Maybe bigger models will always just get progressively better at writing out bigger and bigger programs to the point that we're one-shotting fully functioning operating systems with wide driver support in the future, but again these features just don't matter if that's not what you want to do.
The reason you can't one-shot something complicated is that you haven't generally considered all the important choices. You have to be asked for every choice, and complicated things have a lot of choices. But I think you can do something waterfall-esque, which used to be a big no-no in programming, if you just list out all the decisions meticulously.
I've had some success with doing that, just talking it over and thinking a lot about design decisions, making a document, and then letting it loose on the implementation. Inevitably, you find that there were contradictions or missing choices in the design, and you have to answer some questions.
So, not one-shotting, but it's my fault, not the AI's fault.
You can of course tell it to just fill in the gaps with whatever it thinks is reasonable, and then you'll have something potentially quite different from what you imagined.