Rough stream of thoughts... no AI was used in the process

I tested five code review skills across Claude Code and Codex

A key piece of my development flow is a plugin I call gdev_cycle. It essentially defines a development cycle where the LLM aligns on a plan, proceeds with implementation, goes through up to 3 review rounds where the agent calls a subagent to review the code, and finally signs off with a QA activity.

I’ll write more about my flow in a separate post - but the gist is that forcing my agent through a review cycle and a clear QA activity allows it to get much closer to the end result I’m after.

Since the review cycle is so fundamental to the skill, I decided to take 3 real tasks I had and test five review skills out there to see which one is the best.

The Skills

  1. Vercel Code Review
  2. Thermo-Nuclear Code Quality Review
  3. Adversarial Reviewer
  4. Sentry Code Review
  5. review — plain review skill - a copy of Claude Code built-in /review

The Test

I used Codex to implement all 3 tasks in separate branches, then I reviewed them myself and identified all bugs and key issues I think need to be fixed.

Then I ran each skill on each task across Claude Code and Codex. So overall 30 sessions (5 skills * 3 tasks * 2 agents) - each session ran in isolation so the sessions can’t bias each other.

After all agents were done I ran another agent to compare the review against my own findings and rank each skill on five rubrics.

  1. Real-bug recall
  2. Precision
  3. Restraint against over-engineering
  4. Architecture and design insight
  5. Cost

The Results

The surprising part is that skills ranked consistently across tasks for the same agent. Honestly, I thought that the results would be kind of random. Some skills would be better just by chance and on every run a different skill would shine. That was not the case.

Skills consistently performed better or worse, for the same agent, but for different agents different skills worked better.

Overall scores

Overall scores · SVG

Five rubric scores

Five rubric scores · SVG

If you just want the bottom line: Use the Vercel Code Review skill across Codex and Claude Code.

My three takeaways from this short experiment:

  1. Model-specific prompt engineering still matters. The plain “review” skill that is a copy of Claude Code /review command performed the best across all 15 Claude Code sessions, but performed worst on Codex. Interestingly, that makes the case for the claim that vendor-specific harnesses win - the native Claude Code review performs best in Claude Code.
  2. No skill found all bugs - so a clean code review is not permission to ship on medium-large tasks.
  3. A combination of review + QA found all bugs.

← All posts