skill-insp: A Skill That Scores Other Skills

If you've been building Claude Code skills for a while, you've probably noticed a pattern: every skill author makes the same mistakes the first few times. Vague descriptions that fail to trigger. Workflows that don't say what to do when files are missing. allowed-tools that ask for Bash with no glob restriction. No eval scenarios, so you have no idea if the skill actually works.

I built skill-insp to catch those mistakes automatically. It's a skill that inspects other skills, scores them across 8 dimensions, and tells you what to fix.

Source: github.com/conanttu/skills

What It Does

You point skill-insp at a folder containing a SKILL.md and it gives you:

1. A 0–100 score across 8 weighted dimensions (Structure, Triggering, Usability, Completeness, Progressive Disclosure, Testability, Maintainability, Safety & Trust)
2. A prioritized list of findings (High / Medium / Low) with file:line references
3. An HTML report you can open in a browser
4. The ability to apply high-priority fixes with automatic backup and revert with SHA-256 hash verification
5. A functional eval runner that spawns sub-agents against fixture skills to verify the skill's own logic

✨ skill-insp ✨

Overall: my-skill scores 66/100 — risk low, readiness usable-with-improvements.

Key strengths
- Minimal, appropriate permissions (Read and Write only)
- Clear 3-step workflow
- Clean YAML frontmatter with version tracking

Scorecard
| Dimension              | Score |
|------------------------|-------|
| Structure              |  8/10 |
| Triggering             |  9/15 |
| Usability              |  9/15 |
| Completeness           |  7/15 |
| Progressive Disclosure |  7/10 |
| Testability            |  4/10 |
| Maintainability        |  8/10 |
| Safety & Trust         | 14/15 |
| Total                  | 66/100 |

Recommendations
  Medium  Add error-handling for missing or unreadable files.
  Medium  Create evals/ with at least one input and expected output.
  Low     Expand README with usage examples or remove the placeholder.

HTML report: /abs/path/to/cache/my-skill/latest.html
✨ skill-insp ✨

The Scoring Rubric

The 100 points are weighted toward what actually breaks skills in practice:

Safety & Trust is the dimension where pattern-matching tools fall apart, so skill-insp does semantic analysis here. It distinguishes between documentation and executable code: a SKILL.md that says "check for rm -rf usage" in a safety checklist is not a destructive operation. A scripts/cleanup.sh that actually runs rm -rf "$TEMP_DIR" is, and gets flagged with the file:line reference.

How It's Built

The skill follows the "model is the analyzer" pattern. There's no Python or Node script that parses YAML and counts characters. Instead:

skill-insp/
├── SKILL.md                       # Workflow + the 4 modes (default, detailed, apply, revert) + Run Evals
├── README.md
├── references/
│   ├── rubric.md                  # Scoring dimensions and what good looks like
│   └── output-format.md           # JSON schema for analysis.json
├── scripts/
│   ├── render-html.js             # analysis.json → latest.html
│   └── run-evals.js               # fixture setup + sub-agent prompt generation
├── assets/
│   └── report_template.html       # Self-contained HTML template
└── evals/
    └── evals.json                 # 8 functional eval scenarios

Two deterministic scripts handle the parts that should be deterministic:

• render-html.js turns a JSON analysis into a self-contained HTML report. The template uses CSS custom properties for theming, conic-gradient score rings, and an auto-hidden eval results section.
• run-evals.js creates a fixture skill in cache/_fixtures/<id>/, copies a snapshot of skill-insp itself into _skill_home/ so sub-agents can resolve <this-skill> references, and prints a self-contained sub-agent prompt.

Everything else — reading files, parsing YAML, evaluating the rubric, distinguishing documentation from code — is done by the model. This sounds slower than a parser, but it's actually the only way to do it correctly. A regex doesn't know that rm -rf inside a markdown code fence labeled "examples to flag" is not the same as rm -rf inside an executable script.

Progressive Disclosure in Practice

Earlier versions of skill-insp had a 300-line SKILL.md with the entire rubric inline. That hit context budget hard and made the skill harder to edit. The current layout pushes details to references:

• SKILL.md holds the workflow and mode entry points. ~150 lines.
• rubric.md holds the scoring dimensions, priority levels, compactness rules, and safety guidance. Only loaded when the skill actually scores something.
• output-format.md holds the JSON schema. Only loaded when writing analysis.json.

The result: SKILL.md is small enough to read in one sitting, and the model only loads the rubric when it's about to score.

The Eval Runner

This is the part that took the most iteration. The idea is simple: skill-insp ships with 8 eval scenarios in evals/evals.json, each describing a user prompt, fixture files to create, and expectations to verify. To run them:

node scripts/run-evals.js <skill-path> list
node scripts/run-evals.js <skill-path> setup <id>

setup creates a fixture directory, writes the fixture files, copies skill-insp's own resources into _skill_home/, and prints a JSON payload that includes a sub_agent_prompt. The parent agent reads that JSON, spawns a sub-agent with the prompt, and after the sub-agent finishes, checks each expectation:

• File expectations ("analysis.json is written") → find over the fixture directory.
• Content expectations ("a high-priority finding is raised") → read the output files.
• Behavioral expectations ("the model reports an error and stops") → judge from the sub-agent's text output.

