Self-improving Claude Code skills.
Correct Claude once. Run /skillsmith. The lesson is folded back into the skill's SKILL.md — after you approve the diff. Your skills get sharper every time you use them.
You invoke a skill. Claude gets it 90% right. You correct it — "no, always use pnpm here", "stop adding comments", "match the existing test style". It fixes it. Then you close the session and the lesson is gone. Next time, same correction.
The fix you keep meaning to do: ask Claude to update the SKILL.md so it remembers. skillsmith automates that last step — safely.
Next time you use the skill, it already knows.
- 🔍 Diff-approve, never silent. Every edit is shown as a diff and written only on your explicit yes.
- 🪶 Keeps skills lean. Tightens an existing rule instead of piling on new ones, and flags contradictions instead of stacking them.
- ↩️ Reversible. Keep your skills in git — every update is one revertible commit.
- 🎯 No noise. One-off task details are ignored. Only durable, reusable lessons change the skill.
Run each command separately (one per prompt — don't paste both at once):
-
Add the marketplace:
/plugin marketplace add https://github.com/HarshMehta112/skillsmith.git -
Install the plugin:
/plugin install skillsmith@skillsmith -
Reload so it's available:
/reload-plugins
Run /skillsmith at the end of any task. It:
- Finds every skill used in the current session.
- Scans for corrections — rejected output, added constraints, repeated instructions — and attributes each to the skill it belongs to.
- Proposes a concise
SKILL.mddiff per skill — preferring to tighten existing rules, flagging any contradictions. - Asks yes/no for each skill — and edits that file only on yes.
Used several skills in one session? Each gets its own diff and its own approval. No corrections found? It says so and changes nothing.
| File | Role |
|---|---|
commands/skillsmith.md |
the /skillsmith reflection command |
That's the whole plugin — one command, no background hooks, no state on disk. Easy to read, easy to trust.
- Automatic reflection across sessions — stage a finished session on exit, offer to review it at the next session start.
- Target a specific skill by name.
- Batch-reflect across several past sessions.
Contributions welcome — open an issue or PR.