Self-Improving Skill

What to check before installing/use: 1) Understand data flow — observe.py stores full original/final texts locally; improve.py sends excerpts (and the target SKILL.md) to whichever LLM CLI you have (claude/llm or whatever IMPROVE_LLM_CMD points to), so sensitive content can be transmitted to external services. 2) Prefer manual review — use extract/show/apply workflow and avoid 'auto' until you trust proposals; backups and rollback exist but verify they work in your environment. 3) Configure storage paths to a secure location (set SKILL_BASE_DIR / SKILL_LOG_DIR / SKILL_TARGET_PATH) and restrict file permissions. 4) Inspect the scripts before use — there are a few sloppy/incomplete bits (e.g., inconsistent base-path logic and an apparent bug/truncation in auto_improve that may crash), so test on a copy of a non-critical SKILL.md first. 5) If you must analyze private data, run with a local/private LLM or disable automatic networked LLM CLI usage. 6) If you want stronger safety, require human confirmation for all proposals (avoid cron auto apply) and audit proposals before applying.