I was working on an issue that touched two different repos, and honestly I wasn’t familiar with either of them.
So I put both repos into one folder and opened it in Cursor:
workspace/
├─ repo1/
├─ repo2/
├─ opus.plan.md
└─ codex.plan.md
Then I tried a slightly different AI workflow instead of relying on just one model.
Step 1
I copied the issue description into Claude Code (Cursor extension) using Opus 4.5, and asked it to:
- explain the issue
- generate a comprehensive, detailed plan in ./opus.plan.md
Step 2
Then I switched to Codex 5.1 Max (extra high) and asked it to:
- review ./opus.plan.md
- give its opinion
- generate its own plan in ./codex.plan.md.
Step 3
Next, I opened the same workspace in Antigravity IDE, used Gemini 3 Pro (high), and asked:
- which plan is better and why?
Gemini explained why and said the Codex plan was better
Step 4
Finally, I used Codex again to implement the solution based on its own plan.
What I really liked about this setup:
- no copy-pasting between tools
- plan files live outside the repos, so no risk of accidentally committing them
- cleaner mental model when working across multiple codebases
I’ve noticed that when I rely on only one model, the result is often either:
- not fully completed
- or way over-engineered
But letting multiple models review and judge each other actually worked surprisingly well.
It’s a bit slower, sure — but the quality and confidence level feel much higher.
Curious if anyone else is doing something like this 
