If you want to go all the way, a fully formalized specification is not less work than writing the software, though the potential promise of LLMs for formal methods is interesting.
But for any informal specs, no matter how detailed, while LLMs can be very capable when it comes to avoiding surface errors when translating them into code (e.g. they're more likely than human programmers to write code that compiles on first try), they still don't have a mental model, and most of the problems I've seen are IMO results of that and they increase with the scale of the project.