What this lesson is
Lesson 04 was your first AI PRD — a 24-minute draft from a one-line ask. This lesson is the production-grade version: the canonical template every AI PRD needs once a feature stops being a flag-gated v0 and starts being a thing real customers depend on.
Eight sections. Most PMs skip three of them on their first try, and each skipped section maps cleanly to a class of post-launch incident. The point of having a canonical template is not bureaucracy — it's pattern matching. When the PRD is laid out the same way every time, an engineer, a legal counsel, a data lead and a CFO can each read the section that concerns them in five minutes and not miss anything.
This is the lesson you screenshot, paste into your Notion sidebar, and refer to every time someone asks you to ship an AI feature. It is the reference document.
The 8 sections of a canonical AI PRD
Lesson 04 had seven sections. Production adds an eighth — Rollout plan — because v1 isn't "deploy on Tuesday." It's shadow → canary → 100%, with a launch gate between each step. Below are the eight, in the order they appear in the doc.
Section × skip-if-missing × owner × failure mode
The fastest way to internalise the 8 sections is to know what breaks when you cut each one. This table is what Chidi pins to his Notion at NairaSwift before writing the v1 PRD. Notice the right-most column — every skipped section has a named, recurring failure that production AI teams have learned the hard way.
| Section | Skip if missing? | Primary owner | Common failure |
|---|---|---|---|
| 1. Problem & users | Never | PM | Beautiful feature solving the wrong problem; v2 is a complete rewrite |
| 2. Scope (in/out) | Never | PM | Users type out-of-scope intents; model improvises; legal incident |
| 3. Prompts & behavior | Never | PM (eng implements) | Tone drift across releases; user trust score quietly falls |
| 4. Data & retrieval | Never | PM + Data lead | Model leaks PII or cites stale data; DPDP/GDPR exposure |
| 5. Eval & launch gate | Never | PM + Eng | "We'll know it's working when…" — every exec meeting becomes a vibe check |
| 6. Rollout plan | Only if v0 already at 100% | PM + Eng | Day-one regression on edge case nobody saw in QA; no rollback path |
| 7. Risks & fallbacks | Never | PM + Security | Prompt injection, cost spike, vendor model upgrade — discovered on Twitter |
| 8. Metrics & success | Never | PM + Analytics | Launch happens; nobody knows if it worked; feature quietly deprecated in 9 months |
Scenario — Chidi at NairaSwift Lagos
Chidi is a senior PM at NairaSwift, a Lagos-based remittance startup. Six weeks ago he shipped a v0 dispute-assistant behind a feature flag — when a user files a "money sent but not received" dispute, an AI explains the likely cause (correspondent-bank delay, mismatched name, FX cutoff) and the next step. v0 is at 5% of dispute traffic, has a 67% thumbs-up rate, and saved roughly ₦4.2M in agent time last month. Leadership wants v1 at 100% before the next compliance audit. Chidi has 10 weeks.
v0 was a Notion page and a Slack thread. v1 needs an actual PRD that survives review by engineering, the head of compliance, the CFO, and a sceptical VP of CX who already lost trust in chatbots once. Chidi opens the canonical 8-section template. The rest of this lesson is what each section looks like by Friday.
How Chidi fills each of the 8 sections
Below is the actual sequence — eight step-blocks, one per PRD section, with what Chidi puts inside. Read it as a worked example, not a template to copy verbatim. The sections are universal; the contents are NairaSwift-specific.
Chidi opens with the user pain in user words, not PM words: "NairaSwift senders file ~14k disputes/month. 61% are not actually disputes — they're users panicking because the recipient hasn't received naira yet, and we haven't told them why. Average time-to-resolution is 38 hours. CSAT after a dispute resolution is 18 points lower than the baseline."
Then the goal, with three numbers: deflect 45% of dispute filings via the AI explainer in Q3; thumbs-up rate ≥ 75%; zero false-reassurances on actually-failed transfers (the safety constraint). He names the top three intents from v0 telemetry — correspondent-bank delay, name-mismatch hold, weekend FX cutoff — and notes they cover 78% of v0 traffic. Everything else gets a hard handoff. Personas are spelled out as: Sender (anxious, mobile, English or Pidgin), CX agent (using the AI as a co-pilot, not a replacement).
