Loopwork Skill · Claude Code Edition

A drop-in Claude Code skill that turns a complete beginner's idea into working, continuously-evolving software — through a guided loop workflow.

🟢 OpenAI Codex 用户：请用姊妹版 loopwork-skill-codex。两版方法论同源，围栏按各自平台原生机制实现。

一个拖进 Claude Code 就能用的引导型 skill：不会编程的人说出一个想法，它以向导身份带他走完「访谈 → 白话规格 → 批准计划 → 自主循环开发 → 亲手验收」的全过程，然后把项目装进永续循环——你出方向，AI 跑圈，项目不是做完的，是转起来的。

Why

"Beginner-friendly AI builders" (Lovable / Bolt / Replit) and "engineering-grade Claude Code workflows" (superpowers / spec-kit / BMAD) are two lines that never crossed — the former are closed paid products, the latter assume you are an engineer, and both stop at delivery. Loopwork sits at the intersection and goes one step further: verified engineering discipline (spec-driven + test-first + autonomous loops + human acceptance), translated into plain language, packaged in one folder, with a perpetual improvement loop as the end state instead of a one-shot delivery.

	Lovable / Bolt	spec-kit / BMAD	superpowers	Loopwork
Install	paid website	CLI toolchain (Python/Node)	/plugin install	drag one folder
Audience	beginners	engineers	engineers	beginners
End state	one-shot delivery	one-shot delivery	one-shot delivery	perpetual loop
Discipline	platform-managed	mostly text conventions	workflow-enforced	hook-enforced (physical)

🇨🇳 中文快速上手（3 分钟）

第 1 步 · 下载：点本页绿色 Code 按钮 → Download ZIP → 解压。（会用 git 的话：git clone https://github.com/EricEEEEEEE/loopwork-skill-claude-code.git）

第 2 步 · 安装：把解压出来的 loopwork 文件夹（注意是里面那个 loopwork 文件夹，不是整个仓库）放进：

• macOS / Linux：~/.claude/skills/loopwork
• Windows：%USERPROFILE%\.claude\skills\loopwork

没有 skills 文件夹就自己新建一个。命令行党一步到位：

git clone https://github.com/EricEEEEEEE/loopwork-skill-claude-code.git /tmp/lw && mkdir -p ~/.claude/skills && cp -R /tmp/lw/loopwork ~/.claude/skills/

第 3 步 · 开口：打开 Claude Code（新会话），直接说你的想法：

我想做一个记账工具，但我不会编程

剩下的交给向导。你全程只做三件事：回答问题、看结果、点批准。中途随时可以关掉，回来说「继续」就能接上。

What a session looks like（示例对话）

你：我想做一个每天记录喝水杯数的小工具，但我完全不会编程

向导：收到，我来当你的向导。接下来你全程只做三件事：回答我的问题、看我给你的结果、对重要的事点头。写代码、跑测试、存档这些脏活全部归我。先聊 10 分钟想法—— 第一个问题：这个工具给谁用？ ① 就我自己（推荐，最简单）　② 我和家人　③ 给很多人用　④ 其他

你：1

向导：明白，那就不需要注册登录——少一大块复杂度。主要在哪儿用它？ ① 电脑浏览器（推荐）　② 手机　③ 都要　④ 不知道，你来定

你：4

向导：那我定：做成网页，电脑手机浏览器都能开，不用装任何东西。……（几问之后）……想法钉死了，项目简介 60 字如下……对吗？

你：对

向导：下一步我把它翻译成「做成什么样才算对」的验收清单——每一条都是你将来能亲手点一遍验收的句子。你批准计划之前，我一行代码都不会写。

…of course your session will differ — but the shape is always: interview → plain spec → your approval → autonomous loop → click-through acceptance → the perpetual improvement cycle.

Requires: Claude Code ≥ 2.1.196 recommended.

How it works（两个半场）

上半场 · 首航（直线，走一次）
0 开场体检 → 1 想法访谈 → 2 白话规格+项目规矩 → 3 摊开计划(你批准)
→ 4 循环执行 ⟲ → 5 验收(照单点一遍) → 6 收尾 → 进环仪式

下半场 · 循环模式（圆环，转无限次）
点火(30秒) → AI 跑批(挂机) → 收货(3-10分钟) → 续单(5分钟) → 再点火…

You only ever do three things: answer questions, look at results, approve. 用户全程只做三件事：回答问题、看结果、点批准。

What's inside

loopwork/
├── SKILL.md              # 向导本体：铁律 + 阶段路由（压缩存活区置顶）
├── references/           # 7 个阶段剧本 + 循环模式 + 快速通道 + 白话词典
├── scripts/              # 确定性机器
│   ├── init_project.sh   #   建家：git + 状态机 + 围栏接线
│   ├── verify.sh         #   验收裁判：exit code 说了算（fail closed）
│   ├── progress.py       #   状态机 + 进度卡（SessionStart 自动播报）
│   ├── guard_edits.py    #   围栏：实现期锁死考题/规格/规矩
│   ├── guard_bash.py     #   围栏：拦 rm -rf / force push / shell 绕道改写
│   └── stop_batch.py     #   挂机档：Stop 钩子自动续轮（外部计数器）
└── agents/reviewer.md    # 只读判卷员（验收前预检，N 对 M 点名）

Discipline is enforced, not suggested（纪律是硬的）

Industry lesson (68 documented failure cases): text rules get read, "understood", and ignored under pressure. So:

• Tests must fail first (red) before implementation; during implementation the test files are physically locked by a PreToolUse hook — including the known bypass via sed/echo/tee in Bash;
• Acceptance = verify.sh exit code, never the model's claim; unparseable = not passed (fail closed);
• Iteration caps are counted by external scripts, not by the model;
• Money / deletion / publishing / secrets: unconditional human gates, never mixed into "next step";
• Every task = one git commit ("存档点"), everything reversible.

Status

v1 built; guard machinery fully regression-tested (36-case suite in tests/, run bash loopwork/scripts/verify.sh in this repo); live cold-start scenarios simulated, real-beginner field test pending. Release gates in PROJECT.md §9. Until a real beginner completes a voyage + one solo cycle, treat this as beta.

• Design doc: PROJECT.md（含完整用户旅程、五道锁、设计依据）
• Ecosystem research: docs/research/（官方规范 / spec-driven 框架 / 循环纪律 / 小白引导，4 份调研）

Roadmap

• v1.x — real-beginner feedback, scheduled ignition (闹钟档), English-first pass
• v2 — Codex edition: shipped → loopwork-skill-codex

License

MIT — see LICENSE.