Use one model to build quickly and a second model to review the diff for plan mismatches, code smells, and better architecture before applying the fixes.
Source episode · 37:00
How I AI: CJ Hess on Building Custom Dev Tools and Model-vs-Model Code Reviewswith Claire Vo
CJ sends Claude’s implementation and planning artifacts to Codex, which catches a visual mismatch, a hook issue, and refactoring opportunities before applying the fixes.
A specific review report and a cleaner implementation that still matches the approved behavior.
Anthropic AI assistant
OpenAI's cloud-based AI software engineering agent that can execute code, run tests, and handle complex multi-file tasks autonomously.
Step by step
Follow the sequence once, then adapt the prompts, checks, and handoffs to your own setup.
4 steps
Use your primary coding agent to implement the feature from the available plans or acceptance criteria. Get to a working diff before switching models.
Implement [feature] using the attached plan and mockups. Run the relevant checks, then summarize the files changed, any assumptions you made, and anything that still differs from the plan.
Keep the original planning artifacts in the repository so the reviewer can compare intent with implementation.
Give the current diff and the original planning artifacts to a second model. Ask it to review plan alignment, code quality, and whether a cleaner design would improve the codebase.
Review the current git diff and the attached plan or diagram artifacts. Report on: 1. Does the implementation accurately reflect the plan? 2. Are there correctness issues or code smells? 3. If we rebuilt or refactored this feature, what approach would improve the codebase? Cite the relevant files for each finding and rank findings by impact.
A separate model is useful here because it brings different defaults and is less invested in defending the first implementation.
Separate confirmed defects from optional cleanup. In CJ’s example, the useful findings included a spinner alignment mismatch, a missing hook dependency, and opportunities to extract components and constants.
Reject findings that cannot point to the diff, the plan, a failing check, or a concrete maintenance cost.
Have the reviewing model apply the approved fixes, then rerun tests and compare the result with the original artifacts.
Apply the confirmed correctness fixes and the approved refactors from your report. Preserve the intended behavior. Run the relevant tests, inspect the final diff, and list anything you chose not to change.
Turn an idea into a PRD, user stories, and a plan.
After the steps
How to recover when the loop fails and where human judgment helps.
Recover
Keep building
A guide for solo entrepreneurs to manage business operations using AI. Learn how to source manufacturers, build and maintain an e-commerce site, and replace dashboards with voice commands, all without a human team.
Turn a complex, artistic concept into a manufacturable 3D model. This workflow shows how to use AI to generate a design from a photo and then create CAD files for 3D printing, even for previously 'unmanufacturable' items.
Learn to translate fashion sketches into a full suite of marketing visuals. This workflow covers creating a detailed design spec, generating a core product photo with AI, and iterating for different contexts like runway and influencer shots.
Join 100,000+ product managers who use ChatPRD to write better docs, align teams faster, and build products users love.