distill-feedback - Turn captured user corrections into durable rules
Processes queued session feedback to semantically extract user corrections and merge approved additions into durable rules.
Tags
Updated: 2026-09-20Capabilities
Typical Inputs
Typical Outputs
What this skill does
- Extract queued feedback sessions
- Detect durable user corrections
- Draft deduplicated atomic rules
- Request authorization for changes
- Apply updates using delta merge
- Mark reconciled sessions as processed
Inputs
- Queued feedback in ~/.claude/feedback/queue.jsonl
- Session transcript archives
- Existing rules and memory files
- User change authorization decisions
Outputs
- Updated rule files in ~/.claude/rules/
- Updated project memory files
- Appended entries in processed.jsonl
- Proposed changes review table
Requirements
- Python runtime environment
- extract_feedback_queue.py script
- LLM sub-agent execution capability
- Read and write access to ~/.claude/ directory