What this capstone is
You've made it through 19 lessons. You can argue layers, defend a launch gate, write a refusal string and read a model card. This is the final sit-down — the lesson where every previous concept is put on a single page and assembled into the artifact that decides whether an AI feature ships: a full, production-grade PRD.
The previous capstone-style lesson, Lesson 13, gave you the canonical 8-section template with one or two illustrative paragraphs per section. This one is the inverse. Same eight sections — but each section is filled out with the exact content Chidi at NairaSwift Lagos hands to engineering, compliance, the CFO, and the VP of CX before he gets approval to ship "explain-my-decline" to 100% of senders.
Read it two ways. Pass one — follow along, watch how each section gets written, notice which sentences are doing the work. Pass two — pull up a PRD you've drafted recently and run it side-by-side. The gaps will be obvious. The point of a capstone isn't new theory; it's a forcing function for what you already half-know.
The PRD spine — eight sections, one feature, every concept from the track
Here are the eight sections again, but this time framed as the capstone synthesis. Notice the right column — every section maps to at least one earlier lesson, which is what makes this a capstone instead of a recap.
The scenario — Chidi at NairaSwift Lagos
Chidi is the same senior PM at NairaSwift Lagos you met in Lesson 13 — but this is a different feature, and a fresh 10-week clock. Six weeks ago, exec leadership saw the decline-rate dashboard: ~9% of all NairaSwift transfer attempts get declined upstream (insufficient funds, FX cap reached, recipient bank closed for batch, fraud-rule trip). Of declined senders, 41% retry within 5 minutes. Of those retries, 62% are identical — same amount, same bank — and decline again. Support tickets tagged "why was my transfer declined?" are the #2 driver of CX volume in Lagos. A v0 prototype built by an eng-side hack week saw a 28% reduction in retry-decline loops in a 200-user dogfood. Leadership greenlit the build.
Chidi's brief is one sentence: "When a NairaSwift transfer is declined, show the sender a plain-English reason and the next step, generated by the LLM, before they hit retry." He has 10 weeks to write the PRD, freeze the eval, run shadow + canary, and ship to 100% of declined senders before the Q3 risk review. The CFO wants the unit economics in the PRD. The head of Trust & Safety wants the refusal list. The VP of CX wants the tone reviewed before any prompt change. Engineering wants a single eval gate they can wire into CI. All four read the same document. That document is the rest of this lesson.
The full PRD — as it appears in Notion
Below is what Chidi's v1.0 doc looks like by Friday of week 3. It's not a TOC — each row carries the production content for that section, in the same prose Chidi would defend in a review. Skim it once, then read each step-block in the next section to see the reasoning behind every line.
Owner: Chidi · Target 100%: Week 9 · Reviewers: Eng, T&S, Compliance, CFO, VP CX · Launch gate: ≥ 84% golden set · 0 false-success on declined-and-failed subset · ≤ ₦9 / decline · p95 latency < 2.2s
prompts/decline_explainer_v1.txt at SHA pinned to model version.What Chidi chose at each stack layer
Before zooming into each PRD section, here's the architectural shape — the same five-layer stack from Lesson 01, annotated with what Chidi committed to in this PRD. The PRD is the artifact that pins each one of these to a versioned decision.
How it works — Chidi's reasoning, section by section
Eight step-blocks, one per PRD section. Each block has the production content (above, in the doc preview), Chidi's why, and what he debated cutting. This is the part you read on pass two — when you're auditing your own draft.
What Chidi wrote. The opener leads with three numbers — 9% decline rate, 41% retry within 5 minutes, 62% of those retries identical — because the problem is not "users get declined" (that's a system fact) but "users get declined and we don't tell them why, so they retry the same thing and burn another decline event." That reframing is the entire point of the section. He follows it with a one-line trust claim: declined-but-no-explanation senders churn at 2.3× the rate of declined-with-explanation senders, sourced from the v0 dogfood.
