ls-self-improve

What problem does it solve?

Self-improvement loops often stall because skill descriptions drift or fail to trigger reliably, causing the wrong tool to fire or useful tools to be skipped.

Core Features & Use Cases

• Scans your local skill catalog to inventory each skill’s SKILL.md metadata and whether it already has evals/evals.json.
• Evaluates trigger accuracy by running the skill’s eval set to measure false positives and false negatives.
• Trains and recompiles underperformers by calling /train for low-accuracy skills and optionally recompiling via LiteCLI so help text and matching stay current.

Use it when you want your agent to select the right tools more consistently across repeated sessions, especially after adding new skills or noticing misfires.

Quick Start

Ask your agent to run self-improvement for the worst few skills by saying: "self-improve my skills and retrain the 3 lowest-accuracy skills, then recompile them if possible."