The part I had backwards
For a long time I treated the model as the thing that would supply the thinking. I would describe an outcome at the level of a finished screen, ask for the whole thing, and get back something that looked plausible. It fell apart as soon as separate pieces had to agree with each other.
Git revert became a familiar habit.
The most expensive version of that mistake combined a broad goal, a small cheap model, and enough autonomy to keep going without stopping to report back. It kept going. It produced a great deal of work, very little of which functioned, and I paid for every token. I do not think that was the model's fault. I handed over something no junior could have scoped either, and then left the room.
What the speed does and does not do
It helps when I can describe an outcome and recognise a good answer. Drafting, comparing options, finding the gap in my own notes, carrying out work that is already defined.
What it will not do is decide what should exist, or tell me whether the result holds up. Those need a purpose, some understanding of the subject, a boundary, and a way to check. AI gives all of that more reach, including when it is wrong.
The uncomfortable conclusion
None of this removed a step I would have taken with a team of people. Working out what we are building, agreeing it, dividing it up, reviewing what came back. All still there.
What changed is that skipping those steps is now much faster, and the hole is much deeper by the time you notice you are in it.
I keep expecting to find the shortcut. Every time I think I have, I find out later what it cost.