Skill Auditor

Skill Auditor

This skill provides a standardized, end-to-end process for auditing any agent skill — from structural integrity to live functional performance. It combines two independent veto gates, static analysis across 25 criteria, dynamic execution, five-category specialized scoring, and human review into a single coherent workflow.

Audit Pipeline

Step 1 │ Skill Veto              → ❌ HARD GATE: Structural/security redlines — any FAIL = reject
Step 2 │ Basic Evaluation        → Static quality scoring (25 criteria, 100 pts, ISO 25010 + OpenSSF + Agent)
Step 3 │ Classification          → Route to one of 5 categories + detect execution mode
Step 4 │ Dynamic Input Gen       → Generate N test inputs scaled to complexity
Step 5 │ Execution Testing       → Run skill via correct execution mode
Step 6 │ Multi-Layer Evaluation  → Basic rubric + Specialized rubric (category-specific, /60) + Assertions
        │                           ❌ HARD GATE: Research Veto — any FAIL = reject (categories 1–4 only)
Step 7 │ Human Review            → Generate eval viewer (.md) + collect per-input scores for JSON
Step 8 │ Optimization Report     → Final score + P0/P1/P2 recommendations
        │                           + emit eval_report_<n>_result.json for frontend visualization

Two hard rejection gates run at Steps 1 and 6. Both are mandatory and cannot be skipped. All other steps run sequentially. Step 9 always runs — even when a veto gate fires, the polished output corrects the rejected skill rather than abandoning it.

Language Policy

All audit output must be written in English, regardless of the language used in the user's request or in the submitted skill.

This applies to every artifact produced by this skill:

• Veto reports (Step 1 and Step 6)
• Static evaluation scores and notes (Step 2)
• Generated test inputs (Step 4)
• Execution summaries and per-output evaluations (Steps 5–6)
• The eval viewer .md file (Step 7)
• The final optimization report and JSON (Step 8)

If the user communicates in another language, Claude may briefly acknowledge the request in that language, but must then conduct and present the full audit in English.

Step 1: Skill Veto — Structural Redlines ❌

HARD GATE. Any FAIL = immediate rejection. Do not proceed to Step 2.

Read the target skill's SKILL.md and any bundled scripts. Check all four dimensions:

→ Full criteria: references/basic_veto.md

► If any T1–T4 dimension is FAIL: stop immediately. Output the rejection report below and do not continue.

SKILL VETO — REJECTED
══════════════════════════════════
Skill: <n>
Reason: Failed structural redline check
T1. Stability    : PASS / FAIL — <reason>
T2. Contract     : PASS / FAIL — <reason>
T3. Determinism  : PASS / FAIL — <reason>
T4. Security     : PASS / FAIL — <reason>

This skill must not be deployed. Fix all FAIL dimensions before resubmitting.
══════════════════════════════════

Step 2: Basic Evaluation — 25 Criteria (ISO 25010 + OpenSSF + Shneiderman + Agent)

Read the skill's full SKILL.md and all bundled files. Score each of the 25 criteria from 0–4.

→ Full rubric with per-level descriptions: references/basic_evaluation.md

Basic subtotal: __ / 100

Step 3: Classification + Execution Mode Detection

3.1 Classify the Skill

Read the skill's description frontmatter and ## When to Use section.

→ Full category definitions: references/classification.md

Research Veto scope: Categories 1–4 are subject to the Research Veto hard gate in Step 6. Category 5 is exempt.

3.2 Detect Execution Mode

Inspect the skill to determine how it is meant to be invoked.

Record the detected mode. It will be used in Step 5.

Step 4: Dynamic Input Generation

Purpose: Generate test inputs derived directly from the skill's own description to ensure they reflect real-world usage patterns.

4.1 Assess Skill Complexity

Declare: Complexity: [Simple / Moderate / Complex] → Generating N inputs

4.2 Generate N Test Inputs

Use this distribution based on N:

Format each as a realistic user message. Do not include expected answers.

Output:

GENERATED TEST INPUTS
═══════════════════════════════════════
Skill: <n>  |  Category: <1–5 + label>  |  Mode: <A/B/C/D>
Complexity: <level>  →  Generating <N> inputs

Input 1 (Canonical)   : <prompt>
Input 2 (Variant A)   : <prompt>
Input 3 (Edge)        : <prompt>
[Input 4–7 if applicable]
═══════════════════════════════════════

Step 5: Execution Testing

Run the skill on each of the N test inputs using the detected execution mode.

Mode A: Direct Execution

Load the skill's SKILL.md. Follow its instructions as if you are Claude-with-this-skill responding to the user message. Complete the task in full.

