# Audit Skill

> Evaluate any skill against a quality rubric. Checks trigger crispness, scope discipline, conflict detection, and description quality. Use when reviewing skills, evaluating external skills before adoption, or improving existing skills.

- Skill: `dylanpulver/audit-skill` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add dylanpulver/audit-skill`
- Raw SKILL.md: https://api.skillmd.com/api/skills/dylanpulver/audit-skill/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Security
- Author: dylanpulver (https://skillmd.com/u/dylanpulver)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/dylanpulver/audit-skill

---


# Skill Auditor

Evaluates any Claude Code skill against a quality rubric, checks for conflicts with other installed skills, and provides actionable recommendations.

> [Skill Governance](../skill-governance/SKILL.md) > Audit Skill

## Invoke This Skill When

- User says `/audit-skill <name>` or asks to "audit", "review", "evaluate", or "score" a skill
- User is considering adopting an external or open-source skill
- User just finished writing or modifying a skill and wants quality feedback
- User notices unexpected skill triggering or conflicts between skills

## Input

The argument is either:
- A **skill name** (e.g., `acme-app-ops`) — looked up in the skills index
- A **file path** to a SKILL.md (e.g., `~/Downloads/some-skill/SKILL.md`) — read directly

## Workflow

### Step 1: Load the Target Skill

- If given a name: read `skills-index.md` at the root of your skills repo (generated by `/index-skills full`) to find the skill's file path, then read that SKILL.md
- If given a path: read the SKILL.md directly
- Also read any files in a `references/` subdirectory if one exists alongside the SKILL.md
- If the skill cannot be found, tell the user and suggest running `/index-skills full` first

### Step 2: Load the Skills Index

- Read `skills-index.md` at the root of your skills repo
- Extract all skill names, descriptions, and trigger conditions for conflict checking
- If the index doesn't exist or is empty, warn the user and skip conflict checking

### Step 3: Score Against the Quality Rubric

Read the rubric at `references/quality-rubric.md` (relative to THIS skill's directory).

Evaluate the target skill against each dimension independently. Be honest and specific — vague praise is useless. Cite specific lines from the skill that support your score.

### Step 4: Check for Conflicts

Compare the target skill's description and trigger conditions against every other skill in the index:

- **Keyword overlap:** Do two descriptions share the same action verbs AND domain nouns?
- **Subset triggers:** Is one skill's trigger condition a subset of another's?
- **Ambiguous routing:** Could a reasonable user query trigger both skills?

Rate each potential conflict:
- **NONE** — No meaningful overlap
- **LOW** — Minor overlap, but context/domain makes disambiguation trivial
- **MEDIUM** — Overlap exists and could cause mis-routing on ambiguous queries
- **HIGH** — Skills will frequently compete for the same query

### Step 5: Assess Provenance (for external skills only)

If the skill was provided as a file path (not from the index), also evaluate:
- Does the skill reference specific tools or APIs that exist?
- Is the instruction style precise (step-by-step) or vague (aspirational)?
- Are there signs it was tested in production vs. written speculatively?
- Does it make assumptions about the environment that may not hold?

### Step 6: Output the Audit Report

Format the report as follows:

```markdown
# Skill Audit: [skill-name]

**Source:** [file path]
**Version:** [from frontmatter, or "unversioned"]
**Type:** [router/leaf] | **Hidden:** [yes/no] | **Slash command:** [/cmd or none]

## Scores

| Dimension | Score | Summary |
|---|---|---|
| Trigger Crispness | A/B/C/D | [one-line explanation] |
| Scope Discipline | A/B/C/D | [one-line explanation] |
| Description Quality | A/B/C/D | [one-line explanation] |
| Instruction Clarity | A/B/C/D | [one-line explanation] |
| Tool Permissions | A/B/C/D | [one-line explanation] |
| Conflict Potential | NONE/LOW/MED/HIGH | [one-line explanation] |

**Overall: [A/B/C/D]**

## Detailed Findings

### Trigger Crispness
[2-4 sentences. What triggers it, what shouldn't trigger it, any false positive risk.]

### Scope Discipline
[2-4 sentences. How many distinct tasks does it handle? Could any be split out?]

### Description Quality
[2-4 sentences. Is it concise? Action-oriented? Specific enough for the model to route correctly?]

### Instruction Clarity
[2-4 sentences. Could an LLM follow the instructions unambiguously? Missing steps?]

### Tool Permissions
[2-4 sentences. Are allowed-tools specified? Minimal or overly broad?]

## Conflicts Detected

[Table of conflicts, or "No conflicts detected with N indexed skills."]

| Conflicting Skill | Overlap | Severity |
|---|---|---|
| skill-name | "Both trigger on X" | LOW/MED/HIGH |

## Recommendations

[Numbered, actionable items. Focus on the lowest-scoring dimensions. Each recommendation should be specific enough to implement directly.]

1. [Specific recommendation]
2. [Specific recommendation]
...
```

## Scoring Guidelines

When in doubt, refer to the full rubric at `references/quality-rubric.md`. Key principles:

- **A** = production-quality, no changes needed
- **B** = good with minor improvements possible
- **C** = functional but needs work before scaling
- **D** = needs significant revision

Be calibrated: most well-intentioned skills should land at B or C. Reserve A for genuinely tight, battle-tested skills. Reserve D for skills that are broken or harmful.