| name | skill-benchmark |
| description | Use when asked to evaluate whether a local or external skill is actually effective across representative prompts, multiple models, and baseline-vs-with-skill comparisons. Supports quick checks, formal benchmark runs, skill-vs-skill comparisons, and long-term trend reviews.
|
Skill Benchmark
Goal
Measure whether a skill produces stable, repeatable gains in trigger accuracy, routing clarity, and task outcome quality across representative prompts, multiple models, and baseline comparisons.
Workflow
- Confirm the candidate skill or skills to benchmark.
- Run
scripts/candidate_check.py to verify the benchmark target is usable.
- Choose the benchmark mode with
scripts/benchmark_level.py.
- Select or build evaluation scenarios that cover should-trigger, should-not-trigger, edge, and comparison prompts.
- Run baseline and with-skill comparisons with
scripts/run_real_benchmark.py, extract trace signals with scripts/extract_trace_signals.py, and score the raw run with scripts/judge_real_results.py.
- Aggregate summaries with
scripts/aggregate_results.py and update long-term trends with scripts/write_trend_summary.py.
Decision Tree
- If the user wants a fast confidence signal, run
quick-check.
- If the user wants a formal evaluation of one skill, run
benchmark-run.
- If the user wants to compare two skills or two versions, run
compare-skills.
- If the user wants to inspect accumulated historical results, run
trend-review.
- If the candidate is incomplete or invalid, stop and report the eligibility failure before benchmarking.
Constraints
- Do not claim a skill is universally effective based on one model or one prompt set.
- Do not confuse structural compliance with empirical effectiveness.
- Do not benchmark an incomplete candidate as if it were production-ready.
- Keep baseline and with-skill outputs clearly separated.
- Treat trend review as meaningful only when the result schema and scoring rules stayed consistent.
Validation
- Candidate eligibility must be checked before any formal run.
- Every benchmark run must record candidate identity, model, prompts, baseline results, with-skill results, scores, verdicts, and timestamp.
- Final conclusions must distinguish
effective, partially effective, not proven, and ineffective.
- Aggregated output must summarize verdict counts and preserve the source skill name.
Resources
scripts/candidate_check.py: verify whether a local path, remote URL, or external candidate can enter benchmarking.
scripts/benchmark_level.py: classify the request into quick-check, benchmark-run, compare-skills, or trend-review.
scripts/run_benchmark.py: write simulated raw benchmark runs for quick checks and schema testing.
scripts/run_real_benchmark.py: execute baseline and with-skill runs through a runner adapter and write real raw benchmark runs; defaults to the built-in claude-cli adapter when no custom executor path is provided.
scripts/extract_trace_signals.py: verify trigger and route evidence from real raw runs.
scripts/score_benchmark.py: turn a simulated raw run into a scored summary and verdict.
scripts/judge_real_results.py: judge real raw runs for trigger accuracy, routing clarity, output signal retention (per-signal with synonyms), task completion rate, and outcome quality.
scripts/aggregate_results.py: summarize groups of run or summary files into a stable benchmark result.
scripts/write_trend_summary.py: roll up summary files into a per-skill trend signal.
scripts/sync_benchmark_outputs.py: rebuild summary and trend outputs from a raw benchmark directory in one pass.
references/benchmark-dimensions.md: scoring dimensions for trigger, routing, outcome, robustness, and long-term consistency.
references/eval-scenarios.md: scenario design rules for should-trigger, should-not-trigger, edge, and comparison prompts.
references/scoring-rules.md: scoring and verdict rules.
references/result-schema.md: required result fields for raw runs, summaries, and trend analysis.
references/benchmark-workflow.md: execution order and storage rules for benchmark runs.
references/history-schema.md: long-term summary and trend record schema.
references/trace-verification.md: trace-based rules for validating real trigger and route evidence.
references/output-assertions.md: starter output assertion rules for claim decomposition, disagreement, and arbitration.
references/real-benchmark-scenarios.md: scenario patterns for route-sensitive and output-sensitive real benchmarks.