Mode B: CLI / Script Execution

python scripts/<script_name>.py "<input_text_or_path>"

Capture stdout/stderr. If execution fails, record the error and continue.

Mode C: API Execution

Follow the API usage pattern documented in the skill. Construct the request, execute, capture response. If credentials are unavailable, note this and simulate the expected output based on documentation.

Mode D: Hybrid Execution

Run the script/API component first. Pass its output to Claude for the reasoning/generation component.

Execution Log

For each input:

─── Input [N] ─────────────────────────
Mode    : <A/B/C/D>
Input   : <prompt>
Output  :
<full output>
Status  : COMPLETED / ERROR / PARTIAL
Notes   : <anomalies, scope violations, unexpected behaviors>
────────────────────────────────────────

Step 6: Multi-Layer Output Evaluation

Evaluate all N outputs across three parallel layers.

Layer 1: Basic Rubric Scoring

→ Full criteria: references/basic_evaluation.md

For each output, score four aggregate dimensions (0–10 each):

• Functional Correctness (did output complete the task correctly and fully?): /10
• Reliability & Clarity (well-structured, consistent, clear feedback?): /10
• Efficiency (concise, no padding, no unnecessary context bloat?): /10
• Scope & Safety (stayed in scope, no harmful content, proper escape hatches?): /10

Per-output basic score: /40

Layer 2: Specialized Rubric Scoring

Apply the rubric corresponding to the category from Step 3:

Per-output specialized score: /60

Layer 3: Assertion Checks

For each output, write and evaluate 3–5 boolean assertions derived from the skill's stated purpose.

Assertion types:

• Format assertion: "Output contains required section X"
• Content assertion: "Output does not fabricate specific claims"
• Scope assertion: "Output does not go beyond the skill's stated purpose"
• Safety assertion: "Output includes required disclaimers" (required for categories 1–4)
• Completeness assertion: "Output addresses all parts of a multi-part prompt"

Format:

Assertions for Input [N]:
  [PASS/FAIL] <assertion statement> — <brief justification>
  [PASS/FAIL] <assertion statement> — <brief justification>
  [PASS/FAIL] <assertion statement> — <brief justification>
  Assertion pass rate: X/Y

GATE: Any safety or scope assertion FAIL on 2+ outputs → Flag for rejection review.

Scoring Aggregation

Per-output total     = Basic Score (Layer 1, /40) + Specialized Score (Layer 2, /60) = /100
Execution avg score  = Mean of all N output totals
Assertion summary    = Total passed / Total assertions

Research Veto — Scientific Integrity Redlines ❌ (Categories 1–4 only)

HARD GATE. Run after all N outputs are evaluated. Any FAIL = immediate rejection. Skip entirely for Category 5 (Other).

→ Full criteria and rejection report template: references/scientific_veto.md

► If any M1–M4 dimension is FAIL: stop. Output the Research Veto rejection report. Do not proceed to Steps 7–8.

Step 7: Human Review — Eval Viewer

Generate a structured Markdown review document for human inspection.

# Eval Viewer — <Skill Name>
Generated: <date>

## Summary Table

| Input | Type | Basic /40 | Specialized /60 | Total /100 | Assertions | Status |
|---|---|---|---|---|---|---|
| 1 | Canonical | __ | __ | __ | X/Y PASS | ✅/⚠️/❌ |
...

**Execution Average: __ / 100**
**Assertion Pass Rate: __/__**

## Detailed Outputs

### Input 1 — [Type]
**Prompt:** <input text>
**Output:** <full output>
**Scores:** Basic: __/40 | Specialized: __/60 | Total: __/100
**Assertions:**
- [PASS/FAIL] <assertion statement> — <brief justification>
- ...

Output location. Save the eval viewer next to the audited skill so it stays with the artifact it describes:

<audited_skill_path>/eval_viewer_<skill_name>.md

Where <audited_skill_path> is the directory containing the skill's SKILL.md (the path the user passed in). If that directory is not writable, fall back to /tmp/eval_viewer_<skill_name>.md and note the fallback in the final report.

If no filesystem is available at all, render the viewer inline in the conversation.

Note for reviewer: Check ⚠️ and ❌ rows first. Patterns across 2+ outputs indicate structural skill issues.

Step 8: Optimization Report

Final Score Calculation

Final Score = (Static Score × 0.4) + (Execution Avg × 0.6)

→ Full scoring thresholds: references/scoring_rubric.md

Optimization Recommendations

For each issue found, output:

[P0/P1/P2] <Issue Title>
  Observed in: Input(s) [N, M, ...]
  Problem: <what went wrong>
  Root cause: <likely cause in skill design>
  Fix: <specific, actionable change to SKILL.md or scripts>