Bolt has been amazing at design, but it's a token hog - one instruction can drain a million tokens on a fairly small codebase. But if I were prototyping a loose idea with no idea where to start on UI/UX, the most efficient way would be via Bolt.
Bolt will do app-wide changes. Like recently I'm working on a character generator. I want it to include personality traits after gender but before clothing. It adds the screen, increments the steps, and changes the final screen to reflect this new data type.
Cursor does the engineering very well. I think it's become common recently to have a shared markdown file that different agents can access.
Gemini Pro has the huge context window but tends to make edits you didn't ask for, like simplifying your instructions and removing some of the important ones. Claude sonnet 4 is just nice for most tasks. o3 is the most precise by far, but low creativity. Visually GPT-4.1 seems to be the best. So if you have a Figma design, you can just screenshot and paste it.
Do have a cheat sheet ready for the AI. Tell it how to open modals, where to get the fonts, colors, icons.
I know some people are able to make it fully automated, but that is something I don't trust it to do. It would be an interesting exercise to see how well we can write a PRD then, and it becomes an exercise in commanding genies.