Draft the PRD with AI — and Keep the Judgment Sections

Workflow area: Specs & Docs · Tier: Core · Hands-on time: ~40 minutes

The PRD is where product-management theater is easiest to manufacture: a model will produce a complete, plausible, well-formatted PRD from a two-line prompt, and it will be judgment-free all the way through. The workflow that survives contact with good engineering teams splits the document by its nature. Scaffolding sections — user stories, requirements tables, edge cases, rollout mechanics — are structure applied to evidence, and AI drafts them well from a real evidence pack. Judgment sections — which problem, why this quarter, what we're refusing to build — are the decisions you're paid for, and delegating them produces documents that read fine and steer nothing. Draft with AI, decide by hand, then use AI again for what it's genuinely best at: attacking the finished draft for contradictions and holes.

When to use — and when not to

• Use it when a real evidence pack existssynthesis docs, counted themes (pb-17), a metrics snapshot, constraints. The draft inherits the pack's quality directly.
• Use it for medium-and-up featuresanything that needs a document engineering will interrogate. The workflow's overhead pays off exactly where scrutiny is coming.
• Use the consistency pass on every PRD, even hand-written onesthe adversarial read is the cheapest edge-case discovery you'll ever run (pb-07 goes deeper).
• Skip it with no evidence packa PRD generated from vibes is vibes with headings. Do the discovery first; a beautiful document about an unvalidated problem is the trap, not the win.
• Skip full delegation alwaysa PRD whose problem statement you can't defend without re-reading it is not your PRD, and the first sharp question in review will prove it.

The recipe

1. 1.

   Assemble the evidence pack. The synthesis doc with citations (pb-02), the counted themes if relevant (pb-17), the metrics snapshot with dates, the constraints memo (technical, legal, capacity), and your one-paragraph hypothesis of the bet. If two data points conflict, leave them both in — surfacing that conflict is the draft's job, and smoothing it over is the model's known failure.

2. 2.

   Generate the structured draft with judgment sections stubbed. In ChatPRD, a Claude Project holding your team's template, or Notion AI:

   Copy

   Draft a PRD from the evidence pack below, using our template.
   
   Rules:
   - Every claim must trace to the pack; cite the source doc inline
     like [synthesis §3] or [metrics 2026-07].
   - Where evidence conflicts, surface the conflict in Open Questions
     — do NOT reconcile it yourself.
   - For these sections, output only a stub reading "[PM WRITES THIS]"
     plus 3 bullet points of relevant evidence: Problem Statement,
     Why Now, Non-Goals, Success Metrics rationale.
   - Mark anything you inferred beyond the pack with [INFERRED].
   
   Sections to draft fully: user stories, functional requirements,
   edge cases and error states, dependencies, rollout considerations,
   analytics events, open questions.
   
   Evidence pack:
   [PASTE PACK]
   

3. 3.

   Hand-write the judgment sections. No AI in this step — that's the point of it. The problem statement in your own words, sharp enough that a teammate could repeat it in the hallway. Why now — the real reason this quarter beats next quarter, including the uncomfortable parts. Non-goals — what you are explicitly refusing to build and why; this section has no evidence to summarize, only decisions to make, which is why a model's version of it is always filler. Success metrics — AI can list candidates, but which number you'll stake the feature's continuation on is a commitment, not a completion.

4. 4.

   Run the adversarial consistency pass on the merged draft. Fresh conversation, full document in:

   Copy

   Review this PRD as a skeptical senior engineer seeing it cold.
   
   1. CONTRADICTIONS: any place two sections disagree (scope vs
      stories, metrics vs problem, non-goals vs requirements) —
      quote both lines.
   2. UNTRACEABLE CLAIMS: statements presented as fact with no
      source in the doc.
   3. MISSING EDGE CASES: states, permissions, failure paths, and
      scale conditions the requirements don't cover.
   4. AMBIGUITIES: any requirement two engineers could implement
      differently — show the two readings.
   5. THE HARD QUESTION: the one question you'd ask in review that
      this doc can't currently answer.
   

   Fix what's real, log the rest as open questions. Findings like "non-goals excludes bulk edit, but story 7 implies it" are this pass's bread and butter — and exactly what you'd otherwise discover in review, publicly.

5. 5.

   Do the two-minute honesty retro. At the bottom of your working doc, three lines: what AI drafted, what you decided, what the consistency pass caught. This isn't ceremony — it's calibration data. After five PRDs you'll know which sections you can trust the draft on and which it pads, and your prompt (and template) improve from evidence instead of vibes.

Make it yours

• The standing PRD Project: a Claude Project with your template, three of your best past PRDs, and your product context in the instructions — every draft starts consistent with house style, and the citations rule lives in the instructions instead of every prompt.
• ChatPRD for structure, Claude for attack: some teams get better scaffolding from a purpose-built tool and a better adversarial pass from a general model with a hostile system framing — use different tools for the two AI steps and keep whichever pairing your retros favor.
• The eng-preview variant: before formal review, give the consistency-pass output (not just the PRD) to your tech lead — showing what the attack already caught, and what you fixed, changes review from fault-finding to decision-making.

Quality check

• The judgment sections read like youif the problem statement could have been generated, it was; rewrite until it commits to something a generic draft wouldn't.
• Planted conflicts surfacedcheck that the evidence pack's known tensions appear in Open Questions; if the draft smoothed them over, the citations rule isn't binding and needs tightening.
• Every scaffolding claim tracesspot-check five inline citations against the pack; [INFERRED] tags are resolved or promoted to open questions, never silently kept.
• The consistency pass found somethinga clean report on a first draft means the attacker was polite, not that the draft was perfect; rerun with a harsher framing.
• You can defend the doc without opening itthe review-room test: problem, why-now, non-goals, and the success metric, from memory, under challenge. That's what "keeping the judgment" means.