> The hard part is always going down a rabbit hole in established code bases.

Actually, I found that this is exactly where they shine (I wouldn't trust them with greenfield implementation myself). Exploring existing code is so much easier when you can ask them how something works. You can even ask them to find problems - you can't trust them to be correct, of course, but at least you get some good brainstorming going. And, incredibly, they often do find actual problems in the code. Pretty impressive for language models.