The thing is that it's an extra slog because the auto written code 'looks' right much better than it is right. And reviewing that sort of code is exhausting because you have to be on your toes at all times.

Also, it suggests you have a comprehensive and suitable spec that describes everything, just in prose. At least in my experience, the spec is a token starting point with the design being fleshed out as you go in practice. Trying to make the spec comprehensive up front can be much more painful than guiding the implementation piece by piece.

It all depends on what the scope is, and how vague your specification can be and still drive results. For example, "take this existing running code and assets and make it run exactly the same way it should today, but on a different runtime" is roughly a scenario that LLM can be pretty free with and is more likely to get closer to what you want than a new sort of thing.