庙算 Miaosuan

夫未战而庙算胜者，得算多也。——《孙子兵法·计篇》

“The general who wins the battle makes many calculations in his temple before the battle is fought.”

An open-source daily strategy-review harness for AI agents. 一个每天逼你拍板的开源战略复盘 harness。

This is not another Art-of-War prompt pack. It is a working harness — a daily loop that reads what actually happened, interprets it through a pluggable strategic doctrine, forces every insight into an action, and compresses your stalled decisions into a 30-second decision card. The Seven Military Classics (武经七书) ship as its first doctrine pack.

这不是又一个兵法 prompt 包。它是一套真的在运转的 harness：每天读取真实发生的事 → 用可插拔的 doctrine 逐条解读 → 强制每条解读落到行动 → 把拖着的决策压成一张 30 秒拍板卡。武经七书是它出厂自带的第一套 doctrine。

---

What it looks like / 长什么样

When a decision stalls past 3 days, the harness stops analyzing and hands you this — answerable in one word:

(Example fully fictional. Note what it does: honest triage — it tells you this decision should NOT be forced today, and asks for the one word that actually is needed.)

Why a harness, not a prompt / 为什么是 harness 不是 prompt

Ask any LLM “analyze my company with Sun Tzu” and you get an essay. Essays don’t run companies. What compounds is a loop:

flowchart LR
    A["1 · GATHERwindow activity"] --> B["2 · INTERPRETdoctrine lenses"]
    B --> C["3 · FORCE每条解读必落所以应该 action"]
    C --> D["4 · UPDATEfive rollingstrategy files"]
    D --> E["5 · ESCALATEstalled decisions →30-second digest"]
    E -->|"human repliesin one word"| A

随口问模型"用孙子兵法分析我的公司",你得到一篇作文。作文管不了公司。能复利的是上面这个循环——解读被强制落成行动,行动沉淀进滚动的战略文件,拖住的决策被压成 30 秒拍板卡,人的一个字又喂回下一轮。

What’s in the box / 包里有什么

Quickstart (≤5 min)

Fastest path — let your agent install it. Clone the repo, open your coding agent in your workspace, and paste:

I cloned https://github.com/AI-Nate/miaosuan to ./miaosuan.
Install it: copy SKILL.md + doctrines/ + templates/ into this workspace's
skills location (.claude/skills/miaosuan/ for Claude Code), bootstrap
strategy/ from templates/strategy/, then walk me through filling in my
fronts and bets. Finally run a first review over this week.

Your agent reads the docs and does the rest. 克隆仓库,把上面这段话丢给你的 agent,它自己会装好并带你跑第一次复盘。

Manual path:

1. Clone this repo.
2. Copy templates/strategy/ into your agent’s workspace as strategy/.
3. Install SKILL.md as a skill — per-platform guides: docs/setup-claude-code.md · docs/setup-openclaw.md · docs/setup-any-agent.md (Cursor, Codex CLI, aider, anything that reads markdown).
4. Run it once manually: “Run miaosuan over what happened this week.”
5. Wire it to a daily cron. From now on your agent briefs you like a staff officer, not a chatbot.

The five strategy files / 战略五件套

Rolling files, updated every review — the agent’s institutional memory of strategy, separate from its memory of facts:

• 态势 (Situation) — where each front stands right now
• 主攻 (Main thrusts) — the 1–3 bets everything else serves
• 未决 (Pending) — decisions waiting on the human, with age counters
• 红线 (Red lines) — things the agent must never do or recommend
• 变更记录 (Changelog) — every strategy change, dated, one line

Doctrine packs / doctrine 插件

A doctrine pack = a folder of lenses. Each lens: a source quote, a plain-language restatement, triggering situations, and a 所以应该 action template. The harness picks the lenses that match the day’s events — it never dumps the whole book on you.

Ship your own pack via PR: doctrines//. Spec in doctrines/README.md. Wanted: stoic/, munger/, boyd-ooda/.

FAQ

Does this need any API keys or accounts? No. It’s markdown + your existing agent runtime. Is my data sent anywhere? No. Everything stays in your workspace. The harness reads what you give it. Why Chinese military classics? They are the densest decision-making literature ever written for operating under uncertainty with limited resources — and they’re public domain.

