In my experience these models (including opus) aren’t very good at “improving” existing code. I’m not exactly sure why, because the code they produce themselves is generally excellent.