Why it's written that way. The reviewer audience is split: CFO wants volume × unit-cost, T&S wants user-pain framing, eng wants intent counts. The opening paragraph has all three. The personas are spelled out — Sender (mobile, anxious, EN or Pidgin) and CX agent (using the AI as co-pilot, not a replacement) — because the tone in section 3 has to land for both audiences and the eval in section 5 has to test both. What he debated cutting: the CX-agent persona. v0 was sender-only. He kept it because dogfood feedback showed agents were already pasting the AI explanation into ticket replies; designing for that explicitly is cheaper than letting it happen accidentally.
What Chidi wrote. Two columns. In: five named decline codes — insufficient funds, FX daily-cap, recipient-bank batch closed, fraud-rule trip-low-confidence, KYC-tier limit — covering 91% of decline volume. Each one ships with a paired next-step the AI is allowed to surface ("wait until 6am WAT batch", "raise daily limit in settings", "talk to agent"). Out: refunds, cancellations, high-confidence fraud trips, KYC tier changes, and any answer that names a specific bank. Each out-of-scope intent has a verbatim refusal string written in this section, not buried in the prompt section.
Why. The high-confidence fraud subset is excluded by design — Chidi argued, and T&S agreed, that explaining a fraud-flag in user-facing language is a fraud-vector itself (it teaches the attacker which signal tripped). That single decision required a separate refusal string and a hard route to support. What he debated cutting: the fifth decline code — KYC-tier limit — because it covers only 6% of declines. He kept it because the alternative was a refusal in 6% of cases, which would tank thumbs-up. What he debated adding: "high-confidence fraud" as an in-scope refusal. He left it explicitly out-of-scope because letting the AI even acknowledge "this looks like fraud" is a leakage class.
What Chidi wrote. The full system prompt lives in git, but the PRD shows the colour-coded anatomy block — ROLE, TONE, FORMAT, REFUSAL — so a non-engineer can review the exact behaviour. Role: NairaSwift decline explainer. Tone: calm, factual, never apologise before facts, never use the word "guaranteed", match sender language register (English or Pidgin) but stay neutral. Format: ≤ 50 words, single paragraph, always cite the decline code, always provide one next step, end with a CTA from a closed list of three. Refusal: on high-confidence fraud, respond exactly with the verbatim string from §2 — no rewording, no improvisation.
Why. The PRD-and-git split is non-negotiable: compliance and the VP of CX cannot review a .txt file in a code repo, but they need to sign off on tone and refusal language verbatim. Putting the colour-coded version in the PRD makes the behavioural contract reviewable by non-engineers, and any future prompt edit forces a PRD revision instead of just a pull request. What he debated cutting: the "match sender language register" clause. Eng pushed back on the eval cost (Pidgin examples doubled the golden set). Chidi kept it because dropping it meant the feature would feel imported in Lagos, which kills the trust metric.
What Chidi wrote. A field-by-field allow-list. The model can see: decline code, decline timestamp, attempted amount in NGN, sender country, recipient country, sender wallet balance bucket (not exact balance), sender daily-limit tier, retry count in the last 24 hours. The model cannot see: full names, BVN/KYC data, phone numbers, recipient bank name, any other transfers from the same sender. Freshness: real-time, since the decline event itself is the trigger. Retrieval is keyed off transfer ID only — there is no semantic search across user history. Audit log: every retrieved record is hashed and stored with prompt + response for 90 days, queryable by transfer ID for compliance review. PII boundary: if retrieval returns a name field by mistake, it is stripped before reaching the prompt.
