alt Hacker News> instructing once and for all my setup to do what I want next time.
This works up to a point, but eventually your "setup" gets complicated, some of your demands conflict, or have different priorities, and you're relying on the AI to sort it out the way you expect.
But setups get equally complicated, even with human software engineers. The approach that the OP is talking about applies only to experienced, good architect-level SWEs, and I suspect that the code quality and its problems are going to be the same whether they are directing LLMs vs a set of junior SWEs to write the code.
There is an inherent level of complexity in projects that solve some real world problem, due to all the code handling edge cases that were added incrementally over time.