I was wrong about Claude Design

Last week I wrote Claude Design doesn't need me. The idea was that you could tell Claude Code to use Claude Design, and it would design every page and then build it, with no one at the controls.

I don't think that was true, and I'm taking it back.

What I thought was happening

At the time, it looked like it worked. Claude Code said it was building a design system, drawing pages on a canvas and implementing them. The results were better than what Claude Code usually makes on its own, so I believed it.

Then a couple of pages came out that I wasn't happy with. I opened Claude Design myself to fix them, and the difference was obvious. Working in Claude Design directly gave me much better designs than anything I'd seen come back through Claude Code.

That made me look again at the earlier results. They were better than Claude Code's defaults, but they weren't what Claude Design produces. I don't think Claude Code was really using Claude Design the way I described. It was doing something closer to its usual work, with more care about design, and I gave the credit to the wrong tool.

Why designing first is better

The quality of each page is only part of it. The bigger difference is that the whole app comes out consistent.

When Claude Code designs as it builds, each page gets designed on its own, inside whatever feature it's working on. Each one can look fine and still not match the others.

When I mock up the UI in Claude Design first, I'm looking at all the pages together. The sign-in flow, the settings page and the admin get designed at the same time and share the same choices. Then Claude Code builds from that, instead of making up the design one piece at a time.

What I do now

Everything gets designed before it gets built:

Mock up the UI in Claude Design, myself.
Hand the design to Claude Code.
Let Claude Code build it.

It's one more step, and I'm back at the controls for it. The results are clearly worth it.

The lesson

Last week's lesson was "it's whether it occurs to me to ask." That's still sometimes true. This week's lesson is that a model saying it did something is not proof that it did. I liked the results, so I didn't check how they were made. I should have compared against the real thing before telling anyone it worked.