AI design needs more control
Claude-only web design is wild when you already have a Figma design in front of you. There is very little mystery about what you want. You can see the layout, the spacing, the way the pieces belong together. Getting the code to arrive at that same place can still take a surprising amount of back and forth.And I still typed three paragraphs describing it
I described it on X as three stretches: the first 60% flies, the next 30% takes ages of saying 'no, not like that,' and the final 10% feels like god mode. Those numbers describe the feeling, not a benchmark. The strange part is how quickly the experience swings from impressive to frustrating and back again.
Once the structure is right, changes can feel almost instant. Before that, even a small request can turn into another round of explaining the layout. That middle stretch is where I keep getting stuck.
Close is a difficult place to work from
A rough first version is easy to appreciate. There was nothing, and now there is a page. A version that is almost right asks for a different kind of attention. You start seeing all the relationships that are slightly off: a panel that's too wide, a heading wrapping too early, a mobile layout that technically fits but reads in the wrong order.
These are normal design decisions. In a visual tool, I can often point at the thing and change it. In a conversation, I have to describe the thing, explain the desired result, and then check whether the change affected something else. The request may be short while the verification takes much longer.
That's what I meant when I compared vibe coding to bowling. Some throws land beautifully. Others miss in a way I didn't expect. When I know exactly what I want to adjust, I'd like a little less suspense.
Getting connected is only the beginning
In one of my drafts I called this 'MCP hell.' I'd been working through a v0-to-Claude Code workflow, and it felt like most of my time went into matching the layout and structure from Figma. Once that barrier was crossed, iteration opened up. Getting across it was the expensive part.
That experience makes me cautious about judging a workflow by whether two tools can talk to each other. Access to the design is useful. Understanding which relationships need to survive in the browser is another problem. A connection can carry information while leaving me with plenty of explaining to do.
What I want is the ability to make a precise correction and keep the surrounding decisions intact. Move this. Preserve that. Show me what changed at the smaller size. The more directly I can express those intentions, the more useful the speed becomes.
I want both speeds
My preference in that post was to use Framer for structure and AI for things like shaders, animation, and exploring ideas. That was a response to the friction I was feeling. I wouldn't turn it into a permanent rule about which tool belongs in every designer's workflow.
The balance can change. A visual tool might make generation easier to steer. An agent might give me a better way to select and adjust a specific part of the page. Either would help with the same frustration.Less sure about this than I was in July
I want the speed of asking for a whole possibility and the precision of adjusting one small detail. The exciting moment is when those two things happen in the same workflow, and I can spend more time deciding what feels right than explaining why the last attempt still doesn't.
Acknowledgements
Expanded from two posts of mine on X and a draft written alongside them. The Figma-to-code workflow described here is my own; nothing in it is a claim about how anyone else works.