菜单
Run a repository's README quickstart in a sandbox and report whether it actually works, correcting documentation errors but never masking real code bugs. Use when asked to verify a README, check a quickstart, prove install/usage commands run, or run "/proof verify".
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。
基于 SOC 职业分类
Verify that a repository's README quickstart actually runs, and capture a real demo of the project working. Use when the user says "/proof", asks to verify a README, check that a quickstart works, prove install/usage commands run, or generate a real demo GIF/screenshot for a repo. Routes to the proof-verify and proof-demo skills.
Capture a real demo of a project running — a terminal GIF for a CLI (via vhs) or a screenshot for a web app (via Playwright), with a real text-transcript fallback. Never fabricates or AI-generates the image. Use when asked to make a demo, record the CLI, screenshot the app, or run "/proof demo".
| name | proof-verify |
| description | Run a repository's README quickstart in a sandbox and report whether it actually works, correcting documentation errors but never masking real code bugs. Use when asked to verify a README, check a quickstart, prove install/usage commands run, or run "/proof verify". |
Extract the README quickstart (install + first run command), run it in a disposable sandbox, and classify the result. A documentation typo (wrong filename) is corrected and re-verified; a genuine code failure is reported, never hidden.
python "$HOME/.claude/skills/proof/run.py" verify <repo-path> --jsonoutcome field.verified → confirm it works.fixed → show the proposed_edits diff; apply only with --write after the user agrees.real_code_bug → report the failing step as a real bug in the project's code. Do not
touch the README.needs_input / needs_services / extraction_failed / sandbox_unavailable / timeout
→ relay the notes and suggest the next step.--no-fix — report only, never propose a correction.--write — apply a proposed documentation fix in place (off by default).--mode auto|docker|subprocess — sandbox mode (auto picks Docker when available).--timeout <seconds> — per-step time budget.real_code_bug appear to pass.