Continuing a style isn't the same as starting one

I’m rebuilding my website again. A friend I caught up with last week laughed when I mentioned it, because he remembers several previous versions. Fair.

I’ve built the last couple of versions with Claude too. What’s different this time is that I’m also using it to design the site from the start, and I keep looking at what comes back and thinking: this looks like Claude.

It’s obvious what’s wrong with it. The spacing is fine, the type is fine, the layout is sensible, but for a personal website it has zero personality. Maybe that’s what you get for asking an agent to communicate visually who you are.

What makes this strange is that in my client work, the same tool doesn’t feel like that at all.

On one project, most of the screens are fairly straightforward. Once the style and structure were established, I noticed that when someone asked me to mock up a new page, I could hand it to Claude and get most of the way there. Maybe four rounds back and forth, and the result is as good as if I had designed it myself.

On another, I’ve been building a large design system, and Claude has built something like 90% of it. I’ve barely made any components by hand. That only works because a lot of decisions were made at the start, and every new component has to follow them: always bind to tokens, always use the same structure, always use the same property names. I genuinely think the result is a better system than I would have built manually, and it would have taken me about four times as long.

A third team I work with is building a tool that generates imagery from a brand’s guidelines. You give it a few reference images, it works out the style, and then you can ask for whatever you want. I tried something deliberately odd, a car crash in a city, and the style held. It still looked like the brand.

Three different contexts, and the same thing is true in all of them. The tool is very good at continuing something.

My website has nothing to continue. There’s no established style yet, no structure, no rules. I’m asking it to start, and when it starts, it reaches for the average of everything it has seen. Which is a very competent average. It’s just not a point of view.

Continuing is a different job from starting. Continuing means following decisions. Starting means making them, and making them in a way that the next hundred decisions can follow.

A better model may produce a better first draft, but it doesn’t remove that distinction. The faster continuation becomes, the more consequential the starting point becomes.

I’d occasionally had the feeling that I was cheating when I let AI build so much of the system. I don’t anymore. The prompts aren’t where the knowledge lives. The knowledge is in knowing which decisions need to be made up front, which rules the system needs, and what to look for when something comes back slightly wrong. Without that, the same tools produce a very different result.

When production gets this fast, the first decisions carry further. A weak rule can spread through the whole system just as quickly as a good one. The leverage has moved to the beginning: the style, the structure, the rules, the first few decisions that make everything after them easier to get right.

So I know what the problem with my website is. I’ve been asking it to start something I haven’t started yet.

That’s a slightly uncomfortable thing to admit, because I have some version of this conversation with founders all the time. I argue that foundations and frameworks are what let us move quickly without remaking the same decisions later. The value can be hard to see when there are screens to ship and the payoff arrives later.

Then I approached my own website as if none of that applied. I gave Claude a blank page and expected it to find a point of view I hadn’t defined myself.