A sophisticated orchestration pipeline that employs parallel model runs and structured cross-critique to deliver high-accuracy synthesis for complex reasoning tasks.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install multi-model-critique
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 multi-model-critique 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 Multi-Model Critique?
The Multi-Model Critique skill is a high-performance framework designed for Openclaw Skills to tackle ambiguous, high-stakes, or long-form reasoning challenges. By moving beyond single-model limitations, this skill implements a rigorous four-step deliberation loop: Plan, Execute, Review, and Improve. It facilitates a collaborative environment where multiple AI agents evaluate each other's work, leading to a final output that is significantly more robust than any single-model draft.
This skill is specifically optimized for scenarios where deep reasoning and factual precision are paramount. By integrating a weighted scoring rubric and peer-driven revision rounds, it ensures that the final response is backed by evidence and filtered for hallucinations. When you deploy this as part of your Openclaw Skills library, you enable a multi-agent system capable of self-correction and nuanced synthesis.
Multi-Model Critique Use Cases
- High-stakes strategic planning requiring diverse perspectives and rigorous validation.
- Complex technical documentation where accuracy and internal consistency are critical.
- Deep research tasks involving ambiguous data points and competing hypotheses.
- Architectural reviews or complex coding tasks that benefit from multi-agent peer feedback.
How Multi-Model Critique Works
- Parallel Draft Round: The system spawns multiple ACP sessions in parallel, tasking each model to generate an initial draft using a structured Plan-Execute-Review-Improve sequence.
- Cross-Critique Round: Each model's draft is shared with its peers for structured evaluation, focusing on strengths, weaknesses, and potential hallucinations.
- Revision Round: Models receive their peer critiques and perform a targeted revision to address gaps and incorporate improvements.
- Final Synthesis Round: The refined outputs are integrated into a single user-facing response that highlights the best answer, key improvements, and remaining uncertainties.
Multi-Model Critique Setup
To integrate this skill into your Openclaw Skills workflow, ensure your environment is configured with multiple agent IDs and the necessary Python utility scripts. Use the following commands to prepare your orchestration plan:
# Generate model-specific prompts for the multi-round workflow
python scripts/build_round_prompts.py
# Execute the orchestration local helper to build the run plan
python scripts/run_orchestration.py --models agent_id_1,agent_id_2,agent_id_3
Ensure that the references/ directory contains the required markdown templates for prompts, orchestration flow, and the machine-parseable output schema.
Multi-Model Critique Data Schema & Taxonomy
The skill organizes its orchestration and feedback data through a structured taxonomy to ensure reproducibility within Openclaw Skills.
| Data Component | Description |
|---|---|
| Input Configuration | Boolean complex flag, agentId list, and optional runtime controls like timeoutSec and budgetUsd. |
| Per-Model Scoring | Quantitative scores (1-5) for accuracy, coverage, evidence, and actionability. |
| Orchestration Plan | A JSON artifact documenting model mapping, round prompts, and deterministic execution steps. |
| Output Schema | Final synthesized answer paired with a log of key improvements and uncertainty notes. |
Multi-Model Critique Advanced Features
- Parallel orchestration using ACP sessions for concurrent model execution.
- Weighted scoring algorithm (0.40 Accuracy, 0.25 Coverage, 0.20 Evidence, 0.15 Actionability) for objective ranking.
- Automated cross-critique feedback loops that force models to justify rankings and identify peer hallucinations.
- Fault-tolerant execution that allows the pipeline to continue even if a single model fails.
- Support for machine-parseable JSON outputs defined by a strict schema for downstream automation.
SKILL.md
Loading
Related Openclaw Skills
TreeListy
A powerful hierarchical decomposition engine that transforms complex projects and information into structured, actionable trees using 21 specialized patterns.
Django REST Framework (DRF)
A specialized skill for Openclaw Skills that automates the scaffolding and optimization of Django REST Framework APIs.
OpenClaw Menu Bar
A native macOS menu bar application providing instant, popup access to the OpenClaw AI chat interface.
System Information Utility (opensysinfo-skill)
A lightweight Bash skill providing real-time system metrics including CPU, memory, and disk usage for host environments.
Enumeration
An automated data reporting tool designed to aggregate all active skill data and broadcast it to a designated Discord webhook.
Canvas LMS
A comprehensive integration for accessing Canvas LMS course data, assignments, and grades using the official REST API.
Featured*
