pattern-learning

Pattern Learning Skill

Automatically extract and codify patterns from corrections.

What This Skill Does

1. Monitors Corrections

   • Tracks when user edits Claude-generated code
   • Categorizes correction type
   • Records frequency and context

2. Pattern Detection

   • Triggers when pattern appears 3+ times
   • Validates consistency across occurrences
   • Checks for false positives

3. Rule Generation

   • Creates .claude/rules/{category}/{pattern}.md
   • Includes code examples from actual corrections
   • Sets appropriate priority and tags
   • Adds to CLAUDE.md "Common Patterns"

4. Continuous Learning

   • Updates rule frequency count
   • Adjusts priority as pattern becomes more important
   • Retires rules that stop being issues

Examples

Correction Pattern: Missing rate limiting in API endpoints

Session 5: Claude creates endpoint without @rate_limit
         → User adds: @rate_limit(100, 60)

Session 9: Same correction
Session 14: Same correction again (3rd time)

System generates:
- .claude/rules/backend/api-rate-limiting.md
- Adds to CLAUDE.md: "API endpoints must include rate limiting"
- Next session: Claude automatically includes rate limiting

Output

## New Patterns Learned

### Pattern 1: API Rate Limiting
- Occurrences: 3
- Generated rule: `.claude/rules/backend/api-rate-limiting.md`
- Status: Active (applied to future sessions)

### Pattern 2: React Component Memoization
- Occurrences: 2 (monitoring)
- Status: Pending (needs 1 more occurrence)

Configuration

Tracking file: .claude/.patterns.json

{
  "patterns": {
    "api_rate_limiting": {
      "count": 3,
      "first_seen": "2026-01-05",
      "last_seen": "2026-01-14",
      "rule_generated": true,
      "rule_file": ".claude/rules/backend/api-rate-limiting.md"
    },
    "react_memo": {
      "count": 2,
      "first_seen": "2026-01-12",
      "last_seen": "2026-01-14",
      "rule_generated": false
    }
  }
}

Trigger

• PostToolUse hook: After Edit tool use
• Manual: /learn/from-session
• Auto: After each session when patterns emerge

Performance

• Detection: < 5 seconds per correction
• Rule generation: < 10 seconds
• Total overhead: < 2% of session time

Success Metrics

• ✅ Patterns detected by 3rd occurrence
• ✅ 80%+ accuracy in pattern categorization
• ✅ Generated rules useful (positive feedback)
• ✅ No false positive rules created
• ✅ Rules improve over time