I've spent eighteen months trying AI tools on real work, freelance and in-house. Most didn't stick. A few did, and they've changed what a normal workday looks like for me.
For code
Cursor reads the whole repository. Copilot, at least the way I was using it, only ever seemed to read the file I had open. Cursor knows I set up dependency injection in src/lib/di before it suggests mocking it in a test, and that alone saves a surprising amount of arguing.
Most prompt-to-UI tools overpromise, but v0 by Vercel is useful once you know what it's for. It won't replace a designer. It replaces the thirty minutes of scaffolding a Flexbox layout and typing Tailwind classes. I describe the wireframe, it gives me the first 80%, and the taste is still my job.
For design and research
Figma AI, still in beta, is best at chores: resizing frames with sensible padding, swapping lorem ipsum for realistic data, keeping variables in line.
Search results are full of SEO filler, so for anything technical I start with Perplexity Pro. When I need to understand a new payment provider's webhooks, I get a sourced summary first and skim the docs second.
Claude Projects fixed my most repetitive habit, which was pasting the same background into every chat. I keep one project per client, with the schema, our conventions and that strange decision from last quarter already in it.
What I dropped
Autonomous agents that promise to build the whole app while you sleep kept disappointing me. They ignored edge cases, didn't weigh technical debt, and wrote code as if nothing had existed before them. Generated code looks perfect right up until it meets a case nobody trained it on, and then it makes something up with complete confidence.
So the high-stakes logic and the data models stay with me. The tools take the boilerplate and the wiring, which leaves the hard, ambiguous problems. The job got harder and better at the same time.