Two columns. In scope (v1): explain the three named intents above; offer a "wait X hours" expectation; offer "talk to an agent" handoff; offer "track this transfer" link. Out of scope: initiating a refund, cancelling a transfer, KYC-related questions, anything mentioning fraud, disputes against a specific merchant. Each out-of-scope intent gets a verbatim handoff string written in this section, not in the prompt section — this is so Compliance can review the exact words a user will see. The doc says: "If the AI handles an out-of-scope intent in production, that's a P0 incident and the rollout is reversed."
The full prompt anatomy block (see the next section) goes here. Chidi keeps the prose system prompt in the PRD so non-engineers can read it, and the canonical version lives in git as prompts/dispute_explainer_v1.txt with a version pinned next to the model name. This section also names the explicit behaviours: never quote an exact arrival time (only ranges), never mention any correspondent bank by name unless it's in the retrieved data, never say the word "guaranteed", always include the transfer reference. The refusal strings from Section 2 are referenced — not duplicated.
A field-by-field allow-list. The model can see: transfer ID, amount in NGN and destination currency, sender country, recipient country, status code, status timestamp, correspondent bank code (not name), expected arrival window. The model cannot see: full names, phone numbers, BVN/KYC data, linked cards, any other transfers from the same sender. Freshness: status updated within 5 minutes. Retrieval is keyed off transfer ID only — no semantic search across user history. Audit log: every retrieved record is hashed and stored for 90 days with the prompt and response, for compliance review. PII boundary: if the retrieval somehow returns a name field, it is stripped before reaching the prompt.
Golden set: 200 real disputes from v0, hand-labelled with the ideal one-paragraph response. Two graded axes — correctness (did the AI cite the right status code and arrival window?) and helpfulness (did the user understand and stop escalating?) — each scored 1–5 by an LLM-as-judge with weekly human spot-checks. Launch gate in one sentence: "Ship to next rollout step when ≥ 82% of golden set scores ≥ 4 on both axes, AND a 30-item 'failed-transfer' subset has zero false-reassurances." The fail subset is the zero-tolerance gate — even one "your money is on the way" answer on a transfer that actually bounced means no rollout.
Four steps with explicit gates. Shadow (week 4): AI runs on 100% of disputes, output never shown to users, scored against agent answers. Gate: golden set ≥ 82%. Canary 5% (week 5): live to 5% of senders. Gate: live thumbs-up ≥ 70%, zero P0 fallback firings, cost ≤ ₦12 per dispute. Canary 25% (week 7): Gate: thumbs-up ≥ 73%, agent escalation rate within 10% of baseline. 100% (week 9): Gate: thumbs-up ≥ 75%, full compliance sign-off. Rollback rule named for each: any gate miss for 24 hours triggers automatic step-down to the previous percentage.
A four-column table inside the PRD. Each row: a named risk, how we detect it, the fallback, the owner. Examples: Hallucinated arrival time → format rule restricts to retrieved windows only → log monitor alerts on any answer with a date, owner: Eng. Prompt injection from transfer reference field → strip control chars and reject inputs > 64 chars at retrieval, owner: Security. Vendor silently upgrades model → pinned version + weekly automated re-eval, ≥3-point regression auto-rollbacks, owner: Eng. Cost spike from longer prompts → per-conversation token cap of 1500, hard fallback to templated reply, owner: PM. Compliance audit finds PII leak → audit log + retrieval allow-list + quarterly review, owner: Compliance. Twelve rows total. Each one named, no row of "various edge cases."
Three tiers. North-star: dispute-deflection rate (target 45%). Quality guardrails: thumbs-up rate ≥ 75%, refusal rate stable at 8–12% (too low means the model is over-helping; too high means it's punting), agent-escalation latency unchanged. Cost & ops: cost per dispute ≤ ₦12, p95 latency ≤ 4s, fallback firing rate ≤ 1%. Each metric names the dashboard panel, the alert threshold, and what the on-call PM does if it breaches. Crucially, the section ends with "What success means in 90 days": a one-paragraph statement of the outcome, signed by Chidi and the VP of CX. That paragraph is the PRD's closing argument.
What Chidi's finished v1 PRD looks like
Below is the at-a-glance preview — page one of the Notion doc. If a reviewer reads only this, they should already understand what's shipping, what isn't, the launch gate, and the rollout plan. This is the screenshot Chidi pastes into the v1 launch Slack channel.