The Sandbox Gotcha

In the first version, fixtures were created under os.tmpdir(). This worked when I ran it manually, but sub-agents spawned by the harness were sandboxed to the project root — they got "permission denied" on every Read and Bash call against /var/folders/.../T/eval-skill-insp-*. Three out of eight evals failed for sandbox reasons that had nothing to do with the skill's logic.

The fix was a one-line change: move fixtures into cache/_fixtures/<id>/ inside the project. Now sub-agents inherit the project's filesystem permissions, and the cache directory is .gitignored so it doesn't pollute commits. After the change, all 8 evals run cleanly.

Lesson worth remembering: if you're going to spawn sub-agents, keep their working directory inside the parent's sandbox. Temp directories outside the project tree look like a clean choice but break under tighter permission policies.

Apply and Revert

When you say "apply recommendations", skill-insp:

1. Re-reads the target files (state may have changed since the inspection).
2. Copies each file it's about to modify into <cache_dir>/last-apply/.
3. Computes SHA-256 hashes before and after, recording them in manifest.json.
4. Makes the minimal edits.
5. Re-runs the analysis so you see the new score.

The manifest looks like this:

{
  "applied_at": "2026-05-25T23:16:00Z",
  "recommendations_applied": [
    { "priority": "high", "dimension": "triggering", "text": "Replace vague description..." },
    { "priority": "high", "dimension": "usability", "text": "Add a ## Workflow section..." },
    { "priority": "high", "dimension": "completeness", "text": "Add allowed-tools list..." }
  ],
  "files": [
    {
      "relative_path": "SKILL.md",
      "before_sha256": "e698920f94613f8fc335cd0e941938e0990bedd72cea66e52a6b956d4ff47845",
      "after_sha256":  "10c1fafffbfd2b0089a85e72aafc43432374ef627cb0b41602f8396083fa2800"
    }
  ]
}

Revert is the inverse: read the manifest, verify the current file hash matches the recorded after_sha256 (so we don't blow away edits made after the apply), then restore from backup. If the hash doesn't match, skill-insp reports the conflict instead of overwriting. It never falls back to git reset --hard or git checkout -- — those are blast-radius operations that don't belong in a recovery path.

Self-Inspection

Because skill-insp is itself a skill, it can score itself:

You: 评估 .claude/skills/skill-insp
Claude: ✨ skill-insp ✨
        Overall: skill-insp scores 94/100 — risk low, readiness ready.
        ...

The first time I did this, the report flagged things I'd already half-noticed but not bothered to fix: the cache slug derivation rule was dense without an example, the description didn't mention "fix" or "improve" as triggers, and there was no Node.js version floor documented. All of these became Low/Medium recommendations, which I then applied — and the score went up.

This is the most useful feedback loop I've found for skill authoring: write the skill, run skill-insp against it, apply the high-priority recommendations, repeat. The eval suite then verifies the workflow still works end-to-end.

When NOT to Use It

skill-insp's description explicitly says "Not for general code review." It's not a linter for arbitrary Python or TypeScript. It's specifically tuned to the structure and conventions of Claude Code skills:

• It expects SKILL.md with YAML frontmatter.
• It scores against a skill-specific rubric.
• Its safety analysis is calibrated to skills (permission scoping, undisclosed network access in scripts, prompt injection in instructions).

If you point it at a normal source tree, it'll refuse to score because there's no SKILL.md — that's the intended behavior, not a bug.

Trying It

Clone the repo and drop the folder into your Claude Code skills directory:

git clone https://github.com/conanttu/skills.git
ln -s "$(pwd)/skills/skill-insp" ~/.claude/skills/skill-insp

Then in any Claude Code session:

inspect the skill at ./my-skill

Or in Chinese:

evaluate ./my-skill

The trigger phrases are listed in the description so the skill is invoked automatically. After the inspection, follow the numbered prompts:

1. detailed mode to expand evidence
2. apply recommendations to auto-fix high-priority findings
3. run evals to verify the skill with eval scenarios
4. revert to undo the last apply

What's Next

The current version is 1.0.0. A few things I'd like to add:

• Diff view in the HTML report showing what apply actually changed.
• Cross-skill consistency checks for repos that ship multiple skills.
• Optional schema validation for evals.json and analysis.json.

If you build skills regularly, give it a try and let me know what falls over. The eval scenarios in evals/evals.json are a good place to start if you want to extend it — adding a new scenario is just adding a JSON entry with a prompt, fixture files, and expectations.