skill-refinement

Skill refinement collects PostToolUse feedback across Claude sessions and tool contexts (Read, Grep, Glob, Bash) to attribute outcomes to specific skills. Set up a SQLite feedback database at .claude/voyager/feedback.db and a PostToolUse hook at .claude/hooks/post_tool_use_feedback.py, then run voyager feedback insights to surface performance data and improvement recommendations.

Skill refinement semantically attributes tool execution errors and outcomes to skills using a cascade strategy (Transcript, Learned Associations, ColBERT) and optional LLM inference. Run voyager feedback insights --skill <skill-name> --errors to pinpoint error hotspots and get concrete steps to refine SKILL.md and references.

Yes. Skill refinement applies across Claude sessions and tool contexts, storing execution data in schemas for tool_executions, session_summaries, and learned_associations. This persistence enables feedback-driven refinement over time as patterns emerge from repeated tool use.

You need a SQLite feedback database at .claude/voyager/feedback.db, a PostToolUse hook at .claude/hooks/post_tool_use_feedback.py, an insights CLI, and database schemas for tool_executions, session_summaries, and learned_associations. The attribution cascade handles the rest.

Skill refinement uses a three-layer attribution cascade: first checks Transcript context, then Learned Associations from prior feedback, then applies ColBERT semantic matching. Optional LLM-based inference refines attribution when confidence is low, ensuring accurate feedback routing to the responsible skill.

Skill refinement generates insights on error patterns, tool usage frequency, and improvement opportunities tied to specific skills. Output includes concrete steps to update SKILL.md definitions and references, reducing recurring errors and improving tool reliability.