Tutorial · open-skills

Have your code reviewed by a second LLM

Two eyes see more than one — that holds for AI too. One model writes the code, a second model reviews it independently instead of defending its own work.

More on the why: "How I have my code reviewed by a second LLM"

Agent file as a project rule: the first model sets the roles without cloning the project instructions.
Agent file as a project rule: set the roles without cloning the project instructions.
How it works

Seven steps.

  1. Separate the roles. The reviewer never changes product code — it writes audit logs and plan files.
  2. AGENTS.md into the repo. The role rule stays local per project (each repo has its own protected areas).
  3. Install the skill globally. Copy the review skill once to ~/.codex/skills/codex-claude-review-loop/ — then use it in any repo.
  4. Let it implement. Model A writes or changes code.
  5. Let it review. Model B runs the skill, forms risk clusters (security, architecture, tests, docs, release) and writes an audit + plan.
  6. Chat handoff. Audit and plan go back to Model A — with the included handoff template.
  7. Fusion plan. Model A doesn't follow blindly: it writes its own plan, compares both and only then implements the fusion — verified with tests/lint/build.
The review skill in action: autonomous cluster scans, a file-level done list, a mandatory handoff instead of blind adoption.
The review skill in action: cluster scans, a file-level done list, a mandatory handoff.
Kit

Download the kit

AGENTS.md, the Codex review skill, plan files and the chat-handoff template — ready to take with you.

View on GitHub →