Why. Chidi designed the allow-list before the prompt, on purpose. The prompt is what the model can say; the allow-list is what the model can know. If the allow-list is loose, no prompt rule will save you in audit. The "balance bucket, not exact balance" choice is the kind of detail that costs nothing in eval and saves you in compliance review — knowing a sender has < ₦5,000 versus exactly ₦4,317.50 is the same information for the explanation, but only the latter is PII. What he debated cutting: the 24-hour retry-count field. T&S wanted it removed (retry patterns can be a fraud signal). He kept it because the explanation often references "you've tried this twice in the last hour", which directly improves the retry-loop metric. The compromise: retry count is bucketed (0, 1, 2+) so it can't be used to backwards-infer fraud thresholds.
What Chidi wrote. A 250-item golden set sampled from the last 6 weeks of declines, hand-labelled across the 5 in-scope codes plus a 40-item zero-tolerance subset of declined-and-actually-failed-irrecoverably cases. Two graded axes: code-correctness (did the AI cite the right decline reason?) and helpfulness (sender understood and did not retry-identically?), each scored 1–5 by an LLM-as-judge with weekly human spot-checks of 20 random items. Launch gate, in one sentence: "Ship to next rollout step when ≥ 84% of golden set scores ≥ 4 on both axes, AND the 40-item declined-and-failed subset has zero false-success answers."
Why. The two-threshold gate is the critical move. A single average pass rate would let the 40 zero-tolerance items dilute into the noise — but a "your money is on its way" answer on a transfer that actually bounced is not a 1-point score deduction, it's an incident. AND-ing the two thresholds means a single false-success blocks rollout regardless of average. What he debated cutting: the helpfulness axis. Eng argued it was hard to score and slowed eval. Chidi kept it because code-correctness alone would let a technically-right but unintelligible answer ("DECLINE_CODE_FX_DAILY_CAP_EXCEEDED") pass the gate while nuking the trust metric. What he debated raising: the threshold from 84% to 88%. He kept it at 84% because the v0 prototype's hand-graded baseline was 81% and 88% was statistically out of reach by week 4 — promising it would have meant a missed launch.
What Chidi wrote. Five rollout steps, each with an explicit gate and an explicit rollback rule. Shadow (week 4): AI runs on 100% of declines, output never shown to users, scored against agent-written explanations. Gate: golden set ≥ 84%. Canary 1% (week 5): live to 1% of declined senders. Gate: thumbs-up ≥ 70%, zero P0 incidents, cost ≤ ₦9 / decline. Canary 10% (week 6): Gate: thumbs-up ≥ 73%, retry-decline-loop reduction ≥ 18%. Canary 50% (week 8): Gate: thumbs-up ≥ 76%, loop reduction ≥ 28%, agent escalation latency unchanged. 100% (week 9): Gate: thumbs-up ≥ 78%, loop reduction ≥ 30%, full compliance sign-off. Rollback rule: any gate miss for 24 hours triggers an automatic step-down to the previous percentage.
Why. Five steps instead of three because NairaSwift's traffic is spiky — a single canary at 10% would not generate enough volume in a weekend to detect a regression before Monday. The 1% step exists almost entirely to catch p95 latency and fallback-firing rate before user-facing thumbs-up signal becomes meaningful. What he debated cutting: the shadow phase. Eng wanted to skip it ("we already have v0 data"). Chidi kept it because v0 was a different prompt and a different model — comparing v1 outputs against v0 outputs is not the same as comparing v1 against agent-written ground truth, and the 250-item golden set alone doesn't catch latency or cost. Shadow week is cheap; skipping it is how teams ship at 100% on day one and discover edge cases on Twitter.
What Chidi wrote. Fourteen rows, four columns each. The top five — by impact × likelihood — are: (1) hallucinated decline reason → AI cites a code not in the retrieved data → format rule restricts answers to allow-listed codes only; log monitor alerts on any answer naming a code outside retrieval; owner Eng. (2) prompt injection from the optional "memo" field → user-supplied text reaches the prompt → strip control chars + reject inputs > 64 chars at retrieval; owner Security. (3) vendor model upgrade silently regresses behaviour → pinned SHA + weekly automated re-eval; ≥ 3-point regression auto-rollbacks to prior pin; owner Eng. (4) cost spike from longer prompts → per-call token cap of 800; hard fallback to templated reply if cap is hit; ₦15/decline budget alert; owner PM. (5) compliance audit finds PII leak → allow-list + audit log + quarterly review; owner Compliance.
