Writing PRDs

Scope

Covers

• Turning a product idea into a decision-ready PRD with unambiguous scope, requirements, and success metrics
• Optionally producing a PR/FAQ (press release + FAQ) to force customer-centric narrative first
• For AI features: adding a Prompt Set + Eval Spec so “requirements” are testable and continuously checkable

When to use

• “Write a PRD / product spec / requirements doc for this feature.”
• “Turn these messy notes into a PRD we can align on.”
• “Create a PR/FAQ and then a PRD.”
• “This is an AI feature; I need evals + prompts to define behavior.”

When NOT to use

• You’re still choosing what strategy/market to pursue (do product vision / strategy first)
• You need discovery from scratch (research plan, problem validation) more than requirements -> use problem-definition
• You need a detailed engineering design doc (APIs, schemas, low-level architecture)
• You’re prioritizing among many initiatives -> use working-backwards or roadmap prioritization first
• You need build-ready interaction specs, user flows, or prototype briefs -> use writing-specs-designs
• You need to define or refresh the success metric / North Star before writing requirements -> use writing-north-star-metrics

Inputs

Minimum required

• Product + target user/customer segment
• Problem statement + why now (what changed, what’s broken, or what opportunity exists)
• Goal(s) + non-goal(s) + key constraints (timeline, policy/legal, platform, dependencies)
• Success metric(s) + 2–5 guardrails (quality, safety, cost, latency, trust)

If it’s an AI feature (additionally)

• What the model/system should do vs must never do (policy + safety)
• Concrete examples of desired and undesired outputs
• How correctness will be evaluated (offline tests, human review, online metrics)

Missing-info strategy

• Ask up to 5 questions from references/INTAKE.md.
• If answers are still missing, proceed with clearly labeled assumptions and provide 2–3 options (scope, metric, rollout).

Outputs (deliverables)

Produce a PRD Pack in Markdown (in-chat; or as files if the user requests):

1. Context snapshot (what decision we’re making, constraints, stakeholders)
2. Artifact selection (PR/FAQ vs PRD vs AI add-ons)
3. PR/FAQ (optional) — customer narrative + FAQs
4. PRD — goals/non-goals, requirements (R1…Rn), UX flows, metrics, rollout
5. AI Prompt Set (if AI) — versioned prompts + examples + guardrails
6. AI Eval Spec (if AI) — acceptance tests + judge prompts + pass/fail criteria
7. Risks / Open questions / Next steps (always included)

Templates: references/TEMPLATES.md

Workflow (8 steps)

1) Decide the artifact set (don’t over-document)

• Inputs: User request + constraints.
• Actions: Choose: PR/FAQ only, PRD only, PR/FAQ → PRD, or PRD + AI add-ons (Prompt Set + Eval Spec).
• Outputs: Artifact selection + rationale.
• Checks: The artifacts match the decision being made and the audience.

2) Intake + clarify decision and success

• Inputs: references/INTAKE.md.
• Actions: Ask up to 5 questions; confirm decision owner, timeline, constraints, and success metrics/guardrails.
• Outputs: Context snapshot.
• Checks: You can state “what we’re deciding” and “how we’ll measure success” in 1–2 sentences.

3) Write the customer narrative first (PR/FAQ or PRD narrative)

• Inputs: Context snapshot.
• Actions: Draft a customer-centric narrative (problem → solution → why now). If using PR/FAQ, draft the press release headline/summary and top FAQs.
• Outputs: Narrative section (and PR/FAQ if selected).
• Checks: A stakeholder can restate the customer benefit and urgency without jargon.

4) Lock scope boundaries (goals, non-goals, out of scope)

• Inputs: Narrative + constraints.
• Actions: Define goals, non-goals, and explicit exclusions; call out dependencies and assumptions.
• Outputs: Scope section(s) in the PRD.
• Checks: “What we are NOT doing” is as clear as what we are doing.

5) Convert scope into testable requirements (R1…Rn)

• Inputs: Goals + user journeys.
• Actions: Write numbered requirements with acceptance criteria, edge cases, and non-functional needs (privacy, latency, reliability). Mark “must/should/could”.
• Outputs: Requirements table/list.
• Checks: An engineer or QA can turn requirements into test cases without asking you to interpret intent.

6) Define UX flows + instrumentation plan

• Inputs: Requirements + current product surfaces/events.
• Actions: Describe key user flows/states; specify success metrics, guardrails, and event/data needs (what to log, where, who owns).
• Outputs: UX/flows section + metrics & instrumentation section.
• Checks: Every goal has at least one measurable metric and a realistic data source.

7) If AI feature: ship prompts + evals as “living requirements”

• Inputs: Requirements + examples.
• Actions: Create a versioned Prompt Set and an Eval Spec (judge prompts + test set + pass thresholds). Include red-team/failure modes.
• Outputs: Prompt Set + Eval Spec drafts.
• Checks: The eval suite can fail when behavior regresses and pass when requirements are met.

8) Quality gate + finalize for circulation

• Inputs: Full draft pack.
• Actions: Run references/CHECKLISTS.md and score with references/RUBRIC.md. Add Risks/Open questions/Next steps.
• Outputs: Final PRD Pack (shareable as-is).
• Checks: Decisions, owners, metrics, and open questions are explicit.

Quality gate (required)

• Use references/CHECKLISTS.md and references/RUBRIC.md.
• Always include: Risks, Open questions, Next steps.

Examples

Example 1 (B2B SaaS feature): “Write a PR/FAQ + PRD for ‘Saved views’ in our analytics dashboard for admins.” Expected: PR/FAQ narrative, a scoped PRD with R1…Rn, metrics/guardrails, and a rollout plan.

Example 2 (AI feature): “Write a PRD + Prompt Set + Eval Spec for an ‘AI email reply’ assistant with brand tone constraints.” Expected: requirements that include safety/brand constraints, a prompt set with examples, and an eval spec with judge prompts + pass/fail thresholds.

Boundary example (redirect): “Create a detailed interaction spec with user flows and acceptance criteria for our checkout redesign.” Response: redirect to writing-specs-designs — this request needs build-ready specs with flows/states and prototype briefs, not a decision-level PRD.

Boundary example (insufficient context): “Write a PRD for ‘make onboarding better’ (no product context).” Response: ask the minimum intake questions; if context remains missing, produce 2–3 scoped options + assumptions and recommend discovery before committing to requirements.

Anti-patterns

Avoid these common failure modes when writing PRDs:

1. Requirements-as-solutions — Writing “build a modal dialog” instead of “user must confirm destructive actions before execution.” Requirements should describe what and why, not how.
2. Missing non-goals — A PRD without explicit non-goals invites scope creep. Every PRD must state what is deliberately excluded from this version.
3. Vanity success metrics — Choosing metrics that always go up (e.g., “total signups”) rather than metrics that reflect actual value delivery. Pair volume metrics with quality guardrails.
4. Spec-level detail in a PRD — Including pixel-level UI descriptions, API schemas, or interaction states that belong in a spec/design doc. Keep the PRD at decision level.
5. Stakeholder alignment theater — Listing stakeholders without clarifying who is the decision-maker (DRI) vs. consulted vs. informed. Ambiguous ownership leads to PRDs that never ship.