菜单
Run and interpret pvf headless performance diagnostics. Use for generating benchmark PDF fixtures, running `cargo bench --bench perf`, writing JSON reports, comparing results across PDFs or code changes, and explaining scenario metrics or regressions.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/shiguri-01/preview-pdf --skill bench命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Edit or review pvf developer docs under docs/. Use when changing docs/ files, trimming or migrating stale docs, deciding whether information belongs in docs, code comments, or tests, or checking architecture/reference/testing doc ownership and drift risk.
Add, update, place, or review pvf tests and validation. Use when writing regression tests, protecting behavior changes, choosing unit/module/integration test boundaries, reviewing test quality, changing test policy, or deciding validation commands.
Add, edit, rename, or review pvf palette behavior. Use when changing PaletteKind, PaletteProvider implementations, palette open options, provider snapshots, candidate rows, command-palette visibility, input modes, selection behavior, tab or submit behavior, palette-scoped key behavior, or palette rendering contracts.
基于 SOC 职业分类
正在显示 SKILL.md
| name | bench |
| description | Run and interpret pvf headless performance diagnostics. Use for generating benchmark PDF fixtures, running `cargo bench --bench perf`, writing JSON reports, comparing results across PDFs or code changes, and explaining scenario metrics or regressions. |
Use this skill to run, compare, or explain repo-local performance diagnostics.
Before changing benchmark behavior, report shape, scenario metadata, or
performance-diagnostics policy, read the relevant docs/reference.md
Performance Diagnostics section and docs/testing.md guidance on performance
diagnostics versus correctness tests.
For a straightforward benchmark rerun, inspect only the bench command, fixture,
and report context needed to answer the request.
Use benches/perf.rs as the bench entry point, benches/fixtures/ as the fixture source, and target/bench/ for generated fixtures and reports.
Infer the measurement goal from the conversation, then choose the smallest run that answers it. Run benchmark commands sequentially, not in parallel. For comparisons, keep variables unrelated to the comparison axis stable. Report only the metrics that explain the user's question, and include enough context to reproduce the result: command, fixture, scenarios, warmup, iterations, output path, and comparison axis.