What is the Verifying by Consensus skill?

It is a specialized workflow that dispatches multiple AI agents to review the same code independently, then aggregates their findings to ensure accuracy and minimize false positives.

How many agents are used during the review?

The skill uses heuristics based on scope (2 for single files, 3+ for architecture or security), but you can manually override this using the --count argument.

Where are the review results stored?

All individual reviews, collated reports, and cross-check validations are saved in the .work/ directory as timestamped Markdown files.

What happens if agents disagree on a finding?

Findings that aren't caught by all agents are moved to a Cross-Check phase, where a final agent validates the issue against the ground truth to confirm if it is a real problem.

Verifying by Consensus

Name: Verifying by Consensus
Author: tobyhede

bytobyhede

Security & Testing

Improves code reliability by dispatching multiple independent review agents and validating findings through consensus and cross-checking.

About

The Verifying by Consensus skill automates a multi-perspective quality assurance workflow within Claude Code to reduce false positives and identify subtle bugs. It dispatches a configurable number of agents to independently analyze changes, collates their findings based on frequency (Common vs. Exclusive), and employs a final cross-check agent to validate outliers. This approach is highly effective for high-stakes tasks like architecture shifts or security-sensitive updates where a single review pass might miss critical details.

Key Features

Configurable N-agent parallel dispatch for multi-perspective reviews
Automated cross-check phase to validate unique agent observations
Support for specialized agents and custom review prompts
Automatic consensus categorization (Common vs. Exclusive findings)
0 GitHub stars
Comprehensive audit trails saved as markdown reports in the .work directory

Use Cases

Validating complex multi-file architectural refactors
Conducting deep security audits on sensitive code paths
Reducing noise in automated code reviews by prioritizing consensus-backed issues

About

Key Features

Configurable N-agent parallel dispatch for multi-perspective reviews
Automated cross-check phase to validate unique agent observations
Support for specialized agents and custom review prompts
Automatic consensus categorization (Common vs. Exclusive findings)
0 GitHub stars
Comprehensive audit trails saved as markdown reports in the .work directory

Use Cases

Validating complex multi-file architectural refactors
Conducting deep security audits on sensitive code paths
Reducing noise in automated code reviews by prioritizing consensus-backed issues