Skill Evaluation & Improvement

Measure and improve skill quality through empirical testing — because structure doesn't guarantee behavior, and measurement beats assumption. Also covers head-to-head bake-offs of two peer implementations of the same artifact (Mode F).

Reference Loading Table

Instructions

Phase 1: ASSESS — Determine what to evaluate

Step 1: Identify the skill

# Validate skill structure first
python3 -m scripts.skill_eval.quick_validate <path/to/skill>

This checks: SKILL.md exists, valid frontmatter, required fields (name, description), kebab-case naming, description under 1024 chars, no angle brackets.

Step 2: Choose evaluation mode based on user intent

GATE: Skill path confirmed, mode selected.

Phase 2: EVALUATE — Run the appropriate evaluation

Mode A: Trigger Evaluation

Test whether a skill's description causes Claude to invoke it for the right queries.

Step 1: Create eval set (or use existing)

Create a JSON file with 8-20 test queries. Eval set quality matters — use realistic prompts with detail (file paths, context, casual phrasing), not abstract one-liners. Focus on edge cases where the skill competes with adjacent skills.

Example of good eval queries:

[
  {"query": "ok so my boss sent me this xlsx file (Q4 sales final FINAL v2.xlsx) and she wants profit margin as a percentage", "should_trigger": true},
  {"query": "Format this data", "should_trigger": false}
]

Why: Real users write detailed, specific prompts. Abstract queries don't test real triggering behavior. Overfitting descriptions to abstract test cases bloats the description and fails on real usage.

Step 2: Run evaluation

python3 -m scripts.skill_eval.run_eval \
  --eval-set evals.json \
  --skill-path <path/to/skill> \
  --runs-per-query 3 \
  --verbose

This spawns claude -p for each query, checking whether it invokes the skill. Runs each query 3 times for reliability. Output includes pass/fail per query with trigger rates. Default 30s timeout; increase with --timeout 60 if needed for complex queries.

Constraints applied:

• Always run baseline eval before making improvements
• 3 runs per query ensures statistical reliability
• Verbose output shows per-query pass/fail during eval runs

GATE: Eval results available. Proceed to improvement if failures found.

Mode B: Description Optimization

Automated loop that tests, improves, and re-tests descriptions using Claude with extended thinking.

python3 -m scripts.skill_eval.run_loop \
  --eval-set evals.json \
  --skill-path <path/to/skill> \
  --max-iterations 5 \
  --verbose

This will:

1. Split eval set 60/40 train/test (stratified by should_trigger) — prevents overfitting to test cases
2. Evaluate current description on all queries (3 runs each for reliability)
3. Use claude -p to propose improvements based on training failures
4. Re-evaluate the new description
5. Repeat until all pass or max iterations reached
6. Select best description by test score (not train score — prevents overfitting)
7. Open an HTML report in the browser

Why 60/40 split: Improvements should help across many prompts, not just test cases. Training on failures, validating on holdout ensures generalization.

Why report HTML: Visual reports enable quick review of which queries improved, which regressed, and what the new description looks like.

GATE: Loop complete. Best description identified.

Mode C: Output Benchmark

Compare skill quality by running prompts with and without the skill.

Step 1: Create test prompts — 2-3 realistic user prompts

Step 2: Run with-skill and without-skill in parallel subagents:

For each test prompt, spawn two agents:

• With skill: Load the skill, run the prompt, save outputs
• Without skill (baseline): Same prompt, no skill, save outputs

Why baseline matters: Can't prove the skill adds value without a baseline. Maybe Claude handles it fine without the skill. The delta is what matters.

Step 3: Grade outputs

Spawn a grader subagent using agents/grader.md. It evaluates assertions against the outputs.

Step 4: Aggregate

python3 -m scripts.skill_eval.aggregate_benchmark <workspace>/iteration-1 --skill-name <name>

Produces benchmark.json and benchmark.md with pass rates, timing, and token usage.

Step 5: Analyze (optional)

For blind comparison, use agents/comparator.md to judge outputs without knowing which skill produced them. Then use agents/analyzer.md to understand why the winner won.

GATE: Benchmark results available.

Mode D: Quick Validate

python3 -m scripts.skill_eval.quick_validate <path/to/skill>

Checks: SKILL.md exists, valid frontmatter, required fields (name, description), kebab-case naming, description under 1024 chars, no angle brackets.

Mode E: Self-Improvement Loop

Automatically generate variants of a skill, A/B test them against the original, and promote winners. This is a closed-loop pipeline — baseline, hypothesize, generate, test, promote.

