I find both to be true. I use Claude for most of the implementation, and Codex always catches mistakes. Always. But both of them benefit from being asked if they’re sure they did everything.