PRDs & Project Context

Create product requirements and project context that humans and coding assistants can execute effectively.

Two capabilities:

1. PRDs & Specs - Requirements, specs, stories, acceptance criteria
2. Project Context - Architecture, conventions, tribal knowledge (CLAUDE.md)

Modern Best Practices (Jan 2026): Context engineering (right info, right format, right time), decision-first docs, testable requirements with acceptance criteria, metrics with formula + timeframe + data source, cross-tool portability.

Workflow (Use This Order)

1. Pick the deliverable (PRD, AI PRD, tech spec, story map, CLAUDE.md).
2. Gather inputs (problem evidence, users, constraints, dependencies, risks).
3. Fill the template (write decisions first; keep requirements testable).
4. Validate with checklists (requirements, edge cases, security/compliance as needed).
5. Hand off with next actions (implementation plan, owners, open questions).

Docs Folder + LLM Iteration Option (Any Repo)

Use this when a repository has a docs/ folder with:

• research docs prepared for LLM consumption
• feature docs/specs generated by LLMs during implementation

Run this flow before finalizing PRDs/specs:

1. Classify each file by purpose (Tutorial, How-to, Reference, Explanation) to prevent mixed doc types.
2. Tag each non-canonical file with lifecycle metadata (status, owner, last_verified, integrates_into, delete_by).
3. Pick one canonical doc per feature/decision; merge duplicate drafts into it.
4. Convert long research notes into short evidence-backed claims in canonical docs; keep links/dates for external facts.
5. Maintain a compact canonical library for LLMs with root anchors: AGENTS.md (agent instructions) and README.md (human + AI entrypoint), then link deeper specs from docs/.
6. Delete integrated drafts by delete_by date; do not keep .archive/ mirrors in docs/ unless compliance explicitly requires retention.

Quick Reference

PRDs & Specs

Project Context (CLAUDE.md)

Decision Tree

User needs:
    ├─► AI-Assisted Coding?
    │   ├─ Non-trivial (>3 files)? → Planning checklist + agentic session
    │   └─ Simple (<3 files)? → Direct implementation
    │
    ├─► Repo has a docs folder with LLM-generated research/feature docs?
    │   └─ Use Docs Folder + LLM Iteration Option, then validate with qa-docs-coverage
    │
    ├─► Project Onboarding?
    │   ├─ New to codebase? → Generate CLAUDE.md
    │   └─ Quick context? → Minimal CLAUDE.md
    │
    └─► Traditional PRD?
        ├─ Product requirements? → PRD template
        ├─ AI feature? → AI PRD template
        └─ Acceptance criteria? → Gherkin/BDD

Cross-Tool Context Files

CLAUDE.md / AGENTS.md Guidance

• Start minimal: assets/minimal-claudemd.md
• Add only what’s needed: assets/architecture-context.md, assets/conventions-context.md, assets/key-files-context.md, assets/dependencies-context.md, assets/tribal-knowledge-context.md
• Keep it executable: commands must run; include no secrets; prefer file paths over pasted code

Do / Avoid

Do

• Start with executive summary (decision, users, scope, success)
• Define acceptance criteria in testable language
• Keep requirements unambiguous (must/should/may)
• Link to supporting docs instead of pasting

Avoid

• Vague requirements ("fast", "easy") without definitions
• Mixing draft notes and final requirements
• Metrics without measurement plan
• Docs with no owner or review cadence
• Dual-state wording that mixes live behavior, target behavior, and migration behavior in one statement

LLM Ambiguity Gate (Required for planning docs)

• Label every behavior as exactly one of: Live now, Target, or Transition (with owner + end condition).
• Label every metric as either Reference signal or Release blocker.
• Define one canonical feature-gating contract per feature; all other docs must link to it instead of restating variants.
• Keep assumptions/open questions separate from final decisions.
• If conflicts exist across docs, mark one canonical source and add follow-up tasks to resolve mirrors.

Context Extraction

Use:

• references/architecture-extraction.md for components/data flows
• references/convention-mining.md for naming/patterns
• references/tribal-knowledge-recovery.md for git-history “why”
• references/docs-audit-commands.md for audit commands and tool fallbacks

Quality Checklist

PRD Quality

• Clear problem statement
• Measurable success criteria
• Unambiguous acceptance criteria
• Edge cases documented
• AI can execute without clarification
• Every behavior is labeled Live now, Target, or Transition
• Metrics are labeled Reference signal or Release blocker
• Each feature-gating rule has one canonical source (no conflicting duplicates)

CLAUDE.md Quality

• Architecture reflects actual structure
• Key files exist at listed locations
• Conventions match actual patterns
• Commands actually work
• No sensitive information

Resources

Templates

Related Skills

Fact-Checking

• Use web search/web fetch to verify current external facts, versions, pricing, deadlines, regulations, or platform behavior before final answers.
• Prefer primary sources; report source links and dates for volatile information.
• If web access is unavailable, state the limitation and mark guidance as unverified.