菜单
Use to run a zero-dependency sandbox loop that proves the loop machinery works end to end. Triggers on "run the demo loop", "test the loop", "quick loop test", "demo loop", "make the sandbox tests pass". Fixes intentionally-broken code until `python -m unittest` passes — no GitHub/CI/MCP needed.
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/az9713/loop-engineering --skill demo-loop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。
Use to run the parallel-worktree demo loop — multiple makers fix different modules concurrently, each in its own git worktree, then an integrator merges and runs the full suite. Triggers on "parallel loop test", "run the parallel sandbox", "worktree loop demo", "parallel worktree loop". Zero external dependencies; demonstrates worktrees as a load-bearing block.
Use after an incident/outage to draft a blameless postmortem with a verified root-cause hypothesis. Triggers on "draft a postmortem", "incident writeup", "root cause analysis", "RCA", "what caused the outage". Investigates read-only, red-teams the cause, and requires human sign-off before publishing.
Use when documentation is stale or drifting from the code — to reconcile docs with the current source. Triggers on "reconcile docs", "docs are stale", "update docs to match code", "doc drift", "fix outdated documentation". Rewrites affected pages and verifies every example runs.
正在显示 SKILL.md
基于 SOC 职业分类
| name | demo-loop |
| description | Use to run a zero-dependency sandbox loop that proves the loop machinery works end to end. Triggers on "run the demo loop", "test the loop", "quick loop test", "demo loop", "make the sandbox tests pass". Fixes intentionally-broken code until `python -m unittest` passes — no GitHub/CI/MCP needed. |
A self-contained loop for watching the machinery converge. It fixes the three
intentional bugs in loops/E-demo-sandbox/sandbox/calc.py until the tests pass.
No connectors, no worktrees — just the local filesystem and Python's stdlib
unittest. This is the irreducible core of a loop: automation + skill +
maker/checker + memory + a verifiable stop condition.
python -m unittest test_calc (run inside loops/E-demo-sandbox/sandbox/) exits 0.
That exit code is the whole backpressure signal — RED means keep going, GREEN means stop.
loops/E-demo-sandbox/state/demo-state.md to see prior runs and any already-known fixes.python -m unittest test_calc -v in the sandbox dir and report the failing tests + assertion messages. (If already GREEN, stop — nothing to do.)sandbox/calc.py to fix the bugs (never the test file).demo-state.md./demo-loop (Claude) / $demo-loop (Codex), or on a cadence with /loop 10m /demo-loop to watch it idle GREEN then react after a reset.demo-maker (edits code) vs demo-checker (runs tests, no edit). The maker never grades its own work.state/demo-state.md.| Excuse | Rebuttal |
|---|---|
| "I can see the bug, I'll just fix it without running the tests." | The RED→GREEN exit code IS the loop. Run the checker first and last, or you have no proof and no stop signal. |
| "I'll edit the test so it passes." | Never edit test_calc.py. The test is the spec; fix calc.py. |
| "It passed once, we're done forever." | Memory records the run; a reset re-breaks it. The loop must re-derive GREEN from the current RED, not from last time. |
pwsh loops/E-demo-sandbox/reset.ps1 (Windows) or bash loops/E-demo-sandbox/reset.sh restores the broken state.