Agent Skills
› NeverSight/learn-skills.dev
› benchmarking
benchmarking
GitHub用于测量和改进导入吞吐量、预览延迟、内存使用及导出性能，以验证项目SLO。通过定义场景、硬件配置和阈值，执行基准测试并生成包含指标表格的重复性报告，对比基线并提供优化建议。
触发场景
验证性能目标
比较实现方案或回归测试
生成里程碑的可重复性能报告
安装
npx skills add NeverSight/learn-skills.dev --skill benchmarking -g -y
SKILL.md
Frontmatter
{
"name": "benchmarking",
"description": "Use when measuring and improving import throughput, preview latency, memory usage, and export performance against project SLOs."
}
Benchmarking
Use This Skill When
- Validating performance targets.
- Comparing implementations or regressions.
- Producing repeatable performance reports for milestones.
Required Inputs
- Scenario definition (import, edit preview, export).
- Hardware profile and dataset size.
- Success thresholds (for example preview p95 < 50 ms).
Workflow
- Define benchmark harness and fixed dataset.
- Warm up and run multiple iterations.
- Record p50/p95, max, and memory footprint.
- Compare against baseline commit.
- Recommend changes ranked by expected impact.
Guardrails
- Keep runs reproducible (same resolution, same params, same hardware).
- Separate correctness tests from perf tests.
- Do not declare success from a single run.
Output Contract
- Benchmark report with metrics table.
- Regression/improvement summary.
- Follow-up actions if thresholds are missed.
版本历史
- c3c0a1e 当前 2026-07-23 10:50
