Skill Eval#

A diagnostic instrument for AI skills. Feed it any skill, get back a structured report — what's working, what's broken (Bad Cases), and what to fix — then iterate until it passes.

---

Output Structure (Mandatory)#

Create {target-skill-name}-eval/ at the same level as the target skill:

text

{target-skill-name}-eval/
├── v1/
│   ├── plan.md
│   ├── trigger-results.json
│   ├── cases.json
│   ├── execution-results.json
│   ├── report.md
│   └── optimized-skill/SKILL.md
├── v2/ ...
└── summary.md

Version rule: v1 → v2 → v3. Always create a new v{N}/ for each run.

---

Pipeline & Checkpoints#

PhaseActionOutput FileDetail0-1Analyze skill + design planv{N}/plan.md``agents/planner.md1.5Trigger evaluationv{N}/trigger-results.json``scripts/run_trigger_eval.py2Design test casesv{N}/cases.json``references/test-case-design.md3Execute & recordv{N}/execution-results.json``agents/executor.md4Score & verify_(scores for Phase 5)_agents/judge.md + references/scoring.md5Report + optimizev{N}/report.md + optimized-skill/SKILL.md + summary.md``agents/reporter.md + agents/advisor.md

RULE: Each phase writes its file BEFORE the next begins. No file = not done. Go back.

---

Phase 0-1: Analyze & Plan#

Input: Target SKILL.md Output: v{N}/plan.md

Assess structure (high/medium/low), dissect steps with operation types, design test strategy, identify risks and sandbox requirements.

Full protocol → agents/planner.md

CHECKPOINT: Write v{N}/plan.md. Verify it exists. Do NOT proceed until written.

---

Phase 1.5: Trigger Evaluation#

Input: v{N}/plan.md + target SKILL.md (frontmatter description field) Output: v{N}/trigger-results.json

Design trigger probes (positive + negative queries) based on the skill's description, then test whether each probe correctly triggers or does not trigger the skill.

1. Design probes from the plan's Trigger Probe Strategy:

   • Positive probes (should trigger): direct requests, synonyms, multilingual variants, implicit intent
   • Negative probes (should NOT trigger): similar-sounding but out-of-scope queries, general info requests
   • Minimum: 5 positive + 5 negative probes (quick mode), 8+8 (deep mode)

2. Run probes via scripts/run_trigger_eval.py or manual testing against the platform's skill activation mechanism.

3. Record results as v{N}/trigger-results.json with precision, recall, and per-probe pass/fail.

Full script → scripts/run_trigger_eval.py

CHECKPOINT: Write v{N}/trigger-results.json. Must contain precision + recall metrics. Do NOT proceed until written.

---

Phase 2: Design Test Cases#

Input: v{N}/plan.md + target SKILL.md Output: v{N}/cases.json

Generate per-step expected results with check_types (exact > regex > semantic). Quick mode: 4 cases. Deep mode: 8-12 cases.

Critical: Expected results written BEFORE execution. Never adjust after.

Full protocol → references/test-case-design.md JSON schema → references/schemas.md

CHECKPOINT: Write v{N}/cases.json. At least 4 cases (quick) or 8 (deep). Do NOT proceed until written.

---

Phase 3: Execute & Record#

Input: v{N}/cases.json + target SKILL.md Output: v{N}/execution-results.json

Run each case with the skill active. Record per-step: action_taken, actual_output, tool_calls, tokens, time. Run at least 1 baseline (without skill).

Full protocol → agents/executor.md JSON schema → references/schemas.md

CHECKPOINT: Write v{N}/execution-results.json. Every case must have entries. Do NOT proceed until written.

---

Phase 4: Score & Verify#

Input: v{N}/execution-results.json + v{N}/cases.json Output: Scored data for Phase 5

Three independent scores per step:

• Completion (0/1): Did the operation execute?
• Correctness (0/1/2): Does actual match expected?
• Execution Quality (0/1/2): Did it follow the skill's method?

Rules: Completion=0 cascades. Scores never combined. Every sub-max score needs a reason.

Full protocol → agents/judge.md Scoring definitions → references/scoring.md Rubric templates → references/rubrics.md

CHECKPOINT: All steps scored. Every step has three scores + reasons. Ready for Phase 5.

---

Phase 5: Report & Optimize#

Input: Scored results + Advisor analysis Outputs: v{N}/report.md + v{N}/optimized-skill/SKILL.md + summary.md

Report order: Bad Cases FIRST → Overview → Step Scores → Baseline → Efficiency → Safety → Details.

If Bad Cases exist → produce optimized skill with fixes applied. If Bad Cases = 0 + stop conditions met → PASSED.

Full protocol → agents/reporter.md Root cause analysis → agents/advisor.md Visual formats → references/report-format.md

CHECKPOINT:

• v{N}/report.md exists with Bad Cases + overview + step scores
• Bad Cases > 0 → v{N}/optimized-skill/SKILL.md exists
• summary.md updated

---

Bad Cases#

A step is a Bad Case if ANY: Completion=0, Correctness=0, Quality=0, or safety finding.

---

Stop Conditions (All Must Be True to PASS)#

ConditionThresholdBad Cases = 0No steps with any score = 0Correctness avg>= 1.8/2No regressionsNo previously-passing case now failsUnsafe rate = 0%No safety findings

---

Non-Negotiables#

1. Expected results before execution.
2. Low scores need reasons (expected vs actual).
3. Bad Cases shown first.
4. Three scores stay independent. Never combined.
5. Versions immutable. New v{N} for each run.
6. Every fix traces to a Bad Case.
7. Regressions are zero-tolerance.
8. Structure check before testing.
9. Baseline proves skill's value.
10. Every phase writes its file. No file = not done.

---

File Index#

FilePurposeagents/planner.mdPhase 0-1: Structure assessment + plan generationagents/executor.mdPhase 3: Test execution + recordingagents/judge.mdPhase 4: Scoring engine protocolagents/advisor.mdPhase 5B: Root cause analysis + optimizationagents/reporter.mdPhase 5: Report generation + summaryreferences/test-case-design.mdPhase 2: Case design guide + cases.json schemareferences/schemas.mdAll JSON data structure definitionsreferences/scoring.mdScoring scales, computation, displayreferences/rubrics.mdPer-operation-type rubric templatesreferences/report-format.mdVisual report presentationscripts/score_engine.pyAutomated scoring computationscripts/safety_scanner.pyStatic safety analysisscripts/generate_scorecard.pyHTML report generationscripts/run_trigger_eval.pyPhase 1.5: Multi-platform trigger evaluation

---

Platform Compatibility#

PlatformSkill LocationTriggerClaude Code.claude/commands/``claude -p CLICursor.cursor/rules/Agent modeCodex.codex/skills/CLI/APIOpenClaw.claw/skills/Hub activation

---

Security#

• Sandbox untrusted skills — disposable workspace, mock data, approval mode
• Skill names sanitized — no path traversal in trigger probes
• HTML reports escaped — html.escape() on all interpolated values
• Descriptions neutralized — prompt-injection patterns stripped from probe files

See agents/executor.md Safety Boundary section for full sandboxing protocol.