Pre‑flight PRDs: Pilot discipline for AI specs
Use a pilot’s brief and sterile‑cockpit rules to draft AI PRDs, force readbacks, run failure sims, and hand off with fewer gaps.

This guide borrows from aviation. Use a pre‑flight brief and sterile‑cockpit rules to write AI PRDs and specs. It is for PMs who want predictable handoffs. At the end, you will have a step‑by‑step method and a one‑page PRD example for an AI ranking model.
In brief
- Start with a 2‑minute decision summary: goal, primary metric, must‑not‑break.
- Generate a draft with AI in 15–30 minutes. Keep trade‑offs out of that draft.
- Run a sterile slot to decide trade‑offs. No drive‑by edits.
- Require readbacks from ENG, DATA, DESIGN, SECURITY (30–60 minutes total).
- Cross‑check 8 binary items: testability, acceptance, data sources, fallback, bias, latency SLA, rollback, cost.
- Simulate failure modes before sign‑off. Short scenarios with recovery steps.
- Measure success: ambiguity comments down by ≥40% on first ENG pass.
Borrowed from aviation: translate to PM work
| Aviation practice | Spec writing move |
|---|---|
| Pre‑flight briefing sets intent and key risks | Begin every PRD with a 2‑minute decision summary: goal, primary metric, must‑not‑break |
| Sterile cockpit during taxi, takeoff, landing | Block a finalisation slot: only trade‑off decisions; no unrelated edits or chats |
| Readbacks between crew and ATC | Each discipline reads back acceptance criteria in their own words |
| Checklists with binary states | Eight yes/no cross‑checks before sign‑off |
| Emergency procedures | Failure‑mode simulations with defined recovery steps |
The method: seven steps
Pre‑flight brief
Write a 2‑minute decision summary: the goal, one primary metric, and what must not break.
AI draft
Ask AI to draft the PRD/spec outline and neutral language in 15–30 minutes. Keep trade‑offs and acceptance criteria for humans.
Sterile slot
Block a short window to decide trade‑offs. No side edits. Record decisions and rationale.
Roles & readbacks
Have ENG, DATA, DESIGN, SECURITY read back their acceptance criteria. Clarify, then lock.
Cross‑check
Run the eight binary checks. Anything not a clear yes/no needs rework.
Failure sims
Write short failure scenarios. Define detection and recovery steps.
Sign‑off & handoff
Version the PRD, capture open questions, and hand off. A PRD template helps keep the flow the same each time.
What AI should draft, and what stays human
- Structure for the PRD/spec and neutral, unambiguous prose
- Lists of candidate risks, tests, and data sources to review
- Placeholders for metrics, SLAs, and checklists
- Trade‑offs that hit cost, latency, privacy, and UX
- Acceptance criteria per discipline
- Go/no‑go rules, fallbacks, and rollback triggers
The cross‑check checklist (8 binary items)
Worked example: one‑page PRD for an AI ranking model
Goal: Ship an AI ranking model that orders a list to improve the primary metric without breaking current must‑not‑break behavior.
Primary metric: One north‑star metric agreed for this release. Must‑not‑break: Existing contractual or critical flows remain intact.
- In: Model scoring service, ranking logic, fallback to non‑AI ranking, logging for evaluation.
- Out: New UI surfaces, new data collection, long‑term retraining pipeline.
Use a focused window to set these. No other edits.
- Model vs latency: decision and rationale recorded here.
- Metric vs risk: how we guard the must‑not‑break behavior.
- Compute budget vs recall: bounds agreed here.
Each group reads back their acceptance criteria in their own words.
- ENG readback: API shape is fixed; deployment, rollback, and observability steps are clear; latency SLA is testable.
- DATA readback: Data sources and contracts named; evaluation set and offline checks planned; bias checks and thresholds stated.
- DESIGN readback: Fallback UX defined; empty and error states covered; no surprise ranking shifts in must‑not‑break moments.
- SECURITY readback: Data handling and access reviewed; logging scope approved; recovery paths do not expose sensitive data.
- Testability: Requirements phrased as pass/fail.
- Acceptance: Readbacks complete; sign‑offs captured.
- Data sources: Inputs and contracts named.
- Fallback: Non‑AI ranking path defined and tested.
- Bias: Checks and thresholds defined.
- Latency SLA: SLA stated; monitor in place.
- Rollback: Trigger and procedure defined.
- Cost: Run‑time cost and rollback $/hour estimation noted.
- Scenario 1: Ranking API exceeds latency SLA. Detection: alert on SLA breach. Recovery: route traffic to fallback ranking; start rollback procedure if breach persists.
- Scenario 2: Data source missing or schema change. Detection: contract check fails on ingest. Recovery: switch to last good snapshot or fallback ranking; log event for DATA to remediate.
- Scenario 3: Bias check fails post‑deploy. Detection: scheduled bias evaluation flags threshold breach. Recovery: rollback to last approved model; freeze training data updates; open an incident for root cause.
- Latency SLA: Define the SLA and the test method (e.g., p‑level definition tied to user experience).
- Cost estimate: Note expected compute/runtime $/hour for this SLA.
- Rollback cost $/hour: Describe how to estimate (compute cost/hour + expected impact/hour). Record the number when known.
- AI draft time: 15–30 minutes.
- Readback time: 30–60 minutes total.
- Spec quality: ambiguity comments drop by ≥40% on first ENG pass.
- Versioned PRD shared; open questions labeled; owners named.
- Fallback and rollback runbooks linked.
- Monitoring dashboards and alert thresholds listed.
Mistakes to avoid
- Letting AI write acceptance criteria without human readbacks.
- Relying only on example behaviors from a model instead of testable rules.
- Skipping cost and operations checks; no rollback $/hour noted.
- Making iterative, ad‑hoc edits during sterile decision slots.
Run this on your next spec
- Write the 2‑minute decision summary first.
- Use AI to produce a full draft in 15–30 minutes.
- Book a 20–30 minute sterile slot to decide trade‑offs only.
- Schedule a 15–minute readback and get each discipline to read theirs.
- Complete the 8 binary checks. Anything fuzzy becomes an action.
- Draft three failure scenarios with detection and recovery.
- Sign off, estimate rollback $/hour, and hand off.


