Liangyi

Liangyi — Adversarial Product Workflow v1.0

Liangyi (两仪) /lyahng-ee/ — "The Two Polarities." From the I Ching (易经·系辞): "太极生两仪" — The Supreme Ultimate generates the Two Polarities. In Chinese cosmology, a single unified origin (Taiji) gives rise to two opposing forces (yin and yang), whose productive tension shapes all further complexity.

This workflow is built on the same structure: a single product idea generates two opposing expert views, whose structural disagreement you synthesize into your final decision. The tension is the feature.

Full orchestrator for the Liangyi workflow. Combines 4 modular skills into a complete flow.

This skill bundles all reference files in references/. No additional configuration is needed to get started.

Model compatibility: This skill runs in Claude Code (Model A = Claude). For the independent challenger (Model B), Gemini is recommended, but any capable LLM works — GPT, Qwen, Doubao, DeepSeek, etc. The key is using a different provider so training biases don't overlap. Throughout this document, "Gemini" is used as shorthand for "your chosen Model B."

Quick start

/liangyi

Individual phases can also be run:

/liangyi phase0    # Idea refinement
/liangyi phase1    # Expert brainstorm
/liangyi phase2    # Cross-validation
/liangyi phase3    # Document generation
/liangyi phase4    # Development kickoff

First-time setup (optional):

/liangyi setup

For standalone use of any module, install them individually:

• /expert-debate — just the brainstorm
• /blind-review — just the review
• /doc-generator — just the documents
• /decision-log — just decision tracking

Setup (first-time configuration)

When invoked with setup, guide the user through:

1. Model B selection: "Which LLM will you use as Model B? (Gemini recommended, GPT/Qwen/Doubao/DeepSeek also work)"

   • Save choice to references/config.md

2. Skills inventory: Read references/skills-registry.md and ask:

   • "These are the default recommended Skills. Do you want to add, replace, or remove any? (Enter to keep defaults)"
   • If user has changes, update references/skills-registry.md

3. Quick validation: Confirm the setup by listing:

   • Model A: Claude (fixed)
   • Model B: [user's choice]
   • Active Skills: [list from registry]
   • "Ready to go. Run /liangyi to start."

Setup is optional — the workflow runs with defaults if skipped.

Phase 0: Idea capture

1. Ask the user for their raw idea (voice transcript, rough notes, anything)
2. Generate a Gemini prompt for language refinement:

You are a prompt refinement specialist. I have a rough product idea description.
Please:
1. Remove filler words, repetition, and verbal tics
2. Only refine the language — do NOT change my intent or direction
3. Preserve all specific details and constraints I mentioned
4. Output a clean, structured description (under 200 words)

My raw idea: [paste raw idea]

Auto-copy to clipboard:

cat prompts/gemini-phase0.md | pbcopy 2>/dev/null || cat prompts/gemini-phase0.md | xclip -selection clipboard 2>/dev/null || echo "Prompt saved — please copy manually"

3. After user provides Gemini's refined version, review it together — confirm nothing was lost in refinement
4. Save to idea-raw.md

Phase 1: Adversarial expert brainstorm

Uses the same logic as the standalone expert-debate skill.

1.0 Select expert role pair

Read references/expert-roles.md for recommendations. Help user pick two roles with structural tension. Log to decision-log.md:

### [PROCESS] Decision: Expert role assignment
**Role A (Claude):** [name + why]
**Role B (Gemini):** [name + why]
**Tension point:** [where they conflict]
**What AI can't weigh:** [why this pair matters for *this specific* project]

1A: Generate proposal as Role A

Full instructions in expert-debate skill. Key points:

• Adopt persona completely, preserve professional biases
• Mark conflicts with generic advice explicitly
• Generate complete proposal with 5 components

1B: Generate Gemini prompt for Role B

Create prompt for a separate, clean Gemini session. Auto-copy to clipboard:

cat prompts/gemini-phase1.md | pbcopy 2>/dev/null || cat prompts/gemini-phase1.md | xclip -selection clipboard 2>/dev/null || echo "Prompt saved — please copy manually"

Tell the user: "Prompt copied. Open a new Gemini tab (label it 'Role B') and paste."

