Orchestrator for creating, improving, auditing, and benchmarking Claude Code skills.
-
Clone or download this repo into your skills directory:
git clone https://github.com/pyxl-dev/skillforge.git ~/.claude/skills/SkillForge -
Verify the structure:
~/.claude/skills/SkillForge/ ├── SKILL.md # Main entry point └── Workflows/ ├── Create.md # Create a skill from scratch ├── Improve.md # Improve an existing skill ├── Audit.md # Audit all skills └── Benchmark.md # A/B compare skill versions
skill-creator is Anthropic's official plugin for the skill eval/improve cycle. SkillForge relies on it for creating, testing, improving, and benchmarking skills.
Install it from Claude Code:
/plugin skill-creator@claude-plugins-official
This registers the plugin and makes the skill-creator skill available. You can verify it's installed by running /plugin and checking that skill-creator appears in the list.
Without skill-creator, SkillForge still works in degraded mode — you can run Audit (structural validation) but Create, Improve, and Benchmark workflows won't have access to the eval pipeline.
Any skill that enforces structure conventions (TitleCase naming, flat folders, frontmatter format). Optional if you have your own conventions or want to skip canonicalization.
| Workflow | Trigger | Description |
|---|---|---|
| Create | "forge a skill", "create and validate skill" | Full pipeline: interview, draft, test, iterate, canonicalize |
| Improve | "improve skill X", "fix skill X" | Evaluate current state, improve via feedback, re-canonicalize |
| Audit | "audit skills", "audit all skills" | Scan all skills, produce quality report with scores |
| Benchmark | "benchmark skill X" | A/B comparison with metrics (pass rate, time, tokens) |
In Claude Code, just say:
- "Forge a skill for managing Docker stacks"
- "Improve the Research skill"
- "Audit all my skills"
- "Benchmark the Parser skill"
SkillForge is a pure orchestrator — it delegates everything and duplicates nothing:
- skill-creator handles the content lifecycle (drafting, testing, evaluating, improving)
- Structural validation handles conventions (naming, folder structure, frontmatter format)
- SkillForge routes to the right workflow and chains them in the correct order
MIT