Claude Code: Multi-model ensemble opinions (Claude + Gemini + Codex)

ensemble-opinion.md

description
Get multi-model opinions on a problem (Claude + Gemini + Codex)

You are orchestrating an ensemble of AI models to analyze a problem from multiple angles.

Your Task

1. Identify the problem from conversation context
2. Run ALL THREE models in PARALLEL using the Task tool
3. Synthesize their responses

Step 1: Identify the Problem from Context

Review the conversation history and identify:

• What problem or decision is the user facing?
• What's the relevant context (codebase, constraints, goals)?

Restate the problem clearly and factually in 2-3 sentences. This exact problem statement will be sent to ALL models to ensure consistency.

If CLAUDE.md exists, read it and include relevant context in the problem statement.

Step 2: Formulate the Unbiased Prompt

Create ONE prompt that will be sent to all models. Structure:

Problem: [factual description - same for all models]
Context: [relevant codebase/project details from CLAUDE.md if exists]
Constraints: [any limitations]

Analyze this problem. Provide:
1. Your assessment of the situation
2. Risks and failure modes
3. Alternative approaches
4. Your recommendation

Be direct and critical. Identify blind spots.

Step 3: Run All Three Models in PARALLEL

Use the Task tool to spawn THREE agents simultaneously in a SINGLE message:

1. Claude Critic (subagent_type: claude_critic)

   • Pass: the unbiased prompt
   • Ask for: critical analysis, risks, blind spots, alternatives

2. Gemini Agent (subagent_type: general-purpose)

   • Prompt: "Run this command and return the output: echo '<UNBIASED_PROMPT>' | gemini 2>&1. If Gemini fails due to auth, return the error message."

3. Codex Agent (subagent_type: general-purpose)

   • Prompt: "Run this command and return the output: codex exec '<UNBIASED_PROMPT>' 2>&1 | tail -50. Return the model's response, stripping metadata headers."

IMPORTANT: Launch all three Task tools in a SINGLE response so they run in parallel.

Step 4: Synthesize

Once all three return, synthesize:

Ensemble Synthesis

Problem Analyzed

[The problem you identified]

Panel Responses

• Claude: [key points from claude_critic]
• Gemini: [key points, or "unavailable" if auth failed]
• Codex: [key points from codex response]

Consensus

What do multiple models agree on?

Disagreements

Where did they differ? Why?

Blind Spots

What did none address?

Recommendation

Your integrated recommendation.

Next Actions

• Actionable steps

---

Error Handling

• If Gemini fails (auth): note it, continue with Claude + Codex
• If Codex fails (timeout): note it, continue with available responses
• If only one model responds: still provide that opinion, note limitations
• Always report which models succeeded/failed

Important

• Use the SAME prompt for all models (consistency)
• Launch all 3 in PARALLEL (single message with 3 Task calls)
• Inject CLAUDE.md context directly into prompt (don't rely on AGENTS.md)
• Do NOT pre-bias external models with your opinions
• Your job is synthesis, not validation