AI design needs more control
Building a site with only Claude is wild when you already have the Figma file open. You know exactly what you want. The layout, the spacing, how it all fits together, it's right there. And getting the code to match it still takes a lot of back and forth.And I still typed three paragraphs describing it
On X I put it as three stretches: the first 60% flies, the next 30% is ages of saying 'no, not like that,' and the last 10% feels like god mode. I made those numbers up, they're a feeling. What gets me is how fast it swings from amazing to infuriating and back.
Once the structure is right, changes are almost instant. Before that, every small request turns into explaining the layout again. I always get stuck in the middle part.
Close is a difficult place to work from
A rough first version is easy to love. There was nothing, and now there's a page. Almost right is harder. Now you see everything that's slightly off: a panel that's too wide, a heading that wraps too early, a mobile layout that fits but reads in the wrong order.
None of that is unusual design work. In Figma I'd just grab the thing and move it. In a chat I have to describe it, describe what I want instead, and then check nothing else moved. Asking takes a sentence. Checking takes ten minutes.
That's why I said vibe coding feels like bowling. Some throws are perfect. Some miss in ways I couldn't have predicted. When I know exactly what I want to change, I don't want the suspense.
Getting connected is only the beginning
In one draft I called this 'MCP hell.' I was going from v0 to Claude Code, and most of my time went into getting the layout and structure to match Figma. After that, iterating was fast. Getting there was what cost me.
So I don't get too excited when two tools can finally talk to each other. Sure, the agent can read the design now. Knowing which parts of it have to survive in the browser is a different problem, and I'm still the one explaining that.
What I want is to fix one thing and leave the rest alone. Move this. Keep that. Show me what happened on mobile. The speed only helps if I can say that directly.
I want both speeds
In that post I said I'd use Framer for structure and AI for shaders, animation, and trying ideas. That was me being annoyed that week. It isn't a rule for anyone else's workflow.
It could flip. Visual tools might get better at steering generation. Agents might let me point at part of the page and change just that. Either one fixes my problem.Less sure about this than I was in July
I want to ask for a whole page and then nudge one detail, in the same place. Then I could spend my time deciding what feels right, instead of explaining why the last try still doesn't.
Acknowledgements
From two of my posts on X and a draft I wrote around the same time. It's my own Figma-to-code workflow. I'm not saying anyone else works this way.