In fact, I now prefer to use a purely chat window to plan the overall direction and let LLM provide a few different architectural ideas, rather than asking LLM to write a lot of code whose detail I have no idea about.
But it's far from perfect. Really difficult things/big projects are nearly impossible. Even if you break it down into hundred small tasks.
I've tried to make it port an existing, big codebase from one language to another. So it has all of the original codebase in one folder, and a new project in another folder. No matter how much guidance you give it, or how clear you make your todos, it will not work.
It simply forgets code exists during a port. It will port part of a function and ignore the rest, it will scan a whole file into context and then forget that a different codepath exists.
I would never rely on it for a 1:1 mapping of large features/code transformations. Small stuff sure, but beyond say a few large files it will miss things and you will be scratching your head for why it's not working.
How much of this is buildings versus recalling tutorials in the dataset. For every vibe coded project with 20 lines of requirements, I have a model with 20 different fields all with unique semantic meanings. In focused areas, AI has been okay. But I have yet to see Claude or any model build and scale a code base with the same mindset.