Learned Skills Index

How Skills Are Learned

1. Execution: Claude Code runs a task with

   --loop

   enabled
2. Feedback Collection: Each iteration's quality scores and improvements are recorded
3. Pattern Extraction: Successful patterns are extracted from quality-improving iterations
4. Skill Generation: Patterns are compiled into a learnable skill definition
5. Promotion Gate: Skills must meet quality thresholds before promotion

Skill Status

Quality Thresholds for Promotion

• Minimum quality score: 85.0
• Minimum successful applications: 2
• Minimum success rate: 70%

Directory Structure

learned/
├── SKILL.md                    # This index file
├── learned-backend-auth/       # Example learned skill
│   ├── SKILL.md               # Skill definition
│   └── metadata.json          # Machine-readable metadata
└── learned-frontend-form/      # Another learned skill
    ├── SKILL.md
    └── metadata.json

Integration

• Task description keywords
• File types being modified
• Domain context

from core.skill_persistence import retrieve_skills_for_task

skills = retrieve_skills_for_task(
    task_description="your task here",
    domain="backend"  # optional
)

Safety

• Source session ID
• Source repository
• Timestamp of learning
• Quality progression