使用方法
- 复制安装命令，在自己的智能体环境安装前先检查源码包。
- 从原始文档核对所需工具、凭据和支持的宿主。
- 说明输入材料与期望交付物，并对照来源检查输出。
npx skills add https://github.com/emiliodominguez/flow-skills --skill flow-benchmark开始前需了解
- 安装不包含模型订阅、API 额度或访问文件的权限。
- 仓库星数和安装量反映活跃情况，不等于独立质量或安全测试。
按源仓库内容呈现，保留标题、案例、代码、表格、链接以及原文引用的演示图片。
Benchmark
Define what "better" means before measuring. Keep raw results and compare like with like; never report a single attractive timing.
Process
- State metric, workload, target and the decision it informs. Define the timing boundary and
cache/process state: cold startup includes initialization; steady-state throughput may warm up. Keep cold and warm experiments separate.
- Record baseline revision, runtime/build settings, hardware/environment, fixtures,
concurrency, sample unit and tool/version. Time setup only when it is part of the metric; never warm away the behavior under investigation.
- Run baseline and candidate under controlled conditions, changing one variable per
experiment. Interleave A/B samples when drift matters; isolate data/state between samples.
- Collect enough observations for the claim. Report median, tails such as p95/p99 only when
sample size supports them, variance, outliers and known noise. No false precision, no unexplained outlier removal.
- Check tradeoffs in memory, CPU, responsiveness, bundle size and complexity. Profile when the
result needs an explanation. A lower score is not automatically a better user experience.
- Save the harness or command, raw results and comparison so others can rerun it. State
whether the effect is large relative to noise.
Anti-patterns
- Warming a cold-start benchmark or excluding the initialization under investigation.
- Comparing development and production builds as if only the implementation changed.
- Reporting a best run or p99 from a handful of samples.
- Claiming causation from timing alone.
Done when
- Baseline and candidate share a relevant reproducible workload and timing boundary.
- Raw results, uncertainty and meaningful regressions are visible.
- The evidence supports an optimization decision or names what better measurement is needed.
Use flow-work or flow-refactor only for an optimization the evidence justifies.
