cross-verify

AI Cross-Verification Skill

Collects and cross-verifies answers from Codex CLI, Gemini CLI, and Claude (Lead) on a given question, then synthesizes a comprehensive analysis.

Prerequisites

At least one of the following CLI tools must be installed:

• Codex CLI: npm install -g @openai/codex
• Gemini CLI: npm install -g @anthropic-ai/gemini or via Google's official installation

Workflow

Phase 1: Prompt Preparation

1. Extract the question/task to verify from the user's request
2. If code-related, automatically gather context:
   • git diff (if there are changes)
   • Relevant file contents
   • Project structure
3. Compose the prompt for each agent:
   • Question body
   • Gathered context
   • Instruction to respond in the user's language

Phase 2: CLI Availability Check

Before spawning agents, check which CLIs are available:

which codex 2>/dev/null && echo "CODEX_OK" || echo "CODEX_NOT_FOUND"
which gemini 2>/dev/null && echo "GEMINI_OK" || echo "GEMINI_NOT_FOUND"

• Both unavailable: inform the user that at least one CLI (codex or gemini) is required and stop
• One available: proceed with 2-model comparison (available CLI + Claude)
• Both available: proceed with 3-model comparison

Phase 3: Agent Spawn

Preferred: Agent Teams (parallel execution)

Try creating a team first:

TeamCreate(team_name: "cross-verify", description: "AI cross-verification")

If TeamCreate fails (Agent Teams not enabled):

1. Display: "Agent Teams requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in your settings.json env. Would you like me to enable it?"
2. If user declines, fall back to sequential Agent spawn (see below)

On success, spawn agents in a single message (parallel):

Codex reviewer (only if codex is available):

Agent(
  description: "Run Codex CLI analysis",
  prompt: "<composed prompt with context>",
  name: "codex-reviewer",
  subagent_type: "codex-reviewer",
  team_name: "cross-verify",
  model: "haiku",
  mode: "dontAsk"
)

Gemini reviewer (only if gemini is available):

Agent(
  description: "Run Gemini CLI analysis",
  prompt: "<composed prompt with context>",
  name: "gemini-reviewer",
  subagent_type: "gemini-reviewer",
  team_name: "cross-verify",
  model: "haiku",
  mode: "dontAsk"
)

Fallback: Sequential Agent spawn

If Teams is not available, spawn agents sequentially with run_in_background: true:

Agent(
  description: "Run Codex CLI analysis",
  prompt: "<composed prompt>",
  name: "codex-reviewer",
  subagent_type: "codex-reviewer",
  model: "haiku",
  mode: "dontAsk",
  run_in_background: true
)

Lead (Claude) Analysis

While agents are working, the Lead performs its own analysis on the same question.

Phase 4: Synthesis

After receiving all results, synthesize in this format:

## Cross-Verification Results

### Consensus (all models agree)
- High-confidence conclusions

### Unique Insights
- **Codex (GPT)**: Findings unique to this model
- **Gemini**: Findings unique to this model
- **Claude**: Findings unique to this model

### Conflicts (disagreements)
- Per item: each model's position + Lead's judgment

### Final Conclusion
- Synthesized recommendation

Phase 5: Cleanup

After output, send shutdown to teammates:

SendMessage(to: "codex-reviewer", message: {type: "shutdown_request"})
SendMessage(to: "gemini-reviewer", message: {type: "shutdown_request"})

Error Handling

Scenario	Action
CLI not installed	Skip that model, proceed with remaining
API error (ModelNotFoundError, etc.)	Skip that model, note in results
Timeout (agent doesn't respond)	Synthesize with available results
Both CLIs fail	Compare Lead analysis against error context
TeamCreate fails	Show activation guide, fallback to sequential Agent

Prompt Composition Rules

Request Type	Context to Include
Code review	git diff + relevant file contents
Architecture question	Project structure + key config files
Bug analysis	Error logs + related code
General technical question	Question only