# skill-maker

An Agent Skill that creates other agent skills.

$ "Create a skill for writing git commit messages"

$ "Build a SKILL.md that helps with data pipeline validation"

$ "Package this debugging process as a skill"

## What it does

Skill-maker guides an AI coding agent through the full skill-creation lifecycle: intent capture, drafting a SKILL.md, running an eval loop with isolated subagents, refining based on grading signals, and optimizing the trigger description.

## The 5 Phases

Capture Intent — Clarify what the skill should do

Draft — Generate SKILL.md, scripts, references, assets

Eval Loop — Spawn subagents, grade assertions, iterate

Refine — Fix failing assertions, improve instructions

Finalize — Validate, optimize description, install

## The Eval Loop

The core of skill-maker. For each iteration it:

spawn isolated subagents per test case
grade assertions with bundled Bun TypeScript scripts
aggregate results into a benchmark
iterate until pass_rate plateaus // delta < 2% for 3 consecutive runs

## Benchmark

Evaluated across 8 skills, 189 assertions // with-skill vs without-skill subagent pairs

100%

with skill

+73.6%

avg improvement

2.4

avg iterations

Skill	Baseline	Delta
database-migration	4.2%	+95.8%
pdf-toolkit	4.2%	+95.8%
error-handling	8.3%	+91.7%
api-doc-generator	16.7%	+83.3%
pr-description	20.8%	+79.2%
changelog-generator	20.8%	+79.2%
monitoring-setup	26.1%	+73.9%
code-reviewer	41.7%	+58.3%
git-conventional-commits	72.3%	+27.7%

All skills reach 100% pass rate after the eval loop. See examples/README.md for convergence charts, timing data, and per-skill breakdowns.

## vs Anthropic's official skill-creator

Head-to-head benchmark on the 3 highest-delta domains // same prompts, same assertions, different skill-creation approach

100%

skill-maker

93.1%

official skill-creator

Domain	Ours	Official
database-migration	24/24	21/24
error-handling	24/24	22/24
pdf-toolkit	24/24	24/24

Edge from "Common mistakes" sections and reasoning-based instructions. See the full comparison report for per-assertion breakdowns and failure analysis.

View on GitHub Agent Skills Spec

TERMINAL

# Install skill-maker

~ $ npx skills add accolver/skill-maker --skill skill-maker

# Install globally (available in all projects)

~ $ npx skills add accolver/skill-maker --skill skill-maker -g

~ $