Writing PRDs (Executable)

This skill turns messy ideas into shippable artifacts:

1. PR/FAQ (customer-first narrative)
2. PRD (scope + requirements + success metrics)
3. AI Eval Spec (requirements as executable evals) — AI only
4. Prompt Set / Prototype plan — AI only

Keep outputs crisp, concrete, and copy‑pasteable. Prefer bullets, tables, and numbered requirements over long prose.

When to use

Use this skill when the user requests:

• a PRD / product requirements / product spec / requirements doc
• a PR/FAQ or “working backwards” narrative
• acceptance criteria, rollout plan, risks, dependencies
• AI/LLM behavior requirements, evals, judge rubric, prompt sets

When NOT to use (redirect)

Do NOT use this skill as a substitute for:

• Company strategy / vision (use a vision/strategy skill)
• Roadmap prioritization across many initiatives (use prioritization/roadmap skill)
• Pure UX copywriting or marketing launch copy (use messaging/launch skills)

If legal/privacy/security constraints are material but unknown, first ask for constraints and the approval process.

Quick decision: which artifact(s) to produce

If the user doesn’t specify, choose the minimal artifact that unlocks the next decision:

• Early / unclear idea → PR/FAQ first, then a lightweight PRD outline.
• Clear direction, need alignment / build plan → PRD.
• AI/LLM feature → PRD + Eval Spec + Prompt Set (required).
• “Improve this existing PRD/spec” → rewrite + gaps + open questions.

Templates:

• PRD:

  assets/PRD_TEMPLATE.md

• PR/FAQ:

  assets/PRFAQ_TEMPLATE.md

• Eval Spec:

  assets/EVAL_SPEC_TEMPLATE.md

• Prompt Set:

  assets/PROMPT_SET_TEMPLATE.md

Quality checks:

• Checklist:

  references/QUALITY_CHECKLIST.md

• Rubric:

  references/QUALITY_RUBRIC.md

• Question bank:

  references/QUESTION_BANK.md

Intake: ask only what you must

Ask at most 5 high-leverage questions at a time (use

references/QUESTION_BANK.md

). If the user can’t answer, proceed with explicit assumptions.

Minimum inputs to proceed:

1. Product/feature name + target user
2. Problem statement (what pain / why now)
3. Success metric(s) or a proxy
4. Constraints (time, platform, legal, data, dependencies)
5. Audience for the doc (exec decision? eng build? cross-functional alignment?)

Output contract (always include)

Always include:

• Version + date + owner (or “TBD”)
• Goals and non-goals
• Requirements as numbered bullets (R1, R2…)
• Clear “Out of scope”
• Open questions (with owner / next step)

When producing multiple artifacts, output in this order:

1. PR/FAQ → 2) PRD → 3) Eval Spec → 4) Prompt Set

Workflow A: Draft a PR/FAQ (customer-first)

Use when the idea is early or benefits are fuzzy.

1. Write the press release in external-facing, factual language.
2. Add FAQs for: customer, internal stakeholders, technical/ops, risks.
3. Include a hypothetical launch date (forces planning).
4. Extract “claims” from the PR (what must be true) → convert to requirements.

Workflow B: Draft or rewrite a PRD (shippable requirements)

1. Start with a 5–10 line narrative (why this matters, for whom).
2. Define scope boundaries: goals, non-goals, assumptions.
3. Define users + key journeys (happy path + edge cases).
4. Translate into requirements:
   • Functional requirements (R1…)
   • Non-functional requirements (performance, reliability, accessibility, privacy)
5. Define success metrics + instrumentation.
6. Rollout plan + risks + mitigations.
7. End with open questions + decision log.

Workflow C (AI only): Requirements as executable evals

Goal: make requirements testable.

1. List non‑negotiable behaviors (MUST / MUST‑NOT).
2. Build a test set: golden, adversarial, regression.
3. Write an LLM‑as‑judge rubric (scale + definitions + pass threshold).
4. Connect evals to shipping gates (pre‑launch + post‑launch monitoring).

Workflow D (AI only): Prompt set + prototype plan

1. Provide a small prompt set that exercises core use cases + edge cases.
2. Specify expected outputs and guardrails.
3. Propose a minimal prototype plan if useful (what to mock vs build).

Quality bar (before finalizing)

Run:

• references/QUALITY_CHECKLIST.md

• references/QUALITY_RUBRIC.md

Also include a brief circulation plan: who should review (Eng, Design, Data, Support, Marketing, Legal) and what feedback you need from each.

Confidentiality

Treat user-provided information as confidential. Do not include secrets in outputs. If something looks sensitive, redact or replace with placeholders.