skillet

Evaluation-driven Claude Code skill development.

Three levels of documentation:

1. This README — the concise overview, mirroring the structure below
2. docs/ — full markdown documentation, shipped with the package
3. skillet.run — the rendered docs site, 1:1 with docs/

pip install pyskillet

Anthropic recommends building evaluations before writing skills:

Create evaluations BEFORE writing extensive documentation. This ensures your Skill solves real problems rather than documenting imagined ones.

But they don't provide tooling:

We do not currently provide a built-in way to run these evaluations.

skillet fills that gap.

Capture failures with /skillet:add in Claude Code, then run the loop:

skillet eval my-skill                              # baseline
skillet create my-skill                            # generate skill from evals
skillet eval my-skill ~/.claude/skills/my-skill    # eval with skill
skillet tune my-skill ~/.claude/skills/my-skill    # iteratively improve

Skillet captures failures, runs systematic evaluations, and iterates on skills with quantitative feedback. → docs/index.md

End-to-end walkthrough from your first capture to a tuned skill. → docs/getting-started.md

Skillet evaluates skills (instructions that shape behavior), not agents (the underlying capability). → docs/concepts/skills-vs-agents.md

The same eval format serves two purposes: capability (pass@k, exploratory) during development, regression (pass^k, strict) in CI. → docs/concepts/capability-vs-regression.md

A good eval suite needs negative cases — prompts where the skill should not trigger — to catch overtriggering. → docs/concepts/balanced-problem-sets.md

Use /skillet:add in Claude Code to record failures as YAML eval files. → docs/guides/capture-with-slash-command.md

skillet lint <path> checks a SKILL.md against 14 rules covering naming, frontmatter, body length, and recommended fields. → docs/guides/linting.md

Development setup, testing strategy, code style, and PR conventions. → docs/guides/contributing.md

skillet ships with eval, create, tune, lint, and generate-evals. → docs/reference/cli.md

YAML schema for eval files: required name/prompt/expected, optional domain/setup/teardown. → docs/reference/eval-format.md

Programmatic interface: evaluate(), tune(), create_skill(), generate_evals(), lint_skill(). → docs/reference/python-api.md