I've been building software for twenty years and the last twelve months are the largest single shift I've seen. Not because AI is replacing the work, but because it's changed the shape of every hour I spend at a keyboard. Here is the honest version of what's different.
Code generation is real, and it's not what people think. I don't ask Claude to "build me a feature." I ask it to scaffold a known shape — a Postgres migration, a React form, an API route — and then I edit. The leverage is in the seventy percent that comes back roughly right, not in the thirty percent that comes back wrong. The skill is knowing which part is which.
The model has changed what "boilerplate" means. A year ago, boilerplate was the slow part of a feature. Today, boilerplate is the part I delegate while I'm still thinking about the design. The thinking has moved upstream — into shape, into edge cases, into the harder questions about what to build at all.
Reviews have changed too. AI is genuinely good at telling me what I missed in a 200-line diff. I run my own pull requests through Claude before I open them, asking specifically for security issues, edge cases, and inconsistent naming. It catches things. Not always; not the deep things; but enough of the surface things that I ship cleaner code.
Documentation is the thing AI changed most for me. I used to write docs after the fact, badly, and only when someone asked. Now I generate a first draft from the code, edit it down, and ship it with the feature. Docs went from a chore to a thing that just happens. That alone is worth the subscription.
What hasn't changed: the design work. The thinking. The decisions about what's worth building, who it's for, and what the failure mode looks like. AI is excellent at lowering the cost of execution and unchanged at making strategic calls. If anything, the strategic part of the work matters more now, because everything else is faster.
The trap I'd warn anyone about: treating AI as a replacement for understanding. The teams shipping the worst code I've seen this year are the ones using AI to write code in stacks they don't know. The model will happily generate an answer; it won't tell you whether the answer is right for your codebase, your stack, or your scale. That judgement is still yours.
If I had one piece of advice for a developer building today: use the model heavily, audit it heavily, and never let it ship anything you couldn't have written yourself. The same advice applies to whatever it changes into next year.