Owner: Chidi · Target 100%: Week 9 · Launch gate: ≥ 82% golden set · ≥ 75% thumbs-up · 0 false-reassurance · ≤ ₦12/dispute
prompts/dispute_explainer_v1.txtThe PRD's lifecycle — draft to in-flight
The PRD is not written in one sitting. It moves through six states, and the canonical template is what makes those state transitions reviewable. Skip a state and the document silently degrades into a wishlist again.
Inside section 3 — what "Prompts & behavior" actually contains
This is the section reviewers most often glaze past, and the one that determines whether the model behaves the way the PRD claims. Below is exactly what Chidi pastes into section 3 — a colour-coded prompt anatomy block that compliance, eng and CX can each read in 60 seconds.
TONE Calm. Direct. Never apologise before the facts. Never use the word "guaranteed". Match the user's language register (English or Pidgin) but stay neutral.
FORMAT One paragraph, ≤ 60 words. Always include: the transfer reference, the current status code, the expected arrival window (range only, never an exact time). End with one of two follow-ups: "Track this transfer" or "Talk to an agent".
REFUSAL If the user asks for a refund, mentions fraud, asks about KYC, or names a merchant, respond exactly: "That one needs an agent — I'm connecting you now." Do not attempt to explain or reassure. Do not name the correspondent bank.
.txt file in a code repo, but they need to sign off on the exact tone and refusal language. Putting the colour-coded version in the PRD makes the behavioural contract reviewable by non-engineers — and any change to the live prompt requires a PRD revision, not just a pull request.
The AI connection
How real AI teams use the canonical template
At Anthropic, Stripe, SouthBank India and PhonePe, AI PRDs follow a remarkably similar 7–9 section shape. The PRD is the input to four downstream artifacts: a versioned prompt file in git, a retrieval config, an eval CI job that blocks deploys, and a rollout config in the feature-flag service. PMs who write tight section 3 (prompts), 4 (data), 5 (gate) and 6 (rollout) save eng teams entire sprints of clarification meetings. The canonical template is what turns "AI feature" from a Slack thread into a deployable artifact. prd-template launch-gate rollout-config prompt-repo eval-ci audit-log
Mini challenge — pressure-test your last AI PRD
Take a real PRD you've written and grade it against the canonical template
For each of the eight sections, paste the heading from your PRD that matches. If a section has no match, write "MISSING" in red. Most PMs find 2–3 missing on their first audit. Don't fix them yet — just count.
Using the table from the deep-dive, name the specific failure mode each missing section invites. Be concrete: "skipped Rollout → no canary, day-one regression on edge case" not "could be risky."
Open section 5 and reduce your launch gate to a single sentence with three numbers: pass rate, golden-set size, zero-tolerance subset. If it doesn't fit in one sentence, the gate is still negotiable. Tighten it until a junior engineer could implement it without asking you.
Interview questions
Test your knowledge
Common mistakes writing a canonical AI PRD
Key takeaways
- Canonical AI PRDs have 8 sections: problem & users · scope · prompts & behavior · data & retrieval · eval & launch gate · rollout · risks & fallbacks · metrics & success. Use these names verbatim.
- The 8th section — Rollout plan — is what separates v0 from v1. Eval says the model can do the task; rollout says it survives real traffic. Skip rollout and you ship at 100% on day one with no rollback path.
- Section 3 (Prompts & behavior) lives both in the PRD and in git. The PRD version is for compliance and CX; the git version is for engineering. Any change requires both.
- Launch gates are two thresholds: an average pass rate AND a zero-tolerance subset for the highest-cost error class. Single-number gates hide rare disasters.
- Risks & fallbacks is a table with four columns — risk · detection · fallback · owner. A number or a name in every cell. Prose risks read as fear; tables read as readiness.
- Every metric in section 8 needs a target, an alert threshold, and a named on-call action. Without the third, the dashboard is decoration and the metric is theatre.
You now write AI PRDs that ship.
Next — design the rollout that proves them.
Join the Best AI Program: The AI-Era Agentic Engineer — 5 months, fully online, with placement assistance. Week 1 is free.
Explore The AI-Era Agentic Engineer →