Red Team — Multi-Model Adversarial Review

Run documents, pitches, plans, and ideas through multiple AI models for adversarial critique, then synthesize findings into actionable improvements. Saves Claude Code context by offloading heavy analysis to external APIs.

Available AI Engines

Tier 1: MCP-Connected (Ready Now)

Gemini Flash — mcp__gemini__gemini-query (model: "flash", thinkingLevel: "high")

• Best for: Fast adversarial critiques, regulatory analysis, market sizing
• Quota: Free tier — rate-limited. Use flash model (pro hits 429 quickly)
• Also available: mcp__gemini__gemini-analyze-text, mcp__gemini__gemini-brainstorm, mcp__gemini__gemini-search, mcp__gemini__gemini-structured, mcp__gemini__gemini-analyze-url, mcp__gemini__gemini-analyze-document

Codex (OpenAI o4-mini) — mcp__codex__codex

• Best for: Technical architecture review, code feasibility, implementation analysis
• Config: sandbox: "read-only", approval-policy: "never"
• Paid key — no rate limit concerns

Grok (xAI) — mcp__grok__* (tools TBD after first load)

• Best for: Contrarian analysis, market sentiment, X/Twitter-informed opinions
• Key source: Forge engine .env (/Users/tori/Documents/Repos/Forge/engine/.env)
• Package: @joemccann/xai-mcp-server
• NOTE: Requires Claude Code restart to activate. After restart, use ToolSearch with query "grok" to discover available tools.

Deferred

Kimi (Moonshot) — Not yet working (auth issues with API key)

• Requires Chinese phone verification for platform.moonshot.cn
• OpenAI-compatible API at https://api.moonshot.cn/v1/chat/completions
• Revisit when a valid key is obtained

Attack Protocols

Protocol 1: Full Red Team (6-角 Attack)

Use when: Major decision, grant application, business plan, investment pitch Run ALL of these in parallel:

1. VC Partner (Gemini Flash) — "You are a ruthless blockchain VC partner. Find every weakness..."
2. Technical Architect (Codex) — "You are a senior technical architect. Find technical flaws..."
3. Regulatory Attorney (Gemini Flash) — "You are a fintech regulatory attorney. Identify legal risks..."
4. Market Economist (Gemini Flash) — "You are an AI/crypto economist. Analyze unit economics..."
5. Contrarian / Market Sentiment (Grok) — "You are a contrarian analyst. What's the X/Twitter sentiment? What are people actually saying? What would the crypto community roast about this?"
6. Long-Context Stress Test (Kimi) — "You are a meticulous analyst with perfect recall. Read this ENTIRE document and find internal contradictions, unsupported claims, and logical gaps..."
7. Deep Research (Claude Task agent) — Check GitHub repos, financial papers, competitive landscape

Protocol 2: Quick Critique (3-角 Attack)

Use when: Feature spec, blog post, proposal draft, smaller decisions

1. Devil's Advocate (Gemini Flash) — "Find the 3 biggest flaws in this plan..."
2. Technical Review (Codex) — "Is this technically feasible? What breaks?..."
3. Market Check (Gemini Search) — "Search for competitors and validate claims..."

Protocol 3: Brainstorm & Refine

Use when: Exploring ideas, not yet committed to a direction

1. Gemini Brainstorm — mcp__gemini__gemini-brainstorm with Claude's initial thoughts
2. Codex Alternative — "Propose 3 alternative approaches to this problem..."
3. Synthesis — Claude combines insights from both

Protocol 4: Fact-Check & Validate

Use when: Verifying claims, checking numbers, validating market data

1. Gemini Search — mcp__gemini__gemini-search for real-time web data
2. Gemini URL Analysis — mcp__gemini__gemini-analyze-url for specific sources
3. Claude WebSearch — Cross-reference findings

Execution Pattern

Step 1: Prepare the Document

Read the file to be reviewed. Extract key claims, numbers, and assumptions.

Step 2: Fan Out Attacks

Fire all applicable model calls IN PARALLEL. Each gets a role-specific prompt with:

• The document content (summarized if too long for token limits)
• A specific adversarial persona
• Explicit instruction to be brutal and quantitative
• Request for a final verdict (fund/pass/conditional, score 1-10, etc.)

Step 3: Collect Results

Wait for all models to return. Note: Gemini Flash may hit rate limits — if 429, wait 60s and retry once with shorter prompt.

Step 4: Synthesize

Claude reads all attack results and produces:

• Consensus findings (what ALL models agreed on)
• Divergent opinions (where models disagreed)
• Kill shots (criticisms so severe they require fundamental changes)
• Actionable fixes (specific, implementable changes)
• Revised document (incorporating all valid criticism)

Step 5: Output

Write findings to context/red-team-[topic]-[date].md and the revised document.

Prompt Templates

Gemini Flash — VC Attack

You are a ruthless [DOMAIN] expert evaluating [DOCUMENT TYPE].
Your job is to find EVERY weakness, gap, and red flag. Be brutal.

[DOCUMENT CONTENT]

Attack from these angles:
1. [SPECIFIC QUESTION 1]
2. [SPECIFIC QUESTION 2]
...
Final verdict: [VERDICT FORMAT]
Be specific and quantitative where possible.

Codex — Technical Review

You are a senior technical architect and [DOMAIN] expert.
Critically evaluate this [DOCUMENT TYPE].
Find technical flaws, economic impossibilities, and strategic blind spots.

[DOCUMENT CONTENT]

Score 1-10 on: technical feasibility, market timing, team fit, fundability.

Gemini Brainstorm — Collaborative Refinement

claudeThoughts: "[CLAUDE'S INITIAL ANALYSIS]"
prompt: "Given these initial thoughts, what are we missing? What alternative approaches exist? What would make this 10x better?"
maxRounds: 3

Rate Limit Management

File Conventions

• Attack results: context/red-team-[topic]-[date].md
• Revised documents: context/[topic]-v[N+1].md
• Always preserve previous versions (never overwrite)

Adding New AI Engines

Currently Configured

Grok — MCP server in ~/.claude.json (@joemccann/xai-mcp-server)

• Key stored in ~/.claude.json env block
• Requires Claude Code restart after config change

Kimi — Bash curl (no MCP server needed)

• Key stored in /Users/tori/.claude/skills/red-team/.env
• OpenAI-compatible API, works immediately via Bash

Adding More Engines

For MCP-based engines, add to ~/.claude.json under mcpServers. For OpenAI-compatible APIs, add keys to /Users/tori/.claude/skills/red-team/.env and use curl.

Example Usage

User: "Red team this pitch"

1. Read the pitch file
2. Summarize key claims for token efficiency
3. Fire Gemini Flash (VC attack) + Codex (technical review) + Gemini Flash (regulatory) in parallel
4. Collect all results
5. Synthesize consensus, divergences, kill shots
6. Write revised version incorporating valid criticism
7. Present summary with ASCII charts showing before/after scoring