Every time you and Claude clash (a misunderstanding, a wrong assumption, a preference Claude didn't know about), the lesson disappears when the conversation ends. Next session, same mistake.
This skill fixes that. It watches for friction, extracts the lesson, and writes it down as a concrete rule that persists across sessions. Over time, Claude stops repeating mistakes and starts anticipating how you work.
Friction happens → lesson extracted → rule proposed to you → you approve → saved as a skill patch.
That's it. No magic. Just a systematic way to turn "Claude messed up" into "Claude won't do that again."
meta-learner/SKILL.md # the engine: friction detection + learning loop
Once running, the meta-learner creates and patches other skills automatically:
~/.claude/skills/
meta-learner/SKILL.md
your-defaults/SKILL.md + PATCHES.md
[auto-created-domain-skills]/...
cp SKILL.md ~/.claude/skills/meta-learner/SKILL.md- Patch, don't edit. Lessons go to
PATCHES.md, originals stay clean. Patches override when loaded. - Two files max per skill.
SKILL.md+PATCHES.md. No fragmentation. - Flat directories. No nesting. Name says what it is.
- Always ask. Never auto-saves a lesson. You confirm every rule.
- Works on claude.ai too. Same loop, but stores lessons as memory edits instead of files.
MIT