convez.ai
SkillProdutividadeIntermediário
jedobe

Skill Evaluator

jedobe/skill-evaluator

Avalia skills do Claude Code com rubrica baseada em dados dos top repositórios

Ver no GitHub
InicialPluginScore · de 0 a 100

Ferramenta que avalia skills do Claude Code contra uma rubrica de 7 dimensões baseada em análise dos repositórios mais bem avaliados do GitHub. Fornece scoring estruturado (0-100 pontos) e tier de qualidade (Elite, Gold, Silver, Bronze), eliminando feedback subjetivo sobre qualidade de skills.

Instalação

$npx convez-ai jedobe/skill-evaluator

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
0
Forks
0
Acompanhando
0
Saúde
57%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Validar qualidade de skills antes de publicar ou compartilhar
  • Identificar gaps em trigger clarity, instruction specificity e reference density
  • Benchmarking comparativo entre skills próprias e padrão da comunidade
  • Orientar melhorias incrementais com feedback em 7 dimensões específicas

Veredito

Skill prático que resolve um problema real: como saber se uma skill é boa de verdade. Rubrica bem calibrada e transparente, baseada em análise de projetos reais. Ideal para quem quer elevar o padrão de skills que cria ou usa.

README

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

skill-evaluator

License: MIT Version CI Platform

Everyone's writing skills. Nobody knows if they're any good.

This skill scores yours — against a rubric reverse-engineered from the 9 most-starred Claude Code skill repos on GitHub.


How it works

Most skill feedback is vibes. This isn't.

The rubric was built by analyzing what actually separates top-tier skills (obra/superpowers 229k★, affaan-m/ECC 216k★, anthropics/skills 151k★) from the rest. Seven dimensions. 100 points. A tier you can point to.

## Skill Evaluation: karpathy-guidelines

Scope: Single skill
Category: Guideline skill

| Dimension            | Score | Max |
|----------------------|-------|-----|
| Trigger Clarity      |  16   |  20 |
| Instruction Specificity |  20 | 24 |
| Reference Density    |   5   |   8 |
| Verifiability        |   3   |   5 |
| Tradeoff Transparency|  16   |  18 |
| Portability          |  14   |  18 |
| Maintenance Maturity |   4   |   7 |
| **Total**            | **78**|**100**|

### Tier: Gold
Strong behavioral guidelines, but thin on bundled reference data and evals.

### Top 3 Improvements
1. **Reference Density**: Add a `references/` folder with lookup tables or code examples...
2. **Verifiability**: Define an output spec or add test prompts to an `evals/` folder...
3. **Trigger Clarity**: Add "do NOT use when..." conditions to the description...

Install

/plugin install skill-evaluator

Or clone and point at it manually:

git clone https://github.com/jedobe/skill-evaluator

Usage

Ask Claude to evaluate any skill — by file path, GitHub URL, or pasted content:

evaluate ~/.claude/skills/my-skill/SKILL.md
evaluate this skill: [paste SKILL.md here]

The 7 dimensions

Weights adapt to the skill category. Tool skills produce structured output or automate a task; Guideline skills shape how the model behaves. Both total 100.

#DimensionToolGuidelineThe question it answers
1Trigger Clarity2020Does the description tell the model when to invoke — not just what it does?
2Instruction Specificity1524Is there a concrete procedure, or just a description of desired output?
3Reference Density158Is supporting data bundled in — or does the model rely on training alone?
4Verifiability155Is there a defined output spec, eval suite, or success criteria?
5Tradeoff Transparency1018Does the skill honestly state its limits and when NOT to use it?
6Portability1518Zero-dep? Multi-harness? No hardcoded paths?
7Maintenance Maturity107License, version, CHANGELOG — does it look maintained?

Tiers: Elite (85+) · Gold (70–84) · Silver (50–69) · Bronze (0–49)


Calibration

Scores are grounded in real repos. A few reference points:

SkillStarsScoreTier
anthropics/skills — skill-creator—~88Elite
JuliusBrussee/caveman73k★~85Elite
multica-ai/andrej-karpathy-skills176k★~78Gold
OthmanAdi/planning-with-files23k★~76Gold

If your skill scores 85+, it's in genuinely rare company.

Note: skill-evaluator itself is not in this table. Meta-skills (tools that evaluate other tools) don't fit the rubric — the dimensions were designed for task-performing skills. Scoring a rubric tool against its own rubric is circular.


Why this exists

The skill ecosystem is growing fast. There's no shared standard for what "good" looks like — so most feedback is either "looks fine" or a wall of subjective opinions.

This rubric is an attempt to make that judgment concrete, reproducible, and grounded in what the community has already validated with stars.


License

MIT

Ver o original no GitHub

Bom para

Relacionados