Read the full protocol: ${CLAUDE_SKILL_DIR}/references/self-improve-loop.md

The loop runs 5 phases: BASELINE (establish metrics with 3+ test cases), HYPOTHESIZE (2-3 single-variable changes), GENERATE VARIANTS (minimal diffs), BLIND A/B TEST (paired comparison via agents/comparator.md), PROMOTE OR KEEP (60%+ win rate required, no regressions). All outcomes — wins and losses — are recorded to the learning DB to prevent re-testing failed hypotheses.

GATE: Self-improvement protocol loaded from reference. Proceed through the 5 phases.

Mode F: Head-to-Head Bake-Off

Score two peer implementations of the same artifact (e.g., toolkit voice-feynman vs an external Feynman voice profile) on a numeric rubric and declare a decisive winner. Use when the user says "bake-off", "head-to-head", "compare implementations", "grade these two", or "which X is better".

Read the full protocol: ${CLAUDE_SKILL_DIR}/references/bake-off-methodology.md

The protocol runs 5 phases: PREPARE (read both artifacts in full, pick a verifier that built neither side), RUBRIC (define 5–12 criteria scored 0–10, pre-state the loser-of-each-criterion before reading evidence), GRADE (every score cites a path/line range or quote; build the matrix; apply anti-rationalization gate), FOLD (filter loser-wins through docs/PHILOSOPHY.md before recommending any folds into the winner), REPORT (output to tmp/<topic>-bakeoff-report.md, gitignored).

The Feynman bake-off (toolkit 86 vs external 74 across 11 criteria, 12-point margin) is the canonical worked example carried in the reference.

GATE: Bake-off protocol loaded from reference. Proceed through the 5 phases.

Phase 3: IMPROVE — Apply results

Step 1: Review results

For trigger eval / description optimization:

• Show the best description vs original
• Show per-query results (which queries improved, which regressed)
• Show train vs test scores

For output benchmark:

• Show pass rate delta (with-skill vs without-skill)
• Show timing and token cost delta
• Highlight assertions that only pass with the skill (value-add)

Step 2: Apply changes (with user confirmation)

If description optimization found a better description:

1. Show before/after with scores
2. Ask user to confirm
3. Update the skill's SKILL.md frontmatter
4. Re-run quick_validate to confirm the update is valid

Constraint: Always show results before/after with metrics. This enables informed decisions.

GATE: Changes applied and validated, or user chose to keep original.

Error Handling

Error: "No SKILL.md found"

Cause: Skill path doesn't point to a valid skill directory Solution: Verify path contains a SKILL.md file. Skills must follow the skill-name/SKILL.md structure.

Error: "claude: command not found"

Cause: Claude CLI not available for trigger evaluation Solution: Install Claude Code CLI. Trigger eval requires claude -p to test skill invocation.

Error: "legacy SDK dependency"

Cause: Outdated instructions or an old checkout still expects a direct SDK client Solution: Update to the current scripts. Description optimization now runs through claude -p.

Error: "CLAUDECODE environment variable"

Cause: Running eval from inside a Claude Code session blocks nested instances Solution: The scripts automatically strip the CLAUDECODE env var. If issues persist, run from a separate terminal.

Error: "All queries timeout"

Cause: Default 30s timeout too short for complex queries Solution: Increase with --timeout 60. Simple trigger queries should complete in <15s.

References

Scripts (in scripts/skill_eval/)

• run_eval.py — Trigger evaluation: tests description against query set
• run_loop.py — Eval+improve loop: automated description optimization
• improve_description.py — Single-shot description improvement via Claude API
• generate_report.py — HTML report from loop output
• aggregate_benchmark.py — Benchmark aggregation from grading results
• quick_validate.py — Structural validation of SKILL.md

Bundled Agents (in skills/meta/skill-eval/agents/)

• grader.md — Evaluates assertions against execution outputs
• comparator.md — Blind A/B comparison of two outputs
• analyzer.md — Post-hoc analysis of why one version beat another

Reference Files

• ${CLAUDE_SKILL_DIR}/references/schemas.md — JSON schemas for evals.json, grading.json, benchmark.json
• ${CLAUDE_SKILL_DIR}/references/self-improve-loop.md — Self-improvement loop protocol: variant generation, blind A/B testing, promotion criteria
• ${CLAUDE_SKILL_DIR}/references/bake-off-methodology.md — Head-to-head bake-off protocol: rubric construction, anti-rationalization gate, philosophy-filtered fold-list, Feynman worked example