1C: Human comparison + decision log

Compare divergence points. Log each decision with the [PRODUCT] tag. Every entry must answer: "Why did I choose this? What couldn't AI weigh?"

Phase 2: Cross-validation & fusion

Uses the same logic as the standalone blind-review skill.

2A: Claude critique (investor role)

Drop Role A persona. Adopt skeptical investor. Critique the merged proposal:

• 3 weakest points + hardening suggestions
• Most likely cause of death
• Output improved version

2B: Single-blind Gemini review

Generate prompt for a brand new Gemini session (label it "Blind Review"). Contains ONLY the improved proposal — no original idea, no background.

Auto-copy to clipboard. Tell user which tab to use.

After receiving Gemini's response:

• Compare one-sentence summary vs original intent
• Classify divergence: None / Minor / Major
• Log as [REVIEW] type (AI-assisted finding, not human decision)

2C: Informed re-review (optional)

If 2B found major divergence, after revision → prompt for the original Gemini session (the one with context) to verify fix didn't drift.

2D: Devil's advocate (conditional)

Trigger: Claude + Gemini agree with no substantive disagreement. Complexity gate: Only for core architecture / strategic direction decisions. Skip for UI tweaks, copy changes, or well-scoped small features.

If triggered: hostile critic persona, find assumptions that will fail, identify competitor attack vectors, specify minimum validation requirements.

Output: idea.md (finalized)

Phase 3: Document generation

Uses the same logic as the standalone doc-generator skill.

Read references/doc-templates.md for templates.

3.1: PRD (prd.md)

Generate full PRD from idea.md.

3.1.1: User perspective validation

AI-simulated user reads the PRD. Results logged as [REVIEW] type — clearly labeled as AI simulation, not human decision.

If simulated user's understanding diverges from intent → flag before proceeding.

3.2: Tech spec (tech-spec.md)

Including Skills dispatch table referencing references/skills-registry.md.

3.3: CLAUDE.md

Code conventions, Git workflow, testing, agent boundaries.

3.4: Dev guide (dev-guide.md)

Step-by-step with Skill callouts and human decision point markers.

Phase 4: Development kickoff

1. Load CLAUDE.md and tech-spec.md into Claude Code
2. Follow dev-guide.md phase by phase
3. Log significant decisions with /decision-log
4. For technical dilemmas, use Best Minds instant summon: "Who in the world knows [domain] best? Answer as that person. Give me a recommendation with conviction, not 'both have pros and cons.'"

Output files

File	Purpose	Reader
idea.md	Final product concept	You
prd.md	Product requirements	You + Claude Code
tech-spec.md	Technical architecture	Claude Code
CLAUDE.md	Dev conventions	Claude Code
dev-guide.md	Operator manual	You
decision-log.md	Decision + review record	You / portfolio
prompts/*.md	Ready-to-paste Gemini prompts	You → Gemini

Session management

This workflow uses multiple Gemini sessions. Label your browser tabs:

Tab	Contains	Used in
"Role B"	Phase 1 Gemini (knows your idea)	Phase 1B, optionally 2C
"Blind Review"	Phase 2 Gemini (no context)	Phase 2B only

Never paste into the wrong tab — it defeats the single-blind mechanism.

When to simplify

Not every project needs the full pipeline:

Situation	What to skip
Small feature iteration	Skip Phase 1B. Run expert-debate solo in Claude.
Already have an idea	Skip Phase 0-1. Start at Phase 2 or 3.
Technical prototype	Skip Phase 0-2. Go straight to doc-generator.
Urgent deadline	Only use decision-log throughout. Everything else by instinct.
Quick brainstorm	Just install expert-debate. Skip everything else.