> Some times the true nature of the problem is revealed while implementing it and by deferring everything to an LLM you spend days throwing tokens at the wrong thing.

If you look at the code the LLM spat out (and you really, really should!) you will immediately notice that its shape is not what you thought it should be. You might not notice immediately if you're learning, but if you really "mastered" the domain, you will "just" see it. You will also recognize the problem with your assumptions, and immediately (or after some research) correct the prompt.

Looking at the diffs for everything the LLM does slows you down, of course. Not looking - or looking and not recognizing problems, for one reason or another - can be initially faster, but a single pathological case can eat both the time saving and tokens. For domains you truly "mastered", the current models can generate code as if they read your mind (because you can be that precise in the prompt, and quickly), so it's really glaring and very hard to miss when they somehow misread your mind.

It only works at the "mastered" / "unconscious competence" stages, and only in those narrow domains you truly mastered, but it does seem to work. Is it 10x? No idea, but there is a marked change in the speed boost when crossing from conscious to unconscious competence area, with everything else (harness, model) staying the same.