For six years I wrote every line of every project myself. No autocomplete bigger than a Tab, no chat windows open during deep work, no asking a model to scaffold a route. I was good at it — and I was slowly getting outpaced by people who were not.
The first week of working with Claude as an actual peer was uncomfortable. Not because the suggestions were bad — they were not — but because my mental model was wrong. I was treating it like a fancy autocomplete and getting brittle, generic output. The unlock was switching from "complete this line" to "here is what I'm trying to build, here are the constraints, propose two approaches and tell me which you'd pick."
The benefits showed up in places I did not expect. Boilerplate became a non-issue, sure — but the bigger win was scope reach. I shipped a marketplace, an admin dashboard, and a billing pipeline in the time it used to take me to ship one of them. Not because I wrote less; because I spent more time on the parts where my judgment actually mattered.
There is a quieter benefit too: fewer dumb mistakes. When I describe an approach in plain English to the model and it pushes back — "you'll have a race condition here" — that is free code review on every diff. I have shipped fewer regressions in the past year than in any year before.
What stayed the same: my taste, my architectural calls, my relationship with clients. The model has no opinion on whether your data model is honest. It has no opinion on whether your client is being lied to about timeline. Those calls are still mine, and I have gotten more aggressive about making them, because the model handles the parts I used to dread.
If you are on the fence, the move is not "try AI in your editor for an afternoon." The move is to commit one full project to a new workflow and accept that you will be slower for two weeks. After that, you do not go back.