Draft a PRD from a synthetic evidence pack using AI — but hand-write the judgment sections yourself. The pack plants two conflicting data points the AI draft will smooth over; your PRD's open-questions section must surface both. Then add an edge-case pass and a probabilistic-requirements annex for the AI sub-feature.
Priya is a Senior PM at a Chennai SaaS company whose product handles invoice workflows for mid-market finance teams. Next quarter's committed feature: auto-escalation for stuck approvals, with an AI sub-feature that suggests the right escalation approver. She has an evidence pack from discovery. Her last AI-drafted PRD demoed well and shipped wrong because the draft's confidence papered over the evidence's contradictions — this time the workflow is different: AI drafts the structure, Priya owns the judgment, and every conflict in the evidence gets surfaced instead of smoothed.
Three documents, condensed. Paste-ready.
DOC 1 — DISCOVERY SYNTHESIS (8 interviews, AP clerks + approvers)
- Clerks say invoices "disappear into approval limbo"; no visibility on
who is sitting on what. Top request: escalation when nothing moves.
- Approvers (4 of 4 interviewed) say they clear their queue "same day,
next day at worst" and blame delays on invoices reaching the wrong
approver in the first place.
- 3 of 8 participants asked for delegate-while-on-leave functionality.
- Nobody asked for AI. Two participants said misrouted invoices are the
root cause: "it goes to whoever the PO template names, even if they
left the company."
DOC 2 — METRICS SNAPSHOT (last 90 days, all tenants)
- Invoices entering approval: 148,200
- Median time-to-approve: 3.4 days; p90: 11 days
- Invoices reassigned at least once: 19%
- Invoices auto-matched to PO without human touch: 62%
- Support tickets tagged "approval delay": 214 (up 40% QoQ)
- Approver median queue size at any moment: 14 invoices
DOC 3 — CONSTRAINTS MEMO (from Eng lead + Finance Ops)
- One backend engineer plus one frontend for this quarter. No new
services; must run on the existing rules engine.
- Finance Ops: "auto-match is barely 40% for our biggest three tenants
— their PO formats break the matcher. Any escalation logic that
assumes clean PO data will misfire for exactly the accounts that
complain loudest."
- Compliance: escalation must never skip the amount-based approval
matrix. Audit log required for every reassignment.
- AI suggestions allowed only if a human confirms; no auto-reassign.
Read the pack and write your judgment notes first — no AI yet. Fifteen minutes. Answer in your own words: what problem is actually worth solving, what evidence conflicts with what, and what would you not build. Two data points in this pack cannot both be true as stated — find at least one now, before the AI's fluency makes everything seem consistent.
Generate the skeleton draft. Paste all three docs with: "Draft a PRD for auto-escalation of stuck invoice approvals with an AI-suggested-approver sub-feature. Sections: problem statement, evidence summary, user stories, requirements, out of scope, risks, open questions. Use only the evidence provided. Where the evidence conflicts, say so explicitly instead of reconciling it." That last sentence rarely fully works — which is the point of the next step.
Audit the draft's evidence summary against the pack. Line by line: does the draft state the approver-behavior story and the latency numbers as one coherent narrative? Does it quote one auto-match rate or both? Mark every place the draft asserted consistency the pack does not contain. AIs reconcile by default — smooth prose over conflicting numbers is the failure mode you are hunting.
Hand-write the judgment sections. Rewrite these yourself, using the draft only as scaffolding: the problem statement (what you decided the real problem is — misrouting vs. slow approvers is a decision, not a summary), prioritization of requirements against the two-engineer constraint, and the open-questions section. Every evidence conflict you found goes into open questions with the exact numbers or quotes on both sides, plus how you would resolve it (data pull, follow-up interview, per-tenant analysis).
Run the AI edge-case pass. Prompt: "List 15 edge cases and failure scenarios for this PRD: approver on leave, approver left company, amount-matrix conflicts, circular escalations, tenant with broken PO matching, and any others." Triage its list yourself into: must-handle (add to requirements), explicitly out-of-scope (write it down), and ignore (with a reason). Do not paste all fifteen into the PRD — undigested AI edge-case lists are how PRDs bloat.
Write the probabilistic-requirements annex for the AI sub-feature. Hand-write this — it is the section AI writes worst. It must contain: the suggestion-acceptance target that gates launch (e.g., "suggested approver accepted without change in ≥70% of cases on the eval set"), what happens on low confidence (fall back to showing the rules-engine default with no AI suggestion), the wrong-suggestion cost analysis (a wrong suggestion a human confirms still violates the audit expectation — say how the UI makes the matrix constraint visible), and what gets logged per suggestion for later error analysis.
Red-team your own PRD. Prompt: "You are a skeptical engineering lead with one backend engineer. Attack this PRD: what is underspecified, what contradicts the constraints memo, what will you be asked to build that the evidence does not justify?" Accept or rebut each attack in a margin note. Keep the two best attacks in the PRD's risks section.
Write the retrospective. Ten lines at the end of your doc: which sections AI genuinely wrote, which you rewrote and why, which conflict AI smoothed over and how you caught it, and the one prompt you would reuse. Honest accounting — this is the workflow you are actually training.
The AI drafts structure and coverage; the PM owns the problem statement, the priorities, the open questions, and every number — those are exactly the sections a smooth draft tempts you to skip.
A complete lab has all of the following:
Evidence packs contradict themselves; AI drafts do not. The PM's job in an AI-drafted PRD is to reintroduce the contradictions as open questions with numbers on both sides — and to hand-write the probabilistic annex no model writes credibly.