evaluating-skills-with-models（用模型评估技能）

Evaluating Skills with Models

Evaluate skills across multiple Claude models using sub-agents with quality-based scoring.

Requirement: Claude Code CLI only. Not available in Claude.ai.

Why Quality-Based Scoring

Binary pass/fail ("did it do X?") fails to differentiate models - all models can "do the steps." The difference is how well they do them. This skill uses weighted scoring to reveal capability differences.

Workflow

Step 1: Load Test Scenarios

Check for tests/scenarios.md in the target skill directory.

Default to difficult scenarios: When multiple scenarios exist, prioritize Hard or Medium difficulty scenarios for evaluation. Easy scenarios often don't show meaningful differences between models and aren't realistic for production use.

Required scenario format:

## Scenario: [Name]

**Difficulty:** Easy | Medium | Hard | Edge-case

**Query:** User request that triggers this skill

**Expected behaviors:**

1. [Action description]
   - **Minimum:** What counts as "did it"
   - **Quality criteria:** What "did it well" looks like
   - **Haiku pitfall:** Common failure mode
   - **Weight:** 1-5

**Output validation:** (optional)
- Pattern: `regex`
- Line count: `< N`

If scenarios.md missing or uses old format: Ask user to update following references/evaluation-structure.md.

Step 2: Execute with Sub-Agents (Phase 1)

Spawn Task sub-agents for each model in parallel.

Prompt template:

Execute the skill at {skill_path} with this query:
{evaluation_query}

IMPORTANT:
- Actually execute the skill, don't just describe what you would do.
- Create output directory under Claude Code's working directory ($PWD):
  $PWD/.ai_text/{yyyyMMdd}/tmp/{skill_name}-{model}-{hhmmss}/
  (Example: If $PWD=/path/to/project, create /path/to/project/.ai_text/20250101/tmp/formatting-tables-haiku-143052/)
- Create all output files under that directory.
- If the skill asks questions, record the exact questions, then assume reasonable answers and proceed.

Return ONLY (keep it brief to minimize tokens):
- Questions skill asked: [list exact questions the skill asked you, or "none"]
- Assumed answers: [your assumed answers to those questions, or "n/a"]
- Key decisions: [1-2 sentences on freedom level, structure choices]
- Files created: [paths only, no content]
- Errors: [any errors, or "none"]

Do NOT include file contents or detailed explanations.

Use Task tool with model parameter: haiku, sonnet, opus

After sub-agents complete: Read created files directly using Glob + Read to evaluate file quality (naming, structure, content). The minimal report provides process info (questions, decisions) that can't be inferred from files.

Step 3: Score Each Behavior

For each expected behavior, score 0-100:

Scoring checklist per behavior:

1. Did it meet the Minimum? (No → score ≤ 25)
2. How many Quality criteria met? (Calculate proportion)
3. Did it hit the Haiku pitfall? (Deduct points)
4. Apply Weight to final calculation

Step 4: Calculate Weighted Scores

Behavior Score = base_score  // after applying deductions (e.g., Haiku pitfalls)
Total = Σ(behavior_score × weight) / Σ(weights)

Rating thresholds:

Step 5: Add Results to README

After evaluation, add a table to the skill's README documenting the results:

README section format:

## Evaluation Results

| Date | Scenario | Difficulty | Model | Score | Rating |
|------|----------|------------|-------|-------|--------|
| 2025-01-15 | Standard workflow | Hard | claude-haiku-4-5-20250101 | 42 | ⚠️ Marginal |
| 2025-01-15 | Standard workflow | Hard | claude-sonnet-4-5-20250929 | 85 | ✅ Good |
| 2025-01-15 | Standard workflow | Hard | claude-opus-4-5-20251101 | 100 | ✅ Excellent |

Table requirements:

• Include full model IDs (e.g., claude-sonnet-4-5-20250929) not just short names
• Show evaluation date in YYYY-MM-DD format
• Indicate scenario difficulty level (Easy/Medium/Hard/Edge-case)
• Include both numeric score and rating emoji
• Append new evaluations (don't overwrite previous results)

This creates a historical record of how the skill performs across models and improvements over time.

Step 6: Output Summary

## Model Evaluation Results

**Skill:** {skill_path}
**Scenario:** {scenario_name} ({difficulty})
**Date:** {YYYY-MM-DD}

### Scores by Behavior

| Behavior | Weight | claude-haiku-4-5-20250101 | claude-sonnet-4-5-20250929 | claude-opus-4-5-20251101 |
|----------|--------|---------------------------|----------------------------|--------------------------|
| Asks clarifying questions | 4 | 25 | 75 | 100 |
| Determines freedom level | 3 | 50 | 75 | 100 |
| Creates proper SKILL.md | 5 | 50 | 100 | 100 |

### Total Scores

| Model | Score | Rating |
|-------|-------|--------|
| claude-haiku-4-5-20250101 | 42 | ⚠️ Marginal |
| claude-sonnet-4-5-20250929 | 85 | ✅ Good |
| claude-opus-4-5-20251101 | 100 | ✅ Excellent |

### Observations
- Haiku: Skipped justification for freedom level (pitfall)
- Haiku: Asked only 1 generic question vs 3 specific
- Sonnet: Met all quality criteria except verbose output

### Next Steps
- Add these results to the skill's README (see Step 5)
- Consider model selection based on your quality requirements and budget

Common Pitfalls by Model

Note: Token usage includes both skill execution AND reporting overhead. Sonnet tends to produce detailed reports, which inflates token count. Compare tool uses for execution efficiency.

Quick Reference

Load scenarios (prioritize Hard) → Execute (parallel) → Score behaviors → Calculate totals → Add to README → Output summary

For detailed scoring guidelines, see references/evaluation-structure.md.