三思而后行。想清楚,不等于想太久。
很多项目不是输在执行,而是输在行动以前:把猜测当成事实,把喜欢的方案当成唯一的方案,把一时顺利当成方向正确。
Foolish Think 是一套给智能体使用的项目思考 Skill。它不会急着替你写方案,而是先陪你多想一步。新点子可以用它打磨,成形的方案可以用它压力测试,卡住或失败的项目也可以用它追根溯源。
它的中文内核,可以归成一句老话:三思而后行。
这里的“三思”不是反复犹豫,而是三个不同的问题:
先照见事实,也照见自己的想当然。
Foolish Mirror 会分清哪些是事实,哪些是推断,哪些仍是假设;新项目做失败预演,旧项目找真实根因。信息不够时,它会停下来问,而不是编出一个完整答案。
一个问题被讲述的方式,往往已经偷偷限定了解法。
Foolish Reframe 会逐条挑战原方案的前提,寻找更简单、更稳、更省,也更容易长期维护的路径。每个新建议都必须说清它在解决哪个具体问题,不能拿一串“最佳实践”糊弄过去。
好点子不缺掌声,缺的是能让它尽早失败的实验。
Foolish Proof 把方向拆成一个个可验证的小步骤。上一步交付什么,下一步就接什么;怎么验收、何时复盘、失败损失多大、什么情况应该停,都事先讲明白。
然后才行动。
这也是 Foolish Think 对 “Stay hungry, stay foolish” 的理解:保持饥饿,所以愿意向前;保持愚拙,所以不把自己的第一反应当成真理。
智能体会根据你的真实意图和项目所处阶段自行判断,而不是机械匹配几个关键词:
- 只有需求或朦胧想法时,它进入
Seed,帮助你把问题想明白; - 已有方案、尚未重投入时,它进入
Stress-test,提前寻找脆弱处; - 项目已经卡住或结果不对时,它进入
Diagnose,区分方向错误、机制问题和执行偏差。
只是提到“项目”不会自动触发它。明确的执行任务、普通知识问答和单点技术故障,也不会被强行拖进三阶段分析。
下载 Release 压缩包或克隆仓库,在仓库目录运行:
./install.sh codex
./install.sh claude
./install.sh openclaw同时安装到三个智能体:
./install.sh all默认安装位置:
| 智能体 | 个人 Skill 目录 |
|---|---|
| Codex | ~/.codex/skills/foolish-think |
| Claude Code | ~/.claude/skills/foolish-think |
| OpenClaw | ~/.agents/skills/foolish-think |
安装脚本只复制 foolish-think/ Skill 目录。如果同名版本已经存在,会先自动备份。
也可以直接把仓库交给智能体:
请阅读 README.md,并把 Foolish Think 安装到你当前运行平台的个人 Skill 目录。安装后验证该平台能够发现 SKILL.md。
如果当前会话没有发现新建的顶层 Skill 目录,重启一次智能体。
显式调用最可靠:
请使用 Foolish Think,帮我三思这个项目。先完成第一阶段,与我确认后再继续。
安装后,智能体也可以在判断这套方法确实有帮助时自动调用。
默认情况下,它完成 Foolish Mirror 后会停下来与你对齐事实。只有你明确要求一次性分析,才会连续完成三个阶段。
把 GitHub 仓库或 Release ZIP 发给同事即可。他们不需要重新复制那段长提示词;安装 Skill 后,智能体会直接读取 SKILL.md。
Stay hungry. Stay foolish. Think before reality thinks for you.
Foolish Think is a portable Agent Skill for developing early project ideas, pressure-testing proposed solutions, diagnosing struggling initiatives, and reconstructing failed projects.
It follows three questions:
- Foolish Mirror — Is what we see actually true? Separate facts, inferences, hypotheses, and missing information.
- Foolish Reframe — Is there another path? Challenge the frame and trace every alternative to a concrete finding.
- Foolish Proof — What evidence earns the next action? Build falsification-first experiments and contract-linked steps.
The agent selects Seed, Stress-test, or Diagnose from the user's intent and project state. It does not rely on an exact keyword list, and merely mentioning a project is not enough to trigger the workflow.
Clone or unzip this repository, then run:
./install.sh codex
./install.sh claude
./install.sh openclawInstall for every supported agent:
./install.sh allDefault destinations:
| Agent | Personal skill directory |
|---|---|
| Codex | ~/.codex/skills/foolish-think |
| Claude Code | ~/.claude/skills/foolish-think |
| OpenClaw | ~/.agents/skills/foolish-think |
The installer copies only the foolish-think/ directory and backs up an existing installation before replacing it.
To let an agent install the skill, give it the repository and say:
Read README.md and install Foolish Think for the agent platform you are currently running. Verify that the platform can discover SKILL.md afterward.
Explicit invocation is the most predictable:
Use Foolish Think to examine this project. Complete Stage 1 first and wait for my confirmation.
Installed agents may also invoke it automatically when critical project development or validation is genuinely part of the request.
foolish-think/SKILL.md uses only portable name and description frontmatter plus platform-neutral instructions. foolish-think/agents/openai.yaml supplies optional Codex interface metadata; other agents can safely ignore it.
MIT