Why. Every row has a number or a name in every cell — no "we'll monitor", no "the team owns it". The structural move is what turns the section from "things that worry me" into "things we have already decided how to handle." The section closes with one line: "Any production firing of a fallback is a P1 incident with a 7-day post-mortem." That single sentence shifts the section's tone from scared to ready. What he debated cutting: rows 11–14, which cover edge cases below the 1% likelihood floor. He kept them because compliance asked specifically about regulator response time, vendor data-residency, and audit-log retention — the rows existed to give those questions a written answer.
What Chidi wrote. Three tiers of metrics, each with target, alert threshold, and named on-call action. North-star: retry-decline-loop rate, target -35% YoY by Q4. Quality guardrails: thumbs-up rate ≥ 78%, refusal rate stable in 6–10% (below = model over-helping; above = model under-helping), agent-escalation latency unchanged from baseline. Cost & ops: cost per decline ≤ ₦9, p95 latency ≤ 2.2s, fallback-firing rate ≤ 0.8%, golden-set CI must pass on every deploy. Closing line: a one-paragraph statement of what success means in 90 days, signed by Chidi and the VP of CX. That paragraph is the PRD's closing argument and the criterion the post-launch retro is graded against.
Why. Each metric names three things — target, alert threshold, on-call action — because a metric without a named action is decoration. The refusal rate range (6–10%) is the under-appreciated trick: a metric with a floor, not just a ceiling, catches the failure mode where the model gets too helpful and starts answering questions it should refuse. What he debated cutting: the closing signed paragraph. The CFO thought it was theatre. Chidi kept it because every previous AI feature at NairaSwift that quietly got deprecated nine months in had no signed success statement — without one, "did this work?" becomes a political question instead of a written one. What he debated adding: a separate trust-score metric. He merged it into thumbs-up to avoid metric proliferation; the closing paragraph carries the qualitative trust claim.
Section × what Chidi wrote × why × what he debated cutting
The same eight rows, condensed. This is the table Chidi pasted at the top of the PRD before the first review meeting — it's the fastest possible scan for a stakeholder who has 90 seconds. Read it as the executive summary; read the step-blocks above when you want to defend a choice.
| Section | What Chidi wrote | Why | What he debated cutting |
|---|---|---|---|
| 1. Problem & users | 9% decline · 41% retry · 62% identical retries · trust drop 2.3× | One paragraph that lands for CFO, T&S, eng simultaneously | The CX-agent persona — kept because dogfood agents already paste outputs |
| 2. Scope (in/out) | 5 in-scope decline codes (91% of volume) · 5 verbatim refusals | Closed-set scope makes eval tractable; verbatim refusals make compliance reviewable | KYC-tier code (only 6% volume) — kept to avoid 6% refusal-driven thumbs-down |
| 3. Prompts & behavior | Role · tone · format · refusal anatomy block in PRD; .txt in git |
Non-engineers must review behavioural contract; PRD edit gates prompt edit | Pidgin register — kept; Lagos trust metric collapses without it |
| 4. Data & retrieval | 8-field allow-list · balance bucket not value · 90-day audit log | Allow-list before prompt; bucketing trades zero eval signal for big PII reduction | 24h retry count — kept, but bucketed (0/1/2+) so it can't backwards-infer fraud |
| 5. Eval & launch gate | 250-item golden + 40-item zero-tolerance · two axes · 84% AND 0 false-success | Single thresholds let rare-but-catastrophic answers slip through; AND fixes that | Threshold raise to 88% — left at 84% because v0 baseline was 81% and 88% wasn't reachable in 4 weeks |
| 6. Rollout plan | Shadow → 1% → 10% → 50% → 100% · auto-rollback on any 24h gate miss | Five steps because traffic is spiky; 1% step exists for latency/cost detection | Shadow phase — eng pushed to skip; kept because v0 ≠ ground truth and shadow is cheap |
| 7. Risks & fallbacks | 14-row table · risk · detection · fallback · owner · closing P1 rule | Number-or-name-in-every-cell turns "scared" into "ready" | Rows 11–14 (low-likelihood) — kept; gave compliance written answers to specific questions |
| 8. Metrics & success | 3 tiers · refusal-rate range 6–10% · signed 90-day success paragraph | Floor-and-ceiling on refusal catches over-helping; signed paragraph prevents quiet deprecation | Separate trust metric — merged into thumbs-up; signed paragraph carries qualitative claim |
The rollout ladder, with gates
Section 6 of the PRD reads as prose, but it lives on a dashboard as the ladder below. The dashed gate on each rung is the single number that determines whether the next step happens. Miss any gate for 24 hours and the feature-flag service auto-rolls-back to the previous percentage.
