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

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.