7. Risks and dependencies
8. Out of scope
9. Open questions
</task>
<constraints>
- MVP scope only; anything beyond it goes in Out of scope.
- Every factual claim about users or the market must cite a source ID (S1...).
- If a claim is not supported by the sources, write OPEN QUESTION instead.
- Do not invent numbers, quotes or customer names.
- Plain, concise language. No marketing copy.
</constraints>
Follow-up prompts that sharpen the draft
"Act as the engineering lead. List the 10 questions you would ask before estimating this."
"List empty, loading, error and permission states the acceptance criteria do not cover."
"Which requirements do not move the success metric? Suggest what to cut."
"Rewrite every acceptance criterion so a tester can verify it without asking me."
"List every sentence in the problem section that is not directly supported by S1-S5."
Worked example: an AI-drafted PRD for leave approvals
This example is illustrative, not a real company. Say a B2B HR platform hears that admins approve leave manually.
What the PM writes: "Admins at customers with 50-500 employees approve each leave request by hand, often days late. In 4 of 7 recent interviews, admins said they approve in batches on Fridays. Success: median approval time falls from 3 days to under 1 day within 60 days of launch, with no rise in disputed approvals."
What the PM pastes as sources: the 4 interview excerpts (S1-S4), a query showing 1,200 leave requests last month (S5), and the current approval screen.
What the AI drafts well: a ranked list of 6 user stories, 14 acceptance criteria in Given/When/Then form, edge cases such as overlapping requests and approvers on leave, and non-functional requirements for audit logging.
What the review catches:
A persona described as "busy HR managers juggling 20 tools" with no source. Deleted.
A suggested guardrail, "keep NPS above 40", with no baseline anywhere. Replaced with an open question.
An auto-approval rule for requests under 2 days that nobody had agreed. Moved to Out of scope pending a decision.
The draft took minutes; the review took about an hour and removed three claims that would otherwise have shaped the build. That ratio is the point.
Which AI tool should you use to write a PRD?
Use whichever model your company allows and you know well. Productboard found 88% of product professionals already use two or more AI models, while only 65% said their company had a documented AI policy (Productboard, 2025). Check the policy before pasting customer data into any tool.
Option
Good for
Watch out for
General chat model (ChatGPT, Claude, Gemini)
Flexible drafting with the CTC prompt
You must paste the evidence every time; no memory of past decisions
Free PRD generator
A fast first draft from a feature idea
Output is only as good as the problem you type in
Docs or wiki AI (Notion, Confluence)
Drafting where the PRD will live
Usually cannot see interviews or tickets in other tools
Evidence-connected product tool
Drafts that cite the underlying calls, tickets and feedback
Needs your sources connected first
To try the first steps without setup, use the free PRD generator for a structured draft, or feedback to spec to turn one piece of customer feedback into a problem statement. If the feature you are specifying is itself an AI feature, read the AI PRD guide for model behaviour, guardrails and evaluation sections.
Common mistakes when writing a PRD with AI
Prompting with a feature name only. The model fills the gap with what is statistically typical, which is rarely what your users need.
Letting AI write the problem statement. It will sound right and may be wrong. Write it yourself.
Accepting invented metrics. If the draft proposes a target, ask where the baseline came from. No baseline, no target.
One giant prompt. Long, multi-section outputs drift. Generate and review section by section.
Skipping the engineering read. An AI draft can look complete and still miss the dependency your tech lead knows about.
Pasting confidential data into unapproved tools. Check your AI policy and the tool's data terms first.
Where Specky fits
Specky drafts PRDs from the customer calls, feedback, Slack threads and tickets already in its Product Graph, so the context in step 1 is assembled for you rather than pasted. The PRD Editor shows an inline citation next to each claim, which makes the review in step 6 a matter of clicking through, and drafts land in the PM Inbox for a human to approve.
Frequently asked questions
What is the best prompt to write a PRD with AI?
The best prompt gives the model context (product, users, problem, evidence), a specific task (which sections, which format) and constraints (MVP scope, cite every claim, mark gaps as open questions). Write the problem and success metric yourself and paste them in. The copy-paste prompt above follows that structure and works across major models.
Is it safe to use ChatGPT or Claude for PRDs?
It depends on your company's AI policy and the tool's data terms, not the model. Productboard's 2025 survey found only 65% of companies had a documented AI policy. Remove personal data from transcripts, use business or enterprise plans where required, and never paste customer data into a tool your company has not approved.
How do you stop AI from making things up in a PRD?
Supply your sources, label them, and require a source ID for every factual claim. Instruct the model to write OPEN QUESTION when the sources do not support a statement. Then trace each claim yourself before engineering sees the draft. Models are trained in ways that reward confident guessing, so the instruction to abstain matters.
Should the AI write the user stories or the PM?
Let the AI draft them from an agreed problem and persona, then edit. AI is good at coverage: it lists stories, states and edge cases you might miss. The PM decides ranking, merges duplicates and cuts stories that do not move the success metric. Review acceptance criteria hardest, because engineers build to them.
Do I still need a PRD template if I use AI?
Yes. A template is what makes AI output consistent across features and reviewers. Put your team's sections in the prompt's task block, or start from the free PRD template. A consistent structure also helps engineers, and AI coding agents, find scope and acceptance criteria.
GDPR-aligned. EU workspace data is stored in Frankfurt, Germany. GDPR and EU data Built in Vienna, Austria. A product of influence solutions fm GmbH (FN 675374y). Imprint 14-day trial. Card required, and nothing is charged until day 14. 30-day refund after subscribing. See pricing