benchmark-guardian
Use for JS/TS performance-sensitive changes, benchmark additions, benchmark regressions, benchmark result JSON, or any claim that code is faster, smaller, cheaper, or not slower.
SKILL.mdskills/benchmark-guardian/SKILL.md
GitHubBenchmark Guardian
Do not make performance claims from intuition.
Workflow
State the benchmark hypothesis:
- target behavior,
- metric,
- fixture,
- command,
- acceptance threshold or baseline.
Prefer focused benchmarks while iterating:
npx evidence-kit add-benchmark --name core --language js
npm run bench:evidence
Write structured JSON to
benchmarks/results/*latest.json.Run correctness tests or fuzzers for the same path before saying a performance change is safe.
Update evidence docs:
npm run docs:perf
npm run docs:perf:search -- <benchmark topic>
Result Rules
- Keep timeout, error, unsupported, and bad-result rows visible.
- Report exact command and output file.
- Report fixture-level rows, not only aggregate scores.
- If benchmark scope changed, run
npm run bench:scopeornpx evidence-kit scope. - If the benchmark came from external research, cite the fetch manifest and source note path.
Final Response
Include the command, output path, key rows, and caveats such as dirty worktree, missing baseline, or local machine noise.