Automating AI Code Review with the Claudex Loop

The Claudex Loop Workflow

The Claudex Loop addresses the tendency of LLMs to overlook flaws in their own generated plans and code by introducing a secondary model, referred to as Codex, to act as an independent auditor. The process operates in four distinct phases: reconnaissance, interrogation, review, and build. During the reconnaissance phase, Claude performs web research to validate technical assumptions. The interrogation phase then establishes a shared understanding between the user and the model through a series of guided questions.

Iterative Review and Validation

In the review phase, Claude generates a plan.md file, which is then submitted to the secondary model for critique. This creates a feedback loop where the secondary model identifies issues, and Claude iterates on the plan until an approved state is reached. The author defaults this to a maximum of five rounds, though it can be extended if the models reach a deadlock. Once a plan is finalized, the build phase begins. After the code is generated, the secondary model performs a final audit, comparing the implementation against the original specifications to catch bugs such as concurrency issues, plain-text credential storage, or time-grid drift. This multi-stage verification process aims to reduce token waste by identifying architectural flaws before the code is deployed.