The thing is to understand that a model has "priors" which steer how it generates code. If what you're trying to build matches the priors of the model you can basically surf the gradients to working software with no steering using declarative language. If what you want to build isn't well encoded by the models priors it'll constantly drift, and you need to use shorter prompts and specify the how more (imperative).

In my experience, you need shorter prompts and steering it constantly for any kind of work, novel or not. You can be doing the most basic thing, if you let it iterate long enough, it will start doing something completely stupid. It's really sad watching Gemini CLI debugging something trivial, and trying to change it "one last time" again. Fortunately, Claude is better at this, but you still need to steer it. And try extremely hard not to give it pointers to something you DON'T want it to do, especially if it's easy and in it's training set, because it will so gladly pick it up.