2026.10.28
Compare Claude and Codex implementation plans without revealing their authors
Have both models create plans, compare them without revealing their authors, and draw on the strengths of each.
I had Claude and Codex each create an implementation plan, then asked both Codex and Claude to compare the plans without revealing which model had written which. The result was interesting…
Both preferred Codex’s plan. Their reason was that Codex’s plan followed Apple’s guidelines.
On the other hand, they recommended referring to Opus’s plan for other features, implementation details, and thoughtful touches.
So, for now, having different models make separate implementation plans and comparing them to take the best parts of each seems like a good approach.
Opus’s plan was intuitive, and its architecture was better than I expected. But it did not refer to the documentation very well and tended toward its own approach. Codex steadily goes and checks the right sources.
So even when Claude’s design is good, I think it is important to have Codex help with organizing documentation and reviewing, or to compare implementation plans to identify what is missing…