cross-model-review

What This Skill Does

The cross-model-review skill is an advanced adversarial orchestration tool for OpenClaw AI agents. It performs rigorous, multi-round reviews of technical implementation plans by utilizing two distinct LLMs. By pitting models against each other, the skill identifies logical flaws, edge cases, and architectural oversights that a single-model approach might miss. The core strength of v2 is its 'Alternating Mode,' where models swap the writer and reviewer roles in a fully autonomous, iterative loop. This forces the agent to take ownership of its critiques, ensuring that every piece of feedback is actionable and aligned with the project's technical requirements.

Installation

To integrate this skill into your environment, use the OpenClaw skill manager. Run the following command in your terminal:

clawhub install openclaw/skills/skills/don-gbot/cross-model-review

Ensure your agent environment has access to the configured models (e.g., Claude Opus and GPT-Codex) to allow the orchestration logic to spawn sub-agents effectively.

Use Cases

This skill is engineered for high-stakes software development. Use it specifically when:

• Designing security-critical features involving authentication flows or payment gateway integrations.
• Altering foundational data models where database migrations carry significant risk.
• Architecting complex systems requiring modularity and long-term maintainability.
• Developing implementation plans that are estimated to require more than one hour of execution time.

Example Prompts

1. "I'm planning to refactor our stripe-integration service to handle asynchronous webhooks; please run a cross-model-review on this plan."
2. "We need to add multi-factor authentication to the API. Can you perform an adversarial review to make sure the sequence diagram is air-tight?"
3. "Is this plan solid? I've outlined the schema changes for our user-preferences system, but I need a sanity check from a second perspective."

Tips & Limitations

• Avoid Overuse: Do not activate this for trivial tasks like writing quick scripts, fixing one-line bugs, or general research. The overhead of iterative rounds is inefficient for simple problems.
• Trust the Loop: In Alternating Mode, let the agent run until it reaches an 'APPROVED' state. Only intervene if the process hits 'max-rounds,' indicating a need for manual clarification.
• Context Matters: Always provide clear project context when initializing the skill; a well-calibrated reviewer is significantly more effective at identifying proportionality issues.