PRD Skill

Produce drift-proof Product Requirements Documents through iterative discovery.

Output: .claude/reference/PRD.md (living requirements document for the project).

Do NOT load when

• user has a final spec and only wants implementation planning (use /orchestrate or /prd --to-plan)
• task is a code review or repo health check
• request is a lightweight note or issue, not a full requirements document

NEVER

• Never generate a PRD without scanning the codebase first — requirements that contradict existing architecture produce unimplementable specs
• Never write acceptance criteria as "should work correctly" — use Given/When/Then; untestable criteria cannot be verified at phase completion
• Never skip dependency analysis — untracked dependencies produce phases that can't be parallelized safely
• Never batch all discovery questions into one wall of text — max 4 questions per round; users disengage from interrogations
• Never assume a library version — check package.json and npm; version assumptions produce broken phase plans
• Never write phasing without dependency arrows — phases must be a DAG; implicit ordering creates unmergeable parallel work
• Never leave [NEEDS CLARIFICATION] markers in a finalized PRD — they signal a spec that cannot drive implementation

Mode routing

Create mode

Phase 1 — Codebase scan (automatic, before any questions)

Launch an Explore agent to map relevant codebase areas. Read roadmap/changelog for prior decisions. Check for outdated dependencies in scope. This prevents asking questions the codebase already answers.

Phase 2 — Interactive discovery (2–4 rounds, max 4 questions each)

• Round 1: Problem, who, success criteria, personas
• Round 2: MoSCoW priorities, explicit out-of-scope, interactions
• Round 3: Architecture constraints, database, auth, performance
• Round 4 (if needed): Phasing, parallelization risks

Never draft before Round 2. Never ask what the codebase scan already answered.

Phase 3 — Draft

Read template.md. Write to .claude/reference/PRD.md. Mark gaps [NEEDS CLARIFICATION: ...]. Populate Architecture Decisions (AD-N entries). Build phasing DAG with explicit arrows.

Phase 4 — Validation

Read validation.md, run every check. Present pass/fail. Iterate until critical gates pass.

Phase 5 — Finalize

Remove all markers. Commit: docs(prd): add <feature-name> requirements.

Update mode (--update)

1. Read existing PRD
2. Ask what changed (new requirement, scope change, dependency update)
3. Launch Explore agent to detect codebase drift since last PRD update
4. Generate diff-style updates: [ADDED], [CHANGED], [REMOVED], [DRIFT DETECTED]
5. Present for approval, apply, re-validate

Plan mode (--to-plan)

Transforms PRD phasing into an /orchestrate-ready task plan at docs/plans/<feature-name>-plan.md.

Wave assignment rules:

• Wave 1: Schemas, types, migrations, config (foundation)
• Wave 2: Routes, services, repositories (implementation)
• Wave 3: Wiring, UI, end-to-end flows (integration)
• Wave 4: Error handling, edge cases, docs (polish)

Agent assignment rules:

• haiku: Type definitions, config, simple CRUD, test writing
• sonnet: Business logic, complex integrations, security-sensitive code

Requires V5 (DAG) validation to pass before generating the plan.

Validate mode (--validate)

Read validation.md and run all gates. Print pass/fail with line references.

Audit-deps mode (--audit-deps)

Read drift-prevention.md and run dependency freshness + architectural drift checks.

Scripts

bash scripts/check-prd-assets.sh skills/prd
node scripts/list-prd-phases.js .claude/reference/PRD.md

Arguments

• /prd — interactive creation
• /prd Workflow Designer — start with context
• /prd --update — incremental update
• /prd --validate — run validation only
• /prd --audit-deps — dependency audit only
• /prd --to-plan — generate orchestrate-ready task plan