Guide

How to write a PRD with AI

Ask an AI for "a PRD" and you get a generic template filled with confident guesses. This guide gives you a prompt built around your real inputs, a review checklist that catches what AI drafts usually get wrong, and a way to turn the result into tasks.

Updated 18 October 2026

Quick answer: write a short brief first (the problem, who has it, evidence, goals and constraints), give it to ChatGPT, Claude or Gemini with the prompt below, and make the AI ask its open questions before it drafts. Then review the draft with the checklist, fix the gaps section by section, and ask for the requirements as a numbered task list.

Product details below were checked against OpenAI's, Anthropic's and Google's help pages as of October 2026. The prompt works in any of the three.

Step 1: Write a five-part brief

A PRD is only as good as what the AI knows. OpenAI's prompting guidance stresses precise requests with plenty of background, and for a PRD that's essential: without facts, the AI fills the gaps with plausible inventions. Before opening a chat, jot down:

The problem in one or two sentences, from the user's side.
Who has it: the user type, roughly how many, and the situation they're in.
Evidence: support tickets, interview notes, sales feedback or usage numbers you actually have. Paste or attach them rather than summarizing.
Goals and non-goals: what success looks like and what this release deliberately won't do.
Constraints: deadline, team size, platforms, legal or technical limits, and anything already decided.

If you write PRDs often, set this up once. In ChatGPT, select New project in the sidebar, then add files and use the ••• menu and Project settings to add instructions. In Claude, open Projects, click + New Project, use Set project instructions, and add files to the project knowledge with the + button. Put your product overview, your PRD template and one PRD your team considers good in there.

Step 2: Run the PRD prompt

Paste your brief and evidence into a new chat (or attach the files; Claude, for example, takes up to 20 files per chat), then send:

"You're helping me write a product requirements document for [feature] in [product]. Use only the brief and material above; don't invent users, numbers or decisions. First, list up to 8 questions whose answers would change the requirements, and wait for my answers. Then write the PRD with these sections: problem, target users, evidence, goals with measurable success metrics, non-goals, user stories, functional requirements numbered R1, R2… each with acceptance criteria, edge cases and error states, dependencies and risks, open questions. Mark anything you assumed with [ASSUMPTION]. Keep it under 1,500 words."

Why it's built this way:

Questions before drafting pull out decisions you hadn't made yet, which is half the value of writing a PRD.
[ASSUMPTION] tags turn hidden guesses into a visible list you can confirm or correct.
Numbered requirements with acceptance criteria give engineering and QA something testable, and make Step 4 easy.
A word limit keeps the AI from padding. A PRD people read beats a long one they skim.

Prefer editing in a document view? In Gemini on the web, click Add Files below the text box and choose Canvas before sending the prompt. You can edit the text directly, ask for changes to a selected passage, and when it's ready use Share & export and Export to Docs to get a Google Doc.

Step 3: Review the draft with this checklist

Read the draft as a skeptical engineer would. OpenAI's guidance treats prompting as iterative, so fix each problem with a targeted follow-up rather than starting again.

Every [ASSUMPTION] is resolved. Confirm it, correct it or move it to open questions.
The problem is backed by your evidence, not by general claims about "users". Ask: "Quote the evidence behind each statement in the problem section."
Metrics have a number and a timeframe you can actually measure with your current tools.
Non-goals are real. At least one thing someone will ask for should be on the list.
Each requirement is testable. "Fast" and "intuitive" aren't; "loads in under two seconds on the slowest supported device" is.
Edge cases cover the boring states: empty, loading, error, offline, permissions, very large inputs.
Nothing contradicts a constraint from your brief, such as a platform you said you won't support.
Open questions each have an owner.

A useful follow-up: "Act as the engineering lead reading this PRD. List the ten questions you'd ask before estimating, and point to the requirement each one is about." Answer them in the chat and ask for an updated version of only the affected sections.

Step 4: Turn the PRD into tasks

Ask for a task breakdown: "Break requirements R1 to Rn into tasks. For each: task, the requirement it serves, type (design, frontend, backend, QA, content), dependencies, and a size of S, M or L. Flag tasks that block others."
Check coverage. Every requirement should have at least one task, and every task should point to a requirement. Orphans on either side mean something was missed or invented.
Let the team size the work. The AI's S, M, L sizes are a starting point for discussion, not an estimate.
Move the tasks into your tracker and link back to the PRD, so the requirements stay the reference when questions come up.

For the general habit of turning AI output into a to-do list, see how to turn AI answers into tasks, and for prompt patterns that produce actionable answers, see prompts for actionable answers.

Good to know

Projects on free plans. Claude's free plan allows up to five projects. ChatGPT projects are available to signed-in users, with file limits that depend on the plan. In ChatGPT, instructions you set on a project take priority over your custom instructions, and only for chats in that project.
Projects keep chats separate. In Claude, chats in the same project don't share context unless the material is in the project knowledge, so add your final PRD there if later chats should use it.
Privacy. PRDs often contain unreleased plans. On personal ChatGPT accounts, conversations can be used to train models while Improve the model for everyone is on (in Settings → Data controls); Business, Enterprise and Edu workspace content isn't used for training by default. Claude Free, Pro and Max users choose in Privacy Settings whether chats are used for training; Team and Enterprise are excluded. Follow your company's policy on which account to use.

If it doesn't work

The draft is generic. The brief was too thin. Paste the raw evidence (tickets, notes) instead of your summary, and add one example PRD you like.
It skips the questions and drafts straight away. Send the questions request on its own first, then the drafting instructions in a second message.
Numbers appear that you never gave it. Reply: "Remove every number that isn't in my material, and mark the metric as [NEEDS TARGET]."
Later edits undo earlier fixes. Ask for changes to one named section at a time, or edit the text directly in a Gemini Canvas or your own document.

Keep the review and tasks in one place

In Workbranch, answers come from Google Gemini and each point of an answer becomes a checklist item with progress shown, so a review checklist or task list can be ticked off where it was written. You can open a side thread on a single requirement without derailing the main chat, and flag the chat with a due date that shows on the calendar. There's a 3-day free trial.

Sources

Prompt engineering best practices for ChatGPT (OpenAI Help Center, checked 18 October 2026)
Projects in ChatGPT (OpenAI Help Center, checked 18 October 2026)
Data controls in ChatGPT (OpenAI Help Center, checked 18 October 2026)
How can I create and manage projects? (Claude Help Center, checked 18 October 2026)
What are projects? (Claude Help Center, checked 18 October 2026)
Upload files to Claude (Claude Help Center, checked 18 October 2026)
Create docs, apps & more with Canvas (Gemini Apps Help, checked 18 October 2026)
Updates to Consumer Terms and Privacy Policy (Anthropic, checked 18 October 2026)

Frequently asked questions

Should the AI write the whole PRD in one go?

Usually not. Asking for open questions first, then the draft, then one section at a time for fixes gives you a document you've actually thought through. A single long answer tends to hide weak sections under polished wording.

Which sections can I trust the AI with most?

It's strongest at structure and coverage: listing edge cases, spotting missing states such as empty, error and loading, and phrasing acceptance criteria consistently. Goals, success numbers and scope decisions should come from you, because the AI has no way of knowing your business.

Can I reuse the same setup for the next PRD?

Yes. Keep your product overview, an example of a PRD your team liked and your house template in a ChatGPT or Claude project, so every new PRD starts with the same context and only the feature brief changes.