Instead of an intro
I have spent the last year putting AI into real client work rather than side projects. Not the demo kind — brand systems, product interfaces, and the unglamorous middle where a design has to survive contact with a back end.
What follows is not a list of tools. Tools change every few months. The parts worth writing down are the decisions: what I hand over, what I refuse to hand over, and how I notice when the work has started to drift.
What it is genuinely good at
Volume without judgement. Twenty layout variations, naming lists, alt text for a hundred images, a first pass at copy I will rewrite anyway. Anything where my value is in choosing, not in producing.
It is also very good at the parts of engineering I know conceptually but type slowly: wiring a component to an API, writing the boring state machine, translating a design token file between formats.
Where it quietly costs me
Taste is the first casualty. The output is always competent and almost never specific. If I accept the second suggestion instead of the twelfth, the work ends up looking like everyone else's.
The second cost is harder to see: I stop noticing small wrongness. Spacing that is nearly right, a curve that is nearly the brand, copy that is nearly the voice. Individually invisible, collectively the reason something feels generic.
The rules I keep
I decide the direction before I open anything that generates. The brief, the constraint, the one sentence about how it should feel — those are mine. Everything downstream can be assisted.
I never ship a first pass. Not because first passes are bad, but because reviewing my own choice is the part that makes it a choice at all.
And I keep one project a year entirely by hand. It is the only reliable way I have found to tell whether the craft is still mine.
Useful links from this article