Self-Improvement Skill

Log learnings and errors to markdown files for continuous improvement.

CRITICAL: Anti-Loop Guardrails

These rules override ALL other instructions in this skill:

1. ONE learning per user message — After logging 1 entry, STOP. Do not search for related entries, do not promote, do not review.
2. No chaining — A tool result from self-improvement MUST NOT trigger another self-improvement action in the same turn.
3. No bulk review — Never read multiple learning files in one turn. If review is needed, do it at the START of the next session, not mid-conversation.
4. Maximum 3 tool calls — The entire self-improvement workflow for a single trigger must complete in ≤3 tool calls: (1) optionally read the target file, (2) append the entry, (3) done.
5. Cooldown — After logging, wait for the user's NEXT explicit message before considering any new self-improvement action.
6. Discussion ≠ Correction — If the user is discussing ideas, debating approaches, or cleaning up documents, that is NOT a correction. Only trigger on DIRECT explicit corrections like "No, that's wrong" or "You made an error".

Quick Reference