skill-forge

中文版

A Claude Code plugin that turns skill creation, discovery, iteration, and optimization into a skill itself — a meta-system for skills.

Why

Claude Code skills solve "how to codify workflows into reusable slash commands." But three gaps remain:

Gap	skill-forge's Answer
Don't know when to create a skill	Auto-detects complex tasks, proactively asks
Don't know if a skill is well-written	Built-in 5-dimension evaluator, won't save below threshold
Don't know if a skill will actually trigger	Dedicated description optimization phase, eval-driven

Install

/plugin marketplace add nekocode/skill-forge
/plugin install skill-forge

Commands

Command	What it does
/scan [prompt]	Scan project for skill opportunities. Optional prompt as focus hint
/create <prompt>	Create a new skill from prompt. Name auto-derived
/improve <prompt>	Iterate existing skill from prompt. Target matched from registry
/list	Print skill registry with versions, scores, trigger rates

Auto mode: After complex tasks (5+ tool calls, error recovery, user corrections), the Stop hook detects the pattern and offers to create a skill — no manual invocation needed.

How It Works

Design Principles

1. Hermes Agent — Autonomous creation with concrete trigger conditions; patch over rewrite
2. planning-with-files — File system as persistent working memory (context window = RAM, files = disk)
3. Anthropic skill-creator — Eval-driven quality: description is a separate optimization problem, 20-case trigger evals, explain why not just what

Dual-File Security Model

External content (grep/glob/read output) goes to skill_insights.md (low trust, hooks don't read it). Only after validation does content get promoted to skill_draft.md (high trust, injected by hooks). This prevents prompt injection amplification.

Hooks Architecture

Skill-scoped hooks (SKILL.md frontmatter) — only active when skill-forge is engaged:

• UserPromptSubmit — Inject draft header into attention window
• PreToolUse — Re-read draft before each tool call (prevent goal drift)
• PostToolUse — Prompt draft status update after Write/Edit
• Stop — Check for unprocessed skill opportunities

Global hooks (hooks/hooks.json, auto-registered by plugin system):

• SessionStart — Reset counters + inject skill inventory
• PostToolUse — Tool counting + registry update on SKILL.md writes
• Stop — Detect complex workflows, trigger auto mode
• PreCompact — Mark compact state to prevent false positives
• UserPromptSubmit — Keyword matching for skill creation prompts

Skill Lifecycle

Complex task completed
  -> Stop hook / manual invocation
  -> scan -> create (draft -> research -> SKILL.md -> eval >= 6/8)
  -> .claude/skills/<name>/SKILL.md
  -> improve (diagnose -> content patch / trigger eval loop -> changelog + version bump)
  -> repeat after real usage

Session Catchup

On each new session, skill_catchup.py scans the previous session's JSONL for uncaptured complex tasks (5+ tool calls after last draft write). Solves "forgot to save as skill yesterday."

Evaluation Criteria

Dimension	Max	Checks
Trigger quality	3	Complex scenarios? Pushy coverage? Do NOT use? Under 250 chars?
Step clarity	3	Concrete actions per step? Explains why, not just what?
Completeness	2	Prerequisites / verification / notes?
Discriminability	bonus	Assertions pass both with and without skill -> no discriminability, rewrite

Minimum score to save: 6/8.

Description Writing Rules

1. Complex scenarios, not simple verbs — "Use when adding a new REST endpoint that requires route registration, Zod schema, test file, and index.ts update" not "Generate API endpoints"
2. Pushy coverage — Cover cases where users won't name the skill explicitly
3. Do NOT use when — Prevent trigger overlap with related skills

Comparison

Feature	Hand-written SKILL.md	Anthropic skill-creator	skill-forge
Auto-discover opportunities	-	-	scan
Content quality evaluation	-	eval viewer	5-dim evaluator
Description trigger optimization	-	run_loop.py	improve
Persistent working memory	-	-	draft/insights files
Cross-session memory	-	-	catchup.py
Scoped hooks (no global pollution)	-	-	frontmatter hooks
Injection defense	-	-	dual-file isolation
Self-iteration	-	-	improve skill-forge