After 1年半 with AI coding, a developer says deciding now outweighs writing

The author frames the shift in concrete terms: the first AI tool was a chat UI used for queries and small functions, while today a polished design doc lets whole implementations be handed over. That change moved the work from typing to specifying, and the author describes writing goals, background, ADRs, API specs, data structures, error cases and test design before any implementation request.

The losses are specific. Searching by hand once produced side discoveries and a habit of comparing sources; AI answers the question asked and little else. Hand-written tests once surfaced spec ambiguity and domain-specific nasty cases, and the author recalls rejecting AI-written tests that merely mirrored the implementation and would pass even if it were wrong.

The response is to rebuild thinking places rather than abandon AI: design docs first, verification pushed into formatters, linters, hooks and CI, and AI used as a sounding board whose proposals can be declined. The author notes that the ability to decline came from past experience, and warns that this reserve is finite — a risk concentrated among juniors who may never accumulate it.