Reduce AI hallucinations through multi-model consensus.
English | 中文
Part of CC Suite - The Standard Library for Claude Code Workflows
A Claude Code skill that queries Claude, Codex, and Gemini simultaneously, then has each model critique the others' answers to produce more reliable conclusions.
| Problem | Solution |
|---|---|
| Single AI models hallucinate | 3 models cross-verify each other |
| No way to know if answer is wrong | Disagreements surface errors |
| Blind spots in reasoning | Different models catch different issues |
Result: Higher confidence answers for critical decisions.
git clone https://github.com/leiMizzou/CrossCheck.git && cd CrossCheck && ./install.shTip: Use
./install.sh --forceto reinstall/update CLIs if you have older versions.
Then login and restart:
codex # Login to OpenAI (first time only)
gemini # Login to Google (first time only)
/exit # Restart Claude Code to load MCP tools/crosscheck <question>
│
▼
┌────────────────────────────┐
│ Round 1: Independent │
│ Claude + Codex + Gemini │ ← 3 models answer independently
└──────────────┬─────────────┘
│
┌───────┴───────┐
│ High Consensus?│ ← Smart Short-Circuit
└───────┬───────┘
No │ Yes → Fast Consensus Mode (skip R2/R3)
▼
┌────────────────────────────┐
│ Round 2: Cross Review │ ← Each critiques the other two
│ Find errors & blind spots │
└──────────────┬─────────────┘
▼
┌────────────────────────────┐
│ Round 3: Final Conclusion │ ← Synthesize consensus + corrections
└────────────────────────────┘
| Feature | Description |
|---|---|
| Smart Short-Circuit | Skip Round 2/3 when all models agree (saves time & cost) |
| Fault Tolerance | Continues with 2 models if one fails (degraded mode) |
| Structured Output | Clear tables with model status and verification mode |
| Auto Logging | Saves results to logs/ for audit trail |
/crosscheck Do we still need vertical models in the era of LLMs?
/crosscheck What's the best approach for real-time collaboration?
/crosscheck Is microservices the right architecture for this project?## Cross-Check Results: Do we still need vertical models?
**Mode**: Full Verification
### Round 1: Independent Answers
| Model | Status | Core View | Key Points |
|--------|--------|-----------|------------|
| Claude | OK | Yes, for regulated industries | Data privacy, compliance |
| Codex | OK | Depends on scale | Cost vs accuracy tradeoff |
| Gemini | OK | Hybrid approach | Combine general + specialized |
### Round 2: Cross Review
- Codex found Claude overstated compliance requirements
- Gemini caught outdated cost assumptions in Codex's answer
- Claude identified missing edge cases in both
### Round 3: Final Conclusion
| Consensus | All agree vertical models valuable for healthcare, finance |
| Corrections | Updated cost analysis based on reviews |
| Final | Hybrid approach recommended for most use cases |Click to expand
# 1. Install CLIs (with version pinning)
npm install -g @openai/codex@0.1 @google/gemini-cli@0.1
# 2. Add MCP servers
claude mcp add codex -- codex mcp-server
claude mcp add gemini -- npx -y gemini-mcp-tool@0.1.8
# 3. Copy skill
mkdir -p .claude/skills/crosscheck
curl -sL https://raw.githubusercontent.com/leiMizzou/CrossCheck/main/.claude/skills/crosscheck/SKILL.md \
-o .claude/skills/crosscheck/SKILL.md
# 4. Login & restart
codex && gemini && echo "Now run /exit to restart Claude Code"Note: Requires Node.js and Claude Code CLI installed.
claude mcp list
# codex: codex mcp-server - ✓ Connected
# gemini: npx -y gemini-mcp-tool - ✓ Connected| Problem | Solution |
|---|---|
| MCP "Failed to connect" | Run codex or gemini to login first |
| Tool not available | Restart Claude Code: /exit then claude |
Questions or suggestions? Join our WeChat group:
Scan to Join Group |
Add Helper if QR Expired |
MIT