At the end of (or during) a working session, extract only the durable, non-obvious learnings and write them into the right agent-instruction file — CLAUDE.md / AGENTS.md for project-wide facts, an existing or new skill for reusable workflows — while leaving everything else out. Use this whenever the user says things like "update CLAUDE.md with what we learned", "add this to the skill", "remember this for next time", "improve the agent instructions", "make sure future sessions know this", "don't make that mistake again", or at the natural end of a session where the agent was corrected, hit environment quirks, or discovered project conventions. Also use it proactively when …
