PM Mapped
Working with AI WRITING GREAT SPECIFICATIONS · The PRD

The PRD with AI

The PRD is one of the highest-value places a PM can use AI: it's a long, structured document AI drafts well from your direction. But a PRD captures real decisions — what to build, what to cut, and why — and a fluent PRD full of wrong calls is worse than a rough one with right ones.

← Back to the The PRD tool
1How to use AI for this

A step-by-step way to work

1

Feed it the decisions, not just the topic

Give AI the problem, the goals, the constraints, and the calls you've already made. It assembles; it shouldn't decide.

2

Draft the skeleton

Ask for a structured PRD draft with the sections you use, then correct and sharpen each.

3

Hunt for gaps and contradictions

Ask 'what's underspecified, contradictory, or missing an edge case here?' — a real strength.

4

Generate the exec summary

Have it compress the full PRD into a tight summary for leadership (see the Executive Briefing guide).

5

Own every decision in it

Read it as if you'll defend each line, because you will. AI's confident phrasing can hide a call you didn't actually make.

2Worked examples

Real prompts, and what good output looks like

Finding gaps in your own PRD
Your prompt

Here's my draft PRD for a 'saved searches' feature [pasted]. What's underspecified or missing that engineering will ask about?

What good output looks like

No limit stated on saved searches per user — eng will need a number (and what happens at the limit).

Unclear whether saved searches sync across devices or are per-device.

No behaviour defined for when the underlying data a search points to is deleted.

Success metric is 'engagement' — too vague to instrument; specify the event you'll measure.

These are the questions that otherwise surface mid-sprint and stall it.

This is the payoff: AI catches the unspecified decisions before the sprint does. Each gap it finds is a decision you now make — it just found the hole.
Compressing to an exec summary
Your prompt

Summarise this PRD into a 5-sentence exec summary: the problem, the proposed solution, the expected impact, the main risk, and what you need from leadership.

What good output looks like

Problem: power users lose time re-running the same searches daily. Solution: let them save and re-run searches in one click. Expected impact: faster task completion for our most active 15%, our highest-retention segment. Main risk: low usage if saved searches aren't surfaced prominently. Ask: sign-off to prioritise this in the next cycle over [the alternative].

Useful, with a check: the numbers ('15%', the retention claim) must be ones you actually have — AI will invent plausible figures if you let it.
3Copy-paste template

A prompt you can reuse

Fill in the highlighted parts and paste it into your AI tool of choice. Edit the output — it's a starting point, not a finished answer.

Reusable prompt
Help me draft a PRD. I'll give you the decisions; you assemble and pressure-test — don't invent facts or figures.

Inputs:
- Problem: [the user/business problem]
- Goal & success metric: [what success looks like, and the metric]
- Proposed solution (high level): [what we'll build]
- Constraints & non-goals: [what's out of scope, known limits]
- Decisions already made: [anything settled]

Please:
1. Draft a structured PRD (background, problem, goals/metrics, solution, scope & non-goals, risks, open questions).
2. Flag everything that's underspecified or contradictory, as 'open questions' for me to resolve.
3. Do NOT invent metrics, numbers, or user research — mark those as [TBD] for me to fill.
4Common pitfalls

What AI gets wrong here

Invented facts and metrics

AI fills gaps with plausible-sounding numbers, research, and impact claims that aren't real.

Do this instead: Force [TBD] placeholders for anything you haven't measured; never ship an AI-supplied statistic you can't source.

Decisions smuggled into prose

Fluent drafting can state a scope or design call as settled when you never made it.

Do this instead: Read every line as a decision you're committing to; strike anything you didn't actually decide.

Tidy but generic

AI produces a competent, generic PRD that doesn't reflect your product's specifics.

Do this instead: Inject the real constraints and context; reject sections that could belong to any product.
The judgment that stays yours

AI can assemble and polish a PRD faster than anyone, and it's genuinely good at finding the gaps you'd have discovered painfully mid-sprint. But the document's job is to record decisions — and the decisions, the real metrics, and the willingness to defend them are yours. Let AI build the scaffold; never let it make the calls or invent the evidence.