Editing: skill-verifier

Field: Description

# skill-verifier

Validate a Claude Code skill two ways, in one tool:

1. **Lint** — deterministic static checks on `SKILL.md`: frontmatter compliance
   (name/description limits, kebab-case, reserved words, unsupported fields),
   description quality (third-person, when-to-use, front-loaded triggers),
   structure / progressive disclosure (body ≤ 500 lines, reference links resolve
   and stay one level deep, TOC on long files), and content anti-patterns
   (backslash paths, time-sensitive phrasing, placeholders).
2. **Trigger e2e** — the part nothing else does. It spawns the **real headless
   Claude Code binary** in an isolated config that contains only the skills you
   choose, feeds it a corpus of prompts, and observes whether the candidate
   skill's `Skill` tool actually fires.

## What it measures

Because skill selection routes through the model, each prompt is run N times and
reported as a **fire-rate** — **stats, not pass/fail**:

- **Success** — how often it fires on prompts it *should* (candidate loaded alone).
- **Success under collision** — does it still win when its neighbor skills are
  loaded, and if not, *which* skill takes the prompt.
- **Cleanliness** — how often it correctly stays quiet on prompts it should not fire.

The engineer reads the numbers and decides whether the description is sharp
enough. Opt into a CI gate with `--recall-min` / `--spec-min`.

## Use

```bash
node ~/.claude/skills/skill-verifier/scripts/skill-verifier.mjs lint <skill>
node ~/.claude/skills/skill-verifier/scripts/skill-verifier.mjs gen  <skill>   # scaffold a trigger corpus
node ~/.claude/skills/skill-verifier/scripts/skill-verifier.mjs e2e  <skill> --model haiku
```

It pairs with **skill-creator** as the verification step before publishing.
Dependency-free (Node 18+); reuses your existing Claude Code credentials for the
isolated runs.

Cancel

You must be logged into Adom to save. Edits are applied immediately but subject to review.