The AI connection
How real AI teams use the capstone PRD
At remittance and fintech teams across Lagos, Mumbai and Nairobi, the "explain-my-decline" pattern is one of the most common shipped AI features — high traffic, low creativity, easy to evaluate, easy to roll back. The PRD shape Chidi used here is the same one a senior AI PM would write at any of them. The four downstream artifacts — a versioned prompt file in git, a retrieval allow-list config, an eval CI job that blocks deploys, and a rollout config in the feature-flag service — all derive directly from sections 3, 4, 5 and 6 of this PRD. PMs whose PRDs cleanly emit those four artifacts save engineering teams entire sprints of clarification meetings. capstone-prd launch-gate two-threshold-gate rollout-ladder audit-log refusal-string balance-bucket signed-success
Mini challenge — write your own capstone PRD
Take a real AI feature you'd ship and produce the same 8-section walkthrough
Open a fresh Notion page. Paste eight headings — exactly the names from this PRD. For each one, write three lines: what you wrote · why · what you debated cutting. Resist the urge to skip the third line; that's where the capstone discipline lives.
Write your section 5 with one average pass-rate threshold AND one zero-tolerance subset for the highest-cost error class. If you can't think of a zero-tolerance subset, you don't yet understand the failure modes — interview a CX agent for 20 minutes and try again.
End section 8 with a one-paragraph statement of what success means in 90 days, with a name on it (yours and one stakeholder). Most teams skip this. The signed paragraph is the difference between a feature that gets retroed and a feature that gets quietly deprecated. Don't ship without one.
Interview questions
Test your knowledge
Common mistakes writing a capstone-grade AI PRD
Key takeaways
- A capstone-grade AI PRD has 8 sections filled with production content, not headings — problem & users · scope · prompts · data & retrieval · eval & gate · rollout · risks & fallbacks · metrics & success.
- Two-threshold launch gates are non-negotiable. Average pass-rate AND a zero-tolerance subset for the highest-cost error class — AND'd together, not OR'd. Single-number gates hide rare disasters.
- Allow-list before prompt. Section 4 (data) decides what the model can know; section 3 (prompt) decides what it can say. Loose allow-list + tight prompt = compliance failure.
- Pin the model by SHA. A name is not a version. Without a pin, every passing eval has a one-day shelf life and vendor-side regressions become invisible.
- Refusal-rate needs a floor, not just a ceiling. The dangerous failure mode is the model becoming too helpful — that fails silently unless you wrote a floor metric.
- Sign the 90-day success paragraph. Two names on a one-paragraph statement of what success means is the difference between a retroed feature and a quietly deprecated one.
- The AI-ness of an AI PRD lives in the content of sections 3, 4, 5 and 7 — not in the section names. A great AI PRD reads like a great PM doc on the surface.
You can now write a production AI PRD.
Now review the eval harness that proves it ships.
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 →