convez.ai
Skill
surahli123

autorefine-skill-improvement

surahli123/autorefine-skill-improvement

Ver no GitHub
BomPluginScore · de 0 a 100

Karpathy's autoresearch + Hamel's Three Gulfs, applied to Claude Code skills. Iteratively improve any skill from zero evals to optimized and validated.

Instalação

$npx convez-ai surahli123/autorefine-skill-improvement

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
10
Forks
1
Acompanhando
0
Saúde
42%
Último commit
há 3 semanas
Versão
—
Issues abertas
2
Contribuidores
2

README

A tradução para o português deste README ainda está na fila.

AutoRefine

Karpathy's autoresearch + Hamel's Three Gulfs, applied to Claude Code skills.

AutoRefine is a guided workflow for improving a SKILL.md file from vague instructions to eval-grounded, trust-checked, iterative optimization. It is designed for skill authors who want more than prompt tweaking: it adds human error analysis, judge validation, and a disciplined mutation loop before claiming a skill got better.

Why this exists

Most skill iteration tools are strong at fast optimization but weak at proving that the optimization is real.

AutoRefine's core idea is simple:

  1. Comprehend what the skill actually does wrong by reading outputs.
  2. Specify judges grounded in those observed failures and validate them.
  3. Generalize only after the eval surface is trustworthy.

That turns “prompt iteration” into a more reliable research loop.

Install

Clone the repo and copy the shipped bundle into your Claude Code skills directory:

git clone https://github.com/surahli123/autorefine-skill-improvement.git
cp -r autorefine-skill-improvement/autorefine ~/.claude/skills/autorefine

If you want the bundle-level file breakdown, see autorefine/README.md.

The repo also includes .claude-plugin/ and .codex-plugin/ metadata for plugin/discovery surfaces. Both plugin manifests mirror the same public package identity and point at the shipped runtime bundle under ./autorefine.

Usage

/autorefine /full/path/to/your-skill

AutoRefine creates a workspace, copies your target skill into it, and then guides you through the full refinement pipeline.

Phase 1 audits the skill across gotchas, voice, progressive disclosure, anti-railroading, description quality, and scripts before later gulfs build evals from those findings.

What you should read next

What is shipped

The installable runtime lives under autorefine/. That bundle contains:

  • SKILL.md
  • references/gulf1-comprehension.md
  • references/gulf2-specification.md
  • references/gulf3-generalization.md
  • references.md
  • dashboard.html
  • scripts/record.py
  • scripts/records-to-gulf1.py
  • validate-host.sh

Repo-maintenance tests and internal design material are intentionally kept outside the shipped bundle.

The optional Phase 0.5 live-trace recorder path uses scripts/record.py, which requires Python with httpx available in the runtime environment. The converter script is stdlib-only.

Public vs internal repo surface

This public repo is user-facing:

  • autorefine/ is the shipped bundle
  • docs/ is curated user documentation

Internal engineering material lives in the dev submodule and is not part of the public install surface.

License

MIT

Ver o original no GitHub

Bom para

Relacionados