Liangyi — Productivity skill for Claude Code
One idea, two opposing expert views, one synthesized decision.
How to install Liangyi
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open wenboxia/liangyi and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Liangyi does
One idea, two opposing expert views, one synthesized decision. A Claude Code workflow for adversarial product development — inspired by the I Ching.
Alternatives in Productivity
- Product Team Skills — Product management team skills 5.3k ★
- /strategy — Run an end-to-end strategy workflow with decision-quality outputs 2.3k ★
- Repurpose — Workflow recipe — turn one blog post, video, or idea into a full platform-native content pack (thread, LinkedI 1.3k ★
README
两仪 · Liangyi
**一念生两仪,两仪成决策。**
一个想法进去,一份经过四轮对抗审查的方案出来。
- P1 对抗式方案生成:两位底模坐标不同的专家各自独立出方案,由执笔者融合成 v1。
- P2 四轮批判链:投资人批判 → 零上下文单盲复审 → 方向漂移检测与回退 → 前提拆台与致命论据分级。执笔与四轮批判各开独立窗口,互不共享上下文;单盲复审、前提拆台与专家 B 来自与执笔者不同的厂商、不同对齐路线。
- 多出口循环:拆台判定驱动回退——回 P1 整链重跑、回 P2 重压批判链、结构性死锁退出,两轮封顶。
- HITL 节点:在方向回退与拆台判定两个关键节点提供人工判断选项,auto / hitl 两档可切换。
为什么要这么做:**单个模型自己写、自己审,审不出自己的盲区。**
**在线体验 →** https://liangyi-five.vercel.app 把你的想法输进去,看它走完全部 13 步(演示档,每人每天 3 次)。
[English](README.en.md)
它解决什么
一个人拿着一个想法去问 AI,AI 会顺着他说。再问一次,还是顺着说。三轮之后方案很漂亮,但里面没有一处被认真反对过——**方案的边界是被 AI 的顺从画出来的,不是被人的判断画出来的。**
两仪把这件事拆开:出方案的和批判的必须是不同坐标的模型,四轮批判来自四个新开的窗口,每一轮只处理一条,最后由人(或规则)判「这还是我想要的吗」。
快速开始
pip install -r requirements.txt
cp .env.example .env # 填四个 key:OpenRouter / DeepSeek / Moonshot / Zhipu
python3 -m engine.run --check # 预检:硬规则 + provider
python3 -m engine.run -s scenarios/what-to-wear.yaml # 跑一条完整的链
跑完看结果:
python3 -m engine.inspect runs/<运行目录> # 每步谁在跑、花了多少
python3 -m engine.inspect runs/<运行目录> --diff idea-v1.md idea-v5.md
两个档位:`--mode auto` 全自动一次不停;`--mode hitl` 在两个边界决策点停下来问人,每处给两份旗舰模型的建议。
**网页同样支持两档。** 选「人工判断」后,链条会在「方向漂移回退」(2C)和「拆台判定」(2D)两处停下: 给你看诊断或分级表、「你当初要 X,现在变成 Y」的摘要,三个选项与终端一致,可附一段指令; 选「接受」不多花一次模型调用,改指令则那一步重跑一次。顾问材料包可下载,自己去问 GPT / Claude。 运行中可以停止、继续、失败时只重试那一步,跑完可下载整份运行记录。
在线入口本地起:`python3 web/server.py`,然后打开 `http://127.0.0.1:8765`。
十三步 · 状态机
**auto 档**(全自动一次不停)
flowchart TB
classDef anchor fill:#F0EEE6,stroke:#B0AEA5,color:#141413
classDef divA fill:#E4ECF2,stroke:#8FA3B3,color:#141413
classDef divB fill:#E8F0E6,stroke:#8FB08F,color:#141413
classDef hitl fill:#FBEFE9,stroke:#D97757,stroke-width:2px,color:#141413
classDef route fill:#FFF4D6,stroke:#C9A227,color:#141413
classDef exitA fill:#F8DADA,stroke:#B04141,color:#141413
classDef exitB fill:#DCEEF3,stroke:#3E7A8C,color:#141413
classDef exitC fill:#E9E8E4,stroke:#666,color:#141413
classDef
Related Skills
Dev Flow
A filesystem contract (.workflow/meta.json) + 39 agent skills that take a product from idea to production: web
Apple Product
Apple product development workflow
Product Dev Workflow
A flexible Claude Skill for the full product & software development lifecycle
Objection
Cross-examine a claim or a diff. Opposing counsel mode — one objection at a time, no approval until every obje
Claude Visualizer
Real-time ambient visualizer for Claude Code tool events — Tycho-inspired generative visuals and music. Zero d
ClaudeCode Workflow Lab
Complete Guide 2026: Claude Code Manual – Workflow Pipelines & Adversarial Budget Loops
Related Agents
Decision Loop
Autonomous PIVOT/REFINE/PROCEED decision agent — evaluates iteration results and autonomously decides whether
Kwb
You are inspired by Kent Beck — creator of Extreme Programming and Test-Driven Development, co-author of JUnit
Swe Dev Conceptual Intelligence Architect
Deep concept creation, cross-domain analogy, creative cognition, scientific discovery, design intelligence, ex