skill-judge

---
name: skill-judge
description: Use when reviewing, auditing, or improving an agent SKILL.md so the agent scores design quality across multiple dimensions and emits concrete fixes against the official skill spec.
source: https://github.com/softaworks/agent-toolkit/tree/main/skills/skill-judge
generated: 2026-05-25T05:10:34.966Z
category: tool
audience: ai
---

## When to use

- Auditing a freshly-written SKILL.md before publishing it to a marketplace
- Reviewing a third-party skill to decide whether it's worth installing
- Refactoring an in-house skill that has drifted from the official spec over time
- Comparing two skills covering the same surface to pick the better-designed one

## Key concepts

### Skill as knowledge externalization, not tutorial

Skill Judge enforces the distinction: a skill encodes patterns and triggers the agent activates on, not step-by-step instructions for a human. Mis-framed skills score low here.

### Multi-dimensional scoring

Each skill is graded across activation clarity, trigger specificity, content quality, tool/format hygiene, and anti-pattern detection. The output is a per-dimension score plus an aggregate, not a single thumbs-up.

### Reference library of 17+ examples

Patterns are derived from real official skills, not invented. When the judge flags something, it can point to a reference skill that handles the same problem correctly.

### Actionable fix suggestions

Findings come with the specific change to make — exact wording for triggers, missing gotcha categories, mis-shaped tool blocks — instead of generic 'improve clarity' notes.

### Anti-pattern detection

Catches common drift: tutorial-shaped skills, over-broad triggers, missing failure modes, and skills that duplicate work the base model already does well.

## API reference

```
npx skills add softaworks/agent-toolkit --skill skill-judge
```

Install the skill-judge auditing skill.

```
npx skills add softaworks/agent-toolkit --skill skill-judge
```

## Gotchas

- Don't run the judge against itself or other judge-style skills; the rubric was tuned for action skills
- Low scores don't mean delete; they mean 'rewrite against the spec' — the judge points to which dimension to fix first
- A perfect score is rare and not the goal; aim for high marks on activation and triggers, the highest-leverage dimensions
- Skill versioning matters; re-run the judge after any meaningful edit because triggers drift fast
- Some legitimate niche skills score lower because their patterns are unique; treat the judge as a discussion partner, not a gate

---
Generated by SkillMake from https://github.com/softaworks/agent-toolkit/tree/main/skills/skill-judge on 2026-05-25T05:10:34.966Z.
Verify against source before relying on details.

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