pyskillet 0.1.2

Evaluation-driven Claude Code skill development

skillet

Evaluation-driven Claude Code skill development.

Install

pip install skillet

Why

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.

Workflow

1. Capture gaps with /gap

When Claude does something wrong, capture it:

> Write a code review comment for this SQL query...

● This code has a SQL injection vulnerability...

> /gap

● What did you expect instead?

> Should start with **issue** (blocking): using conventional comments format

● Gap saved to ~/.skillet/gaps/conventional-comments/001.yaml

2. Run baseline eval

skillet eval conventional-comments

Eval Results (baseline, no skill)
==================================
Gaps: 5
Iterations: 3 per gap
Total runs: 15

Pass rate: 0% (0/15)

3. Create the skill

skillet new conventional-comments

Found 5 gaps for 'conventional-comments', drafting SKILL.md...

Created plugins/conventional-comments/
├── .claude-plugin/
│   └── plugin.json
└── skills/conventional-comments/
    └── SKILL.md (draft from 5 gaps)

4. Eval with skill

skillet eval conventional-comments --skill plugins/conventional-comments

Eval Results (with skill)
=========================
Skill: plugins/conventional-comments
Gaps: 5
Iterations: 3 per gap
Total runs: 15

Pass rate: 80% (12/15)

5. Iterate

Edit SKILL.md, re-run eval, repeat.

Commands

skillet eval <name>                    # baseline eval (no skill)
skillet eval <name> --skill <path>     # eval with skill
skillet new <name>                     # create skill from gaps

Gaps

Gaps are stored in ~/.skillet/gaps/<name>/:

# ~/.skillet/gaps/conventional-comments/001.yaml
timestamp: 2025-01-15T10:30:00Z
prompt: "Write a code review comment for..."
actual: "This code has a SQL injection..."
expected: "Should start with **issue** (blocking):"

Roadmap

See ROADMAP.md for future ideas and planned features.

License

MIT