The bar for AI design tools is in hell right now. Generation speed doesn't matter if you can't touch the DOM.
I’m building an AI UI canvas and testing it with early users, and I realized people are completely brainwashed by 10-second viral demos.
A tester sent me a screenshot of a dark-mode layout today saying it was a "game changer" because it generated instantly. Cool. But generating UI is not the hard part anymore.
The real problem: what happens when you actually try to edit it?
What happens when your client says "make that logo 10px smaller" or "change the gap on this flexbox"?
In almost every AI tool out there right now, the second you touch the layout, it shatters like Jenga. You're locked out of manual control. The output isn't a native, editable component-it’s a fragile prototype masquerading as a website.
If a tool doesn’t let me manually edit the grid, adjust padding, or swap divs without triggering a cascade of broken CSS, it's just a toy. I end up just rebuilding the whole thing from scratch in Figma or VS Code anyway.
Am I the only one who thinks we're optimizing for the wrong thing? At what point in using tools like claude , paper or stitch do you usually just give up, throw the AI output away, and code it yourself?