validate-skills

por callstackincubator

Valida skills conforme a especificação agentskills.io e as melhores práticas do Claude Code. Verifica a conformidade dos metadados da skill, incluindo formato do nome, tamanho da descrição e campos opcionais como licença e compatibilidade. Aplica as melhores práticas do Claude Code: descrições em terceira pessoa, limites de tamanho do corpo, referências de nível único e formatação de links em markdown. Fornece relatórios de validação estruturados que sinalizam violações de especificação e desvios das diretrizes de criação. Projetado para uso via comando /validate-skills dentro de um...

npx skills add https://github.com/callstackincubator/agent-skills --skill validate-skills

Baixar ZIP GitHub

1.6k

Validate Skills

Validate all skills in skills/ against the agentskills.io spec and Claude Code best practices.

Validation Checklist

For each skill directory, verify:

Spec Compliance (agentskills.io)

Check	Rule
`name` format	1-64 chars, lowercase alphanumeric + hyphens, no leading/trailing/consecutive hyphens
`name` matches directory	Directory name must equal `name` field
`description` length	1-1024 characters, non-empty
Optional fields valid	`license`, `metadata`, `compatibility` if present

Best Practices (Claude Code)

Check	Rule
Description format	Third person, describes what + when to use
Body length	Under 500 lines
Loading is one-level deep	`SKILL.md` is the only progressive-disclosure entry point: every reference file must be reachable from `SKILL.md`. References may cross-link each other for navigation (see note below).
Links are markdown	Use `[text](path)` not bare filenames
No redundancy	Don't repeat description in body
Concise	Only add context Claude doesn't already have

One-level-deep vs. cross-linking. The one-level-deep rule targets progressive-disclosure loading chains — a reference that can only be discovered by loading another reference first (SKILL.md → a.md → b.md, where b.md is not linked from SKILL.md). That is a defect: it hides content from the loader.

It does not forbid navigational cross-links. Per AGENTS.md, reference files end with a "Related Skills" footer linking sibling references, and this is required. A cross-link is fine as long as both endpoints are also reachable directly from SKILL.md. Only flag a reference that is reachable exclusively through another reference.

How to Run

Find all skill directories:
```
fd -t d -d 1 . skills/
```
For each skill, read SKILL.md and check against the rules above

Report issues in this format:

## Validation Results

### skills/example-skill
- [PASS] name format valid
- [FAIL] name "example" doesn't match directory "example-skill"
- [PASS] description length OK (156 chars)