pyskillet 0.2.26

Evaluation-driven Claude Code skill development

skillet

Evaluation-driven Claude Code skill development.

Install

pip install pyskillet

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 evals with /skillet:add

When Claude does something wrong, capture it:

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

Claude: This code has a SQL injection vulnerability...

> /skillet:add

Claude: What did you expect instead?

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

Claude: Eval saved to ~/.skillet/evals/conventional-comments/001.yaml

2. Run baseline eval

skillet eval conventional-comments

Eval Results (baseline, no skill)
==================================
Evals: 5
Samples: 3 per eval
Total runs: 15

Pass rate: 0% (0/15)

3. Create the skill

skillet create conventional-comments

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

Created ~/.claude/skills/conventional-comments/
└── SKILL.md (draft from 5 evals)

4. Eval with skill

skillet eval conventional-comments ~/.claude/skills/conventional-comments

Eval Results (with skill)
=========================
Skill: ~/.claude/skills/conventional-comments
Evals: 5
Samples: 3 per eval
Total runs: 15

Pass rate: 80% (12/15)

5. Tune the skill

skillet tune conventional-comments ~/.claude/skills/conventional-comments

Round 1/5: Pass rate 80% (12/15)
  Improving skill...
Round 2/5: Pass rate 93% (14/15)
  Improving skill...
Round 3/5: Pass rate 100% (15/15)
  Target reached!

Best skill saved to ~/.claude/skills/conventional-comments/SKILL.md

Commands

skillet eval <name> [skill]         # run evals (baseline or with skill)
skillet create <name>               # create skill from evals
skillet tune <name> <skill>         # iteratively improve skill
skillet compare <name> <skill>      # compare baseline vs skill from cache
skillet show <name>                 # inspect cached eval results
skillet lint <path>                 # lint a SKILL.md for common issues
skillet generate-evals <skill>      # generate candidate evals from a skill

Evals

Evals are stored in ~/.skillet/evals/<name>/:

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

Python API

import asyncio
from pathlib import Path
from skillet import evaluate

async def main():
    # Baseline (no skill)
    baseline = await evaluate("conventional-comments")
    print(f"Baseline: {baseline['pass_rate']}%")

    # With skill
    result = await evaluate(
        "conventional-comments",
        skill_path=Path("~/.claude/skills/conventional-comments").expanduser(),
    )
    print(f"With skill: {result['pass_rate']}%")

asyncio.run(main())

Tune a skill programmatically:

from skillet import tune
from skillet.tune import TuneConfig

result = await tune(
    "conventional-comments",
    Path("~/.claude/skills/conventional-comments").expanduser(),
    config=TuneConfig(max_rounds=10, target_pass_rate=90.0),
)
print(f"Final pass rate: {result.result.final_pass_rate}%")

See the Python API reference for all functions and options.

Documentation

Full documentation available at the docs site:

• Getting Started
• Concepts — Skills vs agents, capability vs regression, balanced problem sets
• CLI Reference
• Eval Format
• Python API

Roadmap

See ROADMAP.md for future ideas and planned features.

License

MIT