skill-auditor

---
name: skill-auditor
description: Use when reviewing a SKILL.md before installation or publication — audit triggering, structure, safety, evidence, freshness, context cost, and testability, then propose bounded fixes.
source: https://github.com/anthropics/skills/tree/main/skills/skill-creator
generated: 2026-07-14T21:19:49.721Z
category: security
audience: engineers
---

## When to use

- Reviewing a new or changed skill before it enters a marketplace or shared repository
- Investigating why a skill under-triggers, over-triggers, wastes context, or produces inconsistent results
- Checking a third-party skill for unsafe commands, hidden side effects, weak provenance, or stale instructions
- Comparing two versions of a skill and deciding whether the revision is safer and more effective
- Producing a repeatable quality report that separates blocking defects from optional polish

## Key concepts

### Trigger contract

The frontmatter description is the routing surface. It must state the job and concrete request contexts, distinguish neighboring skills, and avoid promises that the body cannot fulfill.

### Progressive disclosure

Keep the core workflow concise and route detailed schemas, provider variants, and large examples into directly linked references. Every bundled file must have a clear load condition.

### Safety boundary

Identify commands, credentials, external writes, deployments, purchases, messages, destructive operations, and privilege changes. Require explicit confirmation where consequences exceed the user's stated scope.

### Evidence chain

Technical claims should trace to primary documentation, executable checks, or repository evidence. Flag uncited version-specific APIs, copied secrets, dead links, and instructions that cannot be reproduced.

### Bounded remediation

Preserve the skill's intent and protected sections. Propose the smallest edits that resolve a demonstrated defect, and keep opinionated rewrites separate from blocking fixes.

## API reference

```
Audit gates: structure -> trigger -> safety -> evidence -> efficiency -> evals
```

Run gates in order so malformed structure and unsafe behavior block publication before subjective quality scoring begins.

```
BLOCK: invalid frontmatter, destructive default, exposed credential
WARN: stale source, vague trigger, unbounded reference loading
INFO: naming, examples, optional compression
```

```
Finding format
```

Make every finding actionable with severity, evidence, impact, and a minimal remediation.

```
[BLOCK] Unsafe publish default
Evidence: SKILL.md:84 deploys without confirmation
Impact: external production mutation
Fix: require explicit target and confirmation before deploy
```

```
Publication verdict
```

Return a clear decision based on unresolved findings rather than an opaque numeric score.

```
PASS: no blockers; warnings documented
CONDITIONAL: fix listed blockers and rerun affected gates
FAIL: unsafe, unverifiable, or materially misleading
```

## Gotchas

- Do not label a skill safe merely because it contains no shell commands; external API calls and messages can be equally consequential
- Do not reward length — duplicated guidance and unconditional reference loading reduce skill quality
- Do not rewrite protected or source-faithful sections to satisfy stylistic preferences
- A GitHub star count is provenance context, not proof that instructions are current, safe, or correct
- Separate source defects from local packaging defects so maintainers know where the fix belongs
- Never auto-fix blocking findings and publish in the same step without re-validation

---
Generated by SkillMake from https://github.com/anthropics/skills/tree/main/skills/skill-creator on 2026-07-14T21:19:49.721Z.
Verify against source before relying on details.

File: ~/.claude/skills/skill-auditor/SKILL.md