Multi-AI Debate Request Skill

Purpose

Initiate intensive debate between Claude CLI + Gemini API + Perplexity for critical technical decisions

Debate Protocol

Round Structure

1. •Round 1: Claude proposes → Gemini reviews
2. •Round 2: Gemini alternative → Claude rebuts
3. •Round 3: Both sides compromise → Calculate consensus
4. •Round 4: Perplexity judgment (if consensus < 70%)

Consensus Criteria

• •≥85%: Auto-adopt decision
• •70-85%: User review required
• •<70%: Extend debate or trigger Perplexity

Usage

# Basic debate
/debate "RTL multiplication optimization methods?"

# Expert mode (includes Perplexity from start)
/debate --expert "NoiseComputer architecture design"

# Quick debate (max 2 rounds)
/debate --quick "Variable naming convention"

# Save results to specific file
/debate --output decisions/my_decision.md "Topic here"

Debate Flow

User Question
    ↓
Claude analyzes & proposes
    ↓
Gemini reviews & counters
    ↓
Calculate consensus score
    ↓
┌─────────────┬──────────────┬─────────────┐
│  ≥85%       │  70-85%      │  <70%       │
│  Auto-adopt │  User review │  Perplexity │
└─────────────┴──────────────┴─────────────┘

Consensus Scoring

Formula: 0.6 × embedding_similarity + 0.4 × keyword_overlap

• •Embedding similarity: Cosine distance between final positions
• •Keyword overlap: Shared technical terms / total unique terms

Output

Results saved to:

1. •GitHub: docs/brain/DECISIONS.md
2. •Vertex AI BigQuery: knowledge_base.embeddings
3. •Vertex AI GCS: gs://bucket/decisions/debate_TIMESTAMP.json

Configuration

Edit config/debate_config.yaml:

• •max_rounds: Maximum debate rounds (default: 4)
• •consensus_threshold: Auto-adopt threshold (default: 0.85)
• •expert_threshold: Trigger Perplexity threshold (default: 0.70)

Implementation

Python script: .claude/skills/debate-request/debate_engine.py

• •Claude API via Anthropic SDK
• •Gemini API via google-generativeai
• •Perplexity API via requests
• •Consensus calculation with embeddings + keywords