道 · Ethos & responsible use / 立意与负责任使用

The first of Sun Tzu’s five factors is 道 (dao) — the moral alignment that makes people move as one. It comes before terrain, weather, command, and method (天·地·将·法). A calculation without it is already a losing one, no matter how many sums you run.

孙子五事,道居其首——排在天、地、将、法之前。没有道的算计,算得再多,也是败局。

So the purpose of miaosuan is narrow and deliberate:

• Build, don’t harm. This harness exists to help you optimize your own life and work — and, through that, make the world a little better for the people around you. It is not a tool for harming, deceiving, or setting anyone up.
• Good is the stronger strategy — for you. If you are ever tempted to aim this at someone, remember what the classics actually teach: 得道多助,失道寡助. The legitimate, constructive path is not the “nice” option; it is the winning one. Trust, reputation, and goodwill compound; their opposites collapse and take you with them. Harm is a short trade that pays you back with interest.
• The harness holds this line. Its default doctrine reads your situation through 道 first. Ask it to help you do wrong and it will not draft the play — it will show you, in plain terms, why the better move is also the better business.

庙算的用途是收窄而明确的:帮你把自己的生活和工作做得更好,并借此让身边的世界好一点点——而不是拿来害人、骗人、设局坑人。真动了这个念头,记住兵法本身的教诲:得道多助,失道寡助。行善不是"善良选项",而是对你自己更划算的那条路——信誉、信任、善意会复利,反面会崩,还会把你一起拖下去。

Who built this / 出处

miaosuan is the distilled, sanitized version of the daily strategy-review loop behind AI Nate’s real one-person company — the harness that reads each day’s activity, interprets it through a doctrine, forces every read into an action, and compresses stalled calls into a 30-second card. It ran in production on a real company before it was ever open-sourced. The runtime is ours; the harness — and the doctrine spec — is yours: MIT (skill/code), CC-BY-4.0 (doctrine packs).

庙算是 AI Nate 真实「一人公司」背后那套每日战略复盘循环的提炼与脱敏版——先在一家真公司的生产环境里天天跑,跑通了才开源。运行时是我们的;这套 harness 和 doctrine 规范是你的:MIT(skill/代码)、CC-BY-4.0(doctrine 包)。

Learn to build this — join AI Superpower / 加入 AI 超能力社区

The harness hands you the loop; wiring it into how you actually run things — which lenses to trust, what to escalate, when to let a decision wait — is a craft you learn faster alongside people already doing it. AI Superpower is our community for learning and building with AI. Get the full walkthrough of how a real strategy-review loop is wired end-to-end, trade doctrine packs with other builders, and build alongside a live cohort.

→ Join AI Superpower: course-ai.app

harness 给你的是那个循环,但把它接进你真实的经营方式——信哪几条 lens、什么该升级成拍板卡、什么决策可以再等等——是一门手艺,和已经在做的人一起学会更快。 AI Superpower 是我们一起用 AI 学习和搭建的社区:看一套真实的战略复盘循环从头到尾怎么接线,和其他 builder 交换 doctrine 包,跟着直播 cohort 一起搭。

→ 加入 AI 超能力社区:course-ai.app

Acknowledgments / 致谢

The classical texts in books/ were studied from the Chinese Text Project (中国哲学书电子化计划, ctext.org) — an open-access digital library of pre-modern Chinese texts created by Dr. Donald Sturgeon. Every chapter file links back to its source page there. If this project is useful to you, ctext.org deserves a visit (and support).

books/ 中七书经文的研读底本来自中国哲学书电子化计划(ctext.org)——每个篇章文件头部都注明了对应的原文页链接。特此致谢。

License

Code and skill files: MIT. Doctrine documents (doctrines/): CC-BY-4.0. The classical source texts themselves are public domain; see Acknowledgments for the digital edition used.

---

Built by running it daily on a real company first. The examples here are fictionalized; the structure is exactly what runs in production.

Learning to build with AI? Come build with us → AI Superpower · course-ai.app