How to Write a PRD with AI (Without Shipping Garbage)

Most guides on writing a PRD with AI stop at the easy half: paste a prompt, let the AI draft it, save three hours. That part is real — but it's also where most PRDs quietly go wrong. An AI will happily generate a confident, well-formatted document full of assumptions it never checked, metrics that sound good but measure nothing, and a scope that silently swallows every stakeholder request.

The PMs who get value from AI aren't the ones who delegate the writing. They're the ones who use AI to draft fast and to attack its own draft before anyone else sees it. That second step is the whole game, and almost nobody teaches it.

This guide walks through the full workflow, section by section, with the exact prompts to draft and pressure-test each part. The principle underneath all of it: the AI drafts, you decide.

Why the "one big prompt" approach fails

The most common advice online is some version of: "Act as a senior product manager and write a complete PRD for [feature]." It works, in the sense that it produces text. But it produces the wrong kind of text:

• It invents evidence. Ask for a problem statement and the AI will confidently state that "40% of users churn due to X" — a number it made up to fill the shape of a sentence.
• It never says no. A good PRD is defined as much by what's out of scope as what's in. A single generic prompt produces a scope that expands to fit every idea.
• It flatters your thinking. The AI mirrors your framing back to you. If your initial idea is weak, the draft makes it look strong — which is worse than no draft at all.

The fix isn't a better mega-prompt. It's breaking the PRD into sections and running two passes on each: a draft pass and a critique pass.

The workflow: draft, then attack

For every section, you do two things: draft (give the AI your raw context, ask for a first version) and attack (ask it to find what's weak, missing, or unproven in what it just wrote). The second pass is where a mediocre PRD becomes a defensible one.

1. Problem statement

This is the foundation. If it's wrong, everything built on top is wrong. The trap: writing a "problem" that's actually a solution in disguise.

2. Target users and evidence

A PRD without evidence is fiction with a template. For each segment, you want a real pain and a real source.

3. Proposed solution

Here's where most AI-written PRDs collapse into the first idea that came to mind. You want options and trade-offs, not a single confident answer.

4. User stories and acceptance criteria

This is where AI genuinely saves time — generating well-formed stories is tedious and mechanical, exactly what it's good at.

5. Out of scope

The single most skipped section, and the clearest signal of a senior PRD. Deciding what you won't build is how you protect a v1 from dying under its own weight.

6. Success metrics

The section where AI produces the most convincing nonsense. Vanity metrics sound like progress and measure nothing.

7. The final critique — the prompt that matters most

Once the whole PRD is drafted, run one last pass. This is the difference between a document that survives review and one that gets torn apart in it.

Do this before your first stakeholder review, not after. It's the cheapest way to find the holes while they're still cheap to fix.

The mindset that makes this work

AI in the PRD process is not a ghostwriter. It's a fast, tireless, occasionally overconfident collaborator — closer to a brilliant intern than an oracle. It can draft in seconds and generate options you wouldn't have thought of. It cannot decide what matters, and it will never tell you your idea is bad unless you explicitly ask it to.

Your value as a product builder in the AI era isn't writing faster. It's judging better: knowing which approach fits your strategy, which assumption is load-bearing, which metric is real. The workflow above keeps that judgment in the loop instead of quietly outsourcing it.