Dispute

GODMODE Cross-Model Dispute

Three independent AI models debate until consensus.

Topic: $ARGUMENTS

Participants

• Claude (you, the orchestrator) — Architect #1
• GPT (via mcp__codex__codex) — Architect #2
• Gemini (via mcp__gemini__gemini) — Architect #3

Step 0: Preparation

1. Analyze the topic. Determine what context is needed (files, code, prior decisions).
2. Read relevant files if needed.
3. Formulate a clear dispute question with full context.
4. Prepare the expert prompt for GPT and Gemini — they need FULL context since they don't see our conversation.

Step 1: Independent Positions (Round 1)

All three models give positions independently (no one sees others' answers yet).

Claude: Write your position directly in the conversation.

GPT: Send via mcp__codex__codex with:

• Full context of the dispute topic
• Role: "You are Architect #2 in a 3-way dispute. Give your INDEPENDENT position."
• developer-instructions: architect prompt
• sandbox: read-only

Gemini: Send via mcp__gemini__gemini with:

• Same full context
• Role: "You are Architect #3 in a 3-way dispute. Give your INDEPENDENT position."
• sandbox: read-only

Launch GPT and Gemini in parallel (same message, two tool calls).

Save all three positions. Present to user.

Step 2: Cross-Critique (Round 2)

Each model sees ALL other positions and critiques them.

Claude: Read GPT's and Gemini's positions. Critique both — find at least 2 weaknesses in each.

GPT (via mcp__codex__codex-reply or new session): Send:

• Claude's position (full text)
• Gemini's position (full text)
• Task: "Critique both positions. Find at least 2 weaknesses in each. Defend your own where you disagree."

Gemini (via mcp__gemini__gemini-reply or new session): Send:

• Claude's position (full text)
• GPT's position (full text)
• Task: "Critique both positions. Find at least 2 weaknesses in each. Defend your own where you disagree."

Launch GPT and Gemini in parallel.

Save all critiques. Present to user.

Step 3: Defense & Concessions (Round 3)

Each model responds to criticism and acknowledges valid points.

Claude: Read critiques from GPT and Gemini. Respond: what you accept, what you defend.

GPT: Send both critiques (from Claude and Gemini) of GPT's position. Task: "These are critiques of YOUR position. Respond: what do you concede? What do you defend and why?"

Gemini: Send both critiques (from Claude and GPT) of Gemini's position. Task: "These are critiques of YOUR position. Respond: what do you concede? What do you defend and why?"

Launch GPT and Gemini in parallel.

Step 4: Consensus Check

Analyze all three Round 3 responses. Categorize:

If no OPEN items remain → go to Step 6 (Final Document).

If OPEN items exist → go to Step 5 (Negotiation).

Step 5: Negotiation Rounds (repeat until consensus or max 2 extra rounds)

For each OPEN item:

Send to GPT and Gemini:

• Current state: who agrees on what, what's still open
• The specific point of disagreement
• Task: "Propose a compromise or explain why you cannot concede this point."

Claude also proposes compromise.

After each negotiation round, re-check consensus:

• Resolved → move to CONSENSUS
• Still open after 2 extra rounds → mark as UNRESOLVED, record all positions

Maximum total rounds: 5 (3 standard + 2 negotiation).

Step 6: Final Document

Write a summary document with:

# Dispute: [Topic]
Date: [date]
Participants: Claude (Opus), GPT (Codex), Gemini

## Consensus (all agree)
- [point 1]
- [point 2]

## Majority decisions (2 of 3)
- [point]: [majority position] — [dissenter] disagrees because [reason]

## Unresolved (if any)
- [point]: Claude says X, GPT says Y, Gemini says Z

## Final Recommendation
[synthesized recommendation based on consensus + majority]

## Key Arguments
[most valuable insights from each participant]

Present to user for final decision. Record in project MEMORY.md.

Rules

• All three models MUST participate. If GPT or Gemini MCP is unavailable, inform user and fall back to 2-model dispute.
• Never fake consensus — real disagreements are valuable.
• Each model gets FULL context of others' positions — no "telephone game".
• Orchestrator (Claude) must be intellectually honest — concede when GPT or Gemini make better arguments.
• GPT and Gemini prompts must include ALL relevant context (they don't see our conversation).
• Use multi-turn (threadId via *-reply) when possible to preserve context across rounds.
• Maximum 5 rounds total to avoid infinite loops.
• Present each round's results to user before proceeding to next.