An adversarial review system that uses two different AI models to stress-test and refine complex implementation plans autonomously.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install cross-model-review
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install cross-model-review using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Cross-Model Review?
The Cross-Model Review skill provides a robust framework for adversarial plan validation, ensuring that AI-generated implementation strategies are scrutinized by a second, independent model. By leveraging Openclaw Skills, developers can prevent common pitfalls in high-stakes areas like authentication, payments, and data modeling. This skill facilitates a rigorous 'second opinion' process that identifies edge cases and architectural flaws before a single line of code is written.
At its core, the skill supports both a static mode with fixed roles and a sophisticated alternating mode. In alternating mode, models swap between writer and reviewer roles across multiple rounds. This ensures that each model must eventually implement its own critiques, leading to natural convergence and higher-quality implementation plans that have been vetted by the best available LLMs.
Cross-Model Review Use Cases
- Reviewing implementation plans for sensitive systems like authentication or payment gateways.
- Sanity checking multi-step development tasks that are estimated to take over an hour to implement.
- Challenging complex data model changes to ensure architectural integrity and prevent regressions.
- Automating the refinement of AI-generated code plans through cross-provider model validation using Openclaw Skills.
How Cross-Model Review Works
- Initialize a review workspace using the CLI, specifying the target plan file and the primary models involved.
- The autonomous loop begins by spawning the first model to either write the initial plan or refine it based on context.
- A second model acts as a reviewer, identifying critical or high-priority blockers and outputting structured feedback.
- In alternating mode, the roles swap, forcing the new writer to address the specific concerns raised in the previous round.
- The state machine tracks progress through multiple rounds until both models reach an APPROVED status or hit the round limit.
- The finalization step aggregates the results, producing a verified plan, a changelog, and a technical summary of resolved issues.
Cross-Model Review Setup
To get started with this skill within the Openclaw Skills framework, use the following initialization command:
node review.js init \
--plan ./path/to/plan.md \
--mode alternating \
--model-a "anthropic/claude-3-5-sonnet" \
--model-b "openai/gpt-4o" \
--project-context "Internal API migration" \
--out ./tasks/reviews
Cross-Model Review Data Schema & Taxonomy
The skill maintains a structured workspace to track the review lifecycle. This organization is essential for complex Openclaw Skills implementations:
| File | Purpose |
|---|---|
| meta.json | Stores workspace state, including modes, models, and the current round counter. |
| issues.json | Tracks the full lifecycle of all identified issues, from discovery to resolution. |
| plan-final.md | The finalized, approved version of the implementation plan ready for execution. |
| summary.json | A high-level overview of the review process, including rounds taken and verdict. |
| round-N-response.json | Raw model outputs for each specific round of the adversarial loop. |
Cross-Model Review Advanced Features
- Fully autonomous orchestration that requires zero human intervention between review rounds.
- Cross-provider enforcement which mandates that models come from different families to ensure diverse perspectives.
- Automated issue tracking with built-in deduplication to prevent semantic drift across multiple iterations.
- Configurable token budgets and max-round limits to manage costs and execution time effectively.
- Integration ready for coding agents, allowing them to block implementation until the plan passes the Openclaw Skills review gate.
SKILL.md
Loading
Related Openclaw Skills
Calendly Quick Book
A natural language booking tool that schedules Calendly meetings instantly without manual link sharing or UI navigation.
dm-bot Agent Messaging
A secure, encrypted messaging platform specifically designed for AI agents to interact via DMs, group chats, and public posts.
MarsBit OpenNews
A high-performance MCP-based skill for fetching the latest cryptocurrency news, market flashes, and detailed reports from MarsBit.
MarsBit Crypto News Skill
A real-time Web3 and crypto-native news intelligence skill powered by the MarsBit MCP endpoint.
Repo Analyzer
A comprehensive GitHub repository auditing tool that provides automated trust scores and security due diligence.
Decodo Scraper
A comprehensive web scraping and search utility that provides AI agents with structured data from Google, Amazon, YouTube, and Reddit.
METADATA
Featured*
