If we can't get it right, we who are theoretically trained and experienced in creating useful abstractions and highly cohesive yet non-coupled data processing systems?
Then how is handing off that particular hard problem (and it is a 'hard problem') to an LLM going to make anything better?
IMHO one underlying issue is we stopped teaching good old 'Systems Analysis' as a necessary skill for programmers. (There are other things to consider as well.)