What's this about?
dotey observes that the more programming experience you have, the less comfortable you are letting AI write and verify code unsupervised — like supervising an intern, always worried they will break the codebase, when they are actually quite good. He only genuinely let go in a domain he does not know well: writing Swift and AppKit, where he can read the code but is not more expert than the model. The quality turned out to be good, most details were not worth agonising over, and he now barely reads what it writes — reviewing at a high level, testing, and shipping. Architecture and module design he still decides jointly with the model, and settling those early saves considerable trouble later.
Why does familiarity make people less willing to delegate?
Because expertise gives you a detailed model of everything that could go wrong, and no way to verify it has not without looking. In an unfamiliar domain you have no such model, so you fall back on the checks that actually generalise — does it work, is it fast enough, is it safe. Which are, uncomfortably, the checks that matter. The expert's discomfort is real information about risk, and also partly an inability to trust anything they have not personally read.
Why is it worth noticing?
- It is an unusually honest account of a bias most experienced engineers have and few state.
- The conclusion is not "stop reading code" but "know which checks you are actually relying on".
- His follow-up makes it explicit: hand-writing code, hand-writing prompts and reading AI's output are all *means*; the goal is software quality, and there are other means to it.