Speed, mostly. Not quality.
The biggest shift is velocity. Scaffolding a new page, writing a seed script, generating boilerplate — things that used to take an hour take ten minutes. That compounds quickly across a project.
What hasn't changed: taste, judgment, and the ability to know when something is wrong. AI will generate confident, plausible code that does the wrong thing. Catching that still requires knowing what the right thing looks like.
AI drafts. I decide.
I use AI for a first pass — component structure, CSS layout, copy drafts. Then I read everything line by line and fix what's wrong. The ratio depends on the task: for boilerplate it's maybe 80% usable, for nuanced design decisions it's closer to 40%.
Never accept generated CSS without checking it in the browser at multiple breakpoints. AI has no idea what your design actually looks like — it's pattern-matching from training data, not seeing your screen.
It can't taste. You have to.
AI-generated design suggestions are average by construction — they're trained on what exists, not on what's excellent. For UI, I use it to generate starting points and throw out most of what it produces.
The one place it's genuinely useful in design: copy. Getting a first draft of microcopy, button labels, error messages, and section headlines is fast with AI. I still rewrite most of it, but starting from something beats staring at a blank text field.
Context collapse is the real problem.
The hardest part of working with AI on a real project is that it doesn't hold context the way a collaborator does. It doesn't remember that you decided to use a specific pattern three hours ago, or that the client hates a certain word.
You become the context manager — keeping track of decisions, constraints, and preferences that the AI resets on every session. That's cognitive work that doesn't show up in the velocity numbers.
Tool, not replacement.
I'm faster. The work is still mine — the decisions, the fixes, the judgment calls. AI handles the parts of the job that are mechanical. The parts that require taste, experience, and understanding of the client's actual problem remain entirely human.
That's probably the right division. I don't see it changing soon.
