An AI agent did most of the making of this site. It wrote the code, sized and cropped the images, built every page, and pushed the whole thing live. If you're here for the part where the machine replaces the designer, though, this isn't it.
Here's how it actually went. The words live as plain text files I can open and edit anywhere. I write and rewrite the copy; the agent turns it into pages, keeps the design consistent, handles the fiddly production, and deploys to Github for versioning. We work in a back-and-forth. I say what I want, it makes a version, I react, it adjusts, and round again. Nothing goes live without me saying so. The agent can build and prepare all day, but publishing is a decision, and that one stays with me.
Most of my time went on the things a model is bad at: what's true, what's worth saying, what to cut, and how a sentence sounds read aloud. That last one matters more than it looks. The site runs on a set of voice rules, written down, that the agent has to follow. Short sentences. Plain words. No jargon, no em dashes. Left to its defaults it writes in the fluent, faintly corporate register everyone now recognises, and the rules exist to keep it out. Getting to something that sounds like me, rather than a competent stranger, was hours of editing, not one clever prompt.
The same goes for the facts. Every number on this site traces to a source I can point to: a project deck, a published figure, my own records. If a claim couldn't be checked it didn't go on the page, however well it read. More than one good-sounding line got cut for exactly that reason. An AI will hand you a confident number as happily as a true one, and telling the two apart isn't the model's job. It's mine.
Even the small stuff was a conversation. A header image got recropped three times because the first two cut off a detail that mattered. A headline was rewritten until it stopped sounding like a headline. None of that is the model being clever. It's the model being fast, and me deciding when fast was actually right.
So here's the position. The model is the easy part. What stays human, what's safe to hand over, what good actually looks like, what to do the first time the machine is confidently wrong: those are judgements, and they're the whole game. Get them right and the tool barely matters. Get them wrong and no tool will save you.
This isn't an argument against working this way. It's fast, it's genuinely capable, and I'd recommend it. It's also meant I have a website that's not costing much beyond the cost of the domain. But the speed is a reason to be more careful, not less. When a system can generate a plausible anything in seconds, the harder skill is knowing which plausible thing is the right one, and being willing to throw the rest away.
The agent built the site. The decisions were mine. That's the whole point. If you're working out how to use these tools yourself, that's the question I'd start with. Not "what can it do?", but "what can I do to make this feel more like me?".