两仪 · Liangyi

太极生两仪 — The Supreme Ultimate generates the Two Polarities. — 《易经·系辞》 / I Ching, Great Commentary

🎴 Live Demo → wenboxia.github.io/liangyi — Interactive ink-wash workflow navigator

A multi-model adversarial product development workflow for Claude Code. One product idea generates two opposing expert views. You synthesize the final decision from their structural tension. The disagreement is the feature.

Why "Liangyi"?

In the cosmology of the I Ching (易经), one of the oldest classical Chinese texts, all complexity unfolds from a single origin through structured opposition. Taiji (太极), the unified whole, gives birth to Liangyi (两仪) — the two polarities of yin and yang. From their tension emerges Sixiang (四象), the four dimensions, and from those, the Bagua (八卦), the eight trigrams that describe every situation.

This workflow is a technical implementation of the same philosophical structure:

• One product idea (Taiji) splits into two opposing expert views (Liangyi)
• The two views generate four dimensions of analysis (Sixiang) — PRD, tech spec, dev guide, decision log
• Those ground the full product architecture (Bagua) — from concept to shipped code

Single-AI workflows fail because a single model fills its own gaps with its own biases. Liangyi refuses this. It deliberately creates two models with conflicting expert priors, lets them disagree productively, and puts you — the human — in the role that AI cannot occupy: the one who synthesizes.

The skill family

Liangyi is the orchestrator. It combines four standalone skills that can also be installed independently:

| Skill | What it does | Install alone | |-------|-------------|---------------| | expert-debate | Adversarial brainstorm with conflicting expert roles | npx skills add wenboxia/expert-debate | | blind-review | Context-free document audit + devil's advocate | npx skills add wenboxia/blind-review | | doc-generator | PRD, tech spec, CLAUDE.md, dev guide | npx skills add wenboxia/doc-generator | | decision-log | Human decision tracking + review | npx skills add wenboxia/decision-log |

Install Liangyi for the full pipeline with all modules:

npx skills add wenboxia/liangyi

Or install only what you need — each skill works independently.

How it works

Raw idea (Taiji 太极)
    ↓
Phase 0: Refine language
    ↓
Phase 1: Two AI models debate as conflicting experts (Liangyi 两仪)
         (e.g., Security Architect vs Growth Hacker)
    ↓
Phase 2: Single-blind review (new Model B session, zero context)
       + Devil's advocate on consensus (with complexity gate)
    ↓
Phase 3: Generate PRD → Tech Spec → CLAUDE.md → Dev Guide (Sixiang 四象)
       + AI-simulated user validation of PRD
    ↓
Phase 4: Development with instant expert summoning (Bagua 八卦)
    ↓
Output: 6 documents + ready-to-copy Model B prompts + decision audit trail

What makes this different

| Liangyi | Typical AI workflow | |---------|-------------------| | Structured expert role debates with professional bias | Generic "review my idea" prompts | | Single-blind review (reviewer has zero context) | All reviewers share full history | | Devil's advocate when models agree (with complexity gate) | Agreement = done | | Every human decision logged with reasoning | Human approves/rejects silently | | Modular — install one skill or all five | Monolithic, all or nothing |

Batteries included

All reference files ship with the skill — no configuration needed to start:

• references/expert-roles.md — recommended expert role pairs by product type
• references/skills-registry.md — template for your Skills inventory
• references/doc-templates.md — document structure templates

No API required

Model B prompts are generated as copyable text (auto-copied to clipboard when possible). No API keys needed. Works with free-tier access to any LLM.

Model B recommendation: Gemini is recommended because its training data diverges most from Claude's, producing higher-quality adversarial tension. GPT, Qwen, Doubao, DeepSeek, and other capable LLMs also work. The key requirement: use a different provider from Claude.

Session management

The workflow uses multiple Model B sessions. Label your browser tabs to avoid mixing contexts:

| Tab name | Purpose | Used in | |----------|---------|---------| | "Role B" | Model B with your idea context | Phase 1 | | "Blind Review" | Fresh Model B session, zero context | Phase 2 |

Usage

/liangyi                    # Full pipeline
/liangyi phase1             # Just the expert brainstorm
/liangyi phase3             # Just document generation
/liangyi setup              # First-time configuration

Output

| File | Purpose | |------|---------| | idea.md | Final product concept | | prd.md | Product requirements | | tech-spec.md | Technical architecture | | CLAUDE.md | Development conventions | | dev-guide.md | Step-by-step operator manual | | decision-log.md | Human decisions + AI review findings | | prompts/*.md | Ready-to-paste Model B prompts |

Philosophy

"One idea, split by opposing expertise, synthesized by human judgment. AI models should produce structured disagreement, not smoothed consensus. The I Ching taught this 3,000 years ago: complexity emerges from polarity, not uniformity."

Deep Dive

Read the full methodology whitepaper:

• 📘 Methodology — English
• 📗 完整方法论 — 中文

Includes phase-by-phase breakdowns, decision-gate rules, Model B prompt templates, session-management protocols, and the reasoning behind every design choice.

Changelog

See CHANGELOG.md for version history.

Current version: v1.1.0

License

MIT