skill-forge 🔨

A self-learning loop for Claude Code — your agent turns finished sessions into reusable skills, so it gets sharper every session.

Inspired by Hermes Agent's reflective loop and Anthropic's skills-first practice, built entirely on Claude Code hooks — no external service, no server.

Honest scope. skill-forge does not rewrite your application code. It grows your agent's skill library — the procedures Claude reuses — so day 30 is sharper than day 1. The thing that self-improves is the agent, not your codebase.

The loop

every session ends
   │
 SessionEnd hook ── ≥5 tool calls? ──► queue the transcript     (no model call; cheap)
   │
 next session starts
   │
 SessionStart hook ──► "📌 N sessions queued → /skill-forge:forge"
   │
 /skill-forge:forge ──► reflect:  new skill · "Got X" · refine   + dedup + prune
   │
 apply gate ──► low-risk: auto   ·   high-risk (new skill / safety / deletion): you approve

Claude Code already loads the right skill before a task. skill-forge automates the other half — writing it.

Install

/plugin marketplace add Refusned/skill-forge
/plugin install skill-forge@refusned

Then /reload-plugins (or restart). The SessionEnd/SessionStart hooks register automatically — no settings.json editing.

Try it: after your next real session, run /skill-forge:forge.

How it works

Stage	What happens
Capture	A SessionEnd hook counts tool calls; sessions with ≥5 get queued to ${CLAUDE_PLUGIN_DATA}/inbox.jsonl. Trivial chats are skipped. No model is invoked — it's a few lines of shell.
Nudge	A SessionStart hook reminds you when sessions are waiting. Silent when there's nothing to do.
Reflect	/skill-forge:forge reads the queued transcripts (and the current session from context), finds repeated patterns, recurring fixes, and refined commands.
Curate	Before creating anything, it dedups against your existing skills, disambiguates co-triggering descriptions, and flags stale instructions.
Apply (gated)	Low-risk edits (a "Got X" line, a description tweak) apply automatically. High-risk changes (a new skill, any safety-invariant edit, a deletion) wait for your approval.

Safety by design

• Gated autonomy. It captures automatically, but never silently creates a skill, edits a safety invariant, or deletes anything — those ask first. Tune the gate in the skill.
• No secret leakage. Transcripts can contain pasted tokens; the skill never copies a literal secret into a skill file.
• No sprawl. A conservative ≥5 tool calls threshold skips noise, and the dedup step extends existing skills instead of spawning duplicates.

Configuration

• Threshold — edit ≥5 in scripts/session_end_capture.sh.
• Queue location — ${CLAUDE_PLUGIN_DATA}/inbox.jsonl (per-plugin data dir; falls back to ~/.claude/skill-forge/).
• Autonomy — v0.1 ships auto-capture + manual approval. A scheduled fully-autonomous mode (cron/Routine, like Hermes 24/7) is on the roadmap.

Credits

• Reflective-loop concept: Hermes Agent by Nous Research.
• Skills-first practice & hooks: Anthropic (Claude Code).

License

MIT © Roman Barmin (@Refusned)

---

🇷🇺 По-русски

Самообучающаяся петля для Claude Code: агент сам превращает завершённые сессии в переиспользуемые навыки и умнеет с каждой сессией.

Вдохновлено рефлексивной петлёй Hermes Agent (Nous Research) и подходом «skills-first» от Anthropic, собрано на хуках Claude Code — без внешнего сервиса.

Честно: скил не переписывает твой прикладной код. Растёт библиотека навыков агента — процедуры, которые Claude переиспользует. Саморазвивается агент, а не кодовая база.

Как ставить:

/plugin marketplace add Refusned/skill-forge
/plugin install skill-forge@refusned

Хуки регистрируются сами — settings.json руками править не нужно.

Петля: SessionEnd-хук копит сессии (≥5 тул-коллов) → SessionStart-нудж → /skill-forge:forge (рефлексия → новый скил / «Грабли» / уточнение + дедуп) → гейт: low-risk авто, high-risk (новый скил / правка money-safety / удаление) — с твоим аппрувом.