A multi-model validation layer that uses local LLMs via Ollama to identify factual and logical errors in cloud-generated AI outputs.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install peer-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 peer-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 Peer Review?
The Peer Review skill serves as a critical quality assurance layer for AI-driven workflows. By leveraging local LLM inference through Ollama, it implements a fan-out architecture where multiple models—such as Mistral and Llama 3.1—independently critique the output of cloud models like Claude. This approach allows developers using Openclaw Skills to catch hallucinations, logical inconsistencies, and overconfident assertions before they impact production environments.
This skill is designed for high-stakes scenarios where accuracy is paramount. It doesn't just provide a second opinion; it orchestrates a swarm of local reviewers that aggregate their findings through consensus logic. This significantly improves the reliability of agentic outputs by providing a transparent, local, and cost-effective method for verifying complex technical analyses and creative content.
Peer Review Use Cases
- Validating high-stakes financial trade analyses or technical reports produced by cloud agents.
- Reviewing agent output quality in automated content pipelines before publishing.
- Testing the accuracy and reasoning capabilities of local models against cloud benchmarks.
- Detecting hallucinated citations or non-existent sources in long-form generated text.
- Providing a safety net for autonomous agents performing multi-step logical reasoning.
How Peer Review Works
- The primary cloud model (e.g., Claude) produces an initial analysis or response.
- The skill triggers a fan-out process, sending the input text to three distinct local models: Drift (Mistral 7B), Pip (TinyLlama), and Lume (Llama 3.1).
- Each model acts as a skeptical reviewer, analyzing the text for factual, logical, and structural errors based on a specialized prompt.
- Critiques are returned as structured JSON, identifying specific quotes, issue categories, and confidence levels.
- An aggregator script synthesizes the results, deduplicating flags and weighting them by model confidence to determine consensus.
- A final report is generated with a recommendation to publish, revise, or flag for human intervention.
Peer Review Setup
To integrate this into your library of Openclaw Skills, ensure you have Ollama installed and the required models pulled locally.
# Pull required local models via Ollama
ollama pull mistral:7b
ollama pull tinyllama:1.1b
ollama pull llama3.1:8b
# Ensure dependencies are available
sudo apt-get install jq curl
# Run a peer review on a single document
bash scripts/peer-review.sh <input_file> [output_dir]
Peer Review Data Schema & Taxonomy
The Peer Review skill generates structured metadata for every critique to facilitate automated decision-making.
| Attribute | Type | Description |
|---|---|---|
| category | String | Classification (factual, logical, missing, overconfidence, hallucinated_source) |
| quote | String | The specific excerpt from the source text being flagged |
| issue | String | Detailed explanation of the identified error or concern |
| confidence | Integer | The model's certainty score ranging from 0 to 100 |
All review logs and performance metrics are stored in the experiments/peer-review-results/ directory for long-term tracking of True Positive Rates (TPR).
Peer Review Advanced Features
- Consensus-based weighting that requires at least two models to agree before a high-confidence flag is raised.
- Batch processing scripts to run peer reviews across large corpora of text for benchmarking.
- Synthetic error generation via seed-test scripts to evaluate the efficacy of the reviewer swarm.
- Integration with the Reef API for programmatic access to the
/reviewendpoint. - Configurable model selection and error category thresholds to tune for specific domains.
SKILL.md
Loading
Related Openclaw Skills
Ops Hygiene
A comprehensive framework of standard operating procedures for maintaining AI agent security, memory health, and system vitals.
Objection Handling Framework
A specialized skill designed to transform sales objections into buying signals using data-backed frameworks and tactical empathy.
Lead Magnets: Free-to-Paid Conversion Framework
A strategic framework for designing high-value lead magnets that solve specific problems and drive prospects into paid conversion funnels.
Cold Outreach Intent-Based Conversion Framework
A high-performance framework for generating cold outreach sequences based on battle-tested conversion psychology and intent-based signals.
Pricing Psychology Strategic Framework
A strategic framework for designing pricing structures that convert using cognitive biases and proven psychological principles.
Proposal Writing — Confirmation-First Closing Framework
A strategic proposal generation tool that transforms verbal agreements into high-converting formal contracts using the McKinsey SCR framework.
Featured*
