---

name: compound description: | This skill should be used when the user says "/compound", "compound this", "document learnings", "save what we learned", or after completing a PR. Extracts knowledge from PR context and saves to docs/learnings/. allowed-tools:

• Read
• Grep
• Glob
• Bash
• Write
• Edit
• AskUserQuestion

---

Compound Skill

Extracts knowledge from PR context and saves structured documentation to docs/learnings/.

Workflow

Phase 1: Context Collection

1. Identify PR number/branch

   • Use PR number if provided as argument
   • Otherwise, find PR from current branch: gh pr view --json number,body,title
   • If no PR exists: Prompt user to enter PR number directly or confirm proceeding without PR

2. Extract Plan path

   • Find Plan path pattern in PR body: .dev/specs/{name}/PLAN.md
   • Regex: \.dev/specs/[^/]+/PLAN\.md
   • If no Plan path found: Prompt user to enter spec name directly or select from .dev/specs/ directory listing

3. Derive Context path

   • Extract spec name from Plan path
   • Context directory: .dev/specs/{name}/context/

4. Parallel collection (run following commands simultaneously, skip if files don't exist)

   # Context files (treat as empty if not found)
   cat .dev/specs/{name}/context/learnings.md 2>/dev/null || echo ""
   cat .dev/specs/{name}/context/decisions.md 2>/dev/null || echo ""
   cat .dev/specs/{name}/context/issues.md 2>/dev/null || echo ""
   
   # PR comments and reviews (collect as JSON for stability)
   gh pr view {pr_number} --json comments,reviews
   

Error Handling:

• If no context files exist AND no PR comments → Notify user and request manual input
• At least 1 source required to proceed with document generation

Phase 2: Knowledge Extraction & Classification

2.1 Extract Valuable Feedback from PR Comments

Criteria for valuable feedback:

• Code improvement suggestions
• Bug/issue identification
• Pattern/best practice mentions
• "This would be better" type advice
• Comments left with approval

Filter out:

• Simple questions ("What is this?")
• Confirmation requests ("Is this correct?")
• Approval-only comments ("LGTM", "Approved")
• Bot comments

Extraction keywords:

• "suggest", "recommend", "better", "instead"
• "pattern", "practice", "convention"
• "issue", "bug", "fix"
• "learned", "TIL", "note"

Extracted information:

• author
• body
• file_path (if inline comment)
• created_at

2.2 Analyze Context Files

| File | Purpose | |------|---------| | learnings.md | Direct learnings | | decisions.md | Decision rationale | | issues.md | Out of scope issues (for future reference) |

2.3 Synthesize

1. Assess documentation value from collected sources
2. Check for duplicates: Search docs/learnings/
3. Classify problem type - Refer to references/problem-types.md (relative to this skill directory)
4. Generate tags

Phase 3: Document Generation

1. Generate YAML frontmatter

   pr_number: {PR_NUMBER}
   date: {YYYY-MM-DD}
   problem_type: {TYPE}
   tags: [{TAGS}]
   plan_path: {PLAN_PATH}
   

2. Write document using template

   • Template location: templates/LEARNING_TEMPLATE.md (relative to this skill directory)
   • Read template and substitute placeholders

3. Determine filename

   • Format: {YYYY-MM-DD}-{short-title}.md
   • Example: 2024-01-15-api-error-handling.md

4. Save

   • Path: docs/learnings/{filename}.md

5. Add cross-references (if related documents exist)

   • Add new document link to Related section of existing documents

Usage Examples

# Specify PR number
/compound 123

# Use PR from current branch
/compound

Output

Outputs the created document path and summary:

Created: docs/learnings/2024-01-15-api-error-handling.md

Summary:
- Problem Type: error-handling
- Tags: api, typescript, validation
- Sources: learnings.md, 2 PR comments

---

<!-- TODO: Future extensions --> <!-- - [ ] Session ID based user feedback collection --> <!-- - [ ] CLAUDE.md auto-update suggestions --> <!-- - [ ] Detect existing document UPDATEs --> <!-- - [ ] Auto-categorization by problem_type (docs/solutions/{type}/) -->