Claude is surprisingly bad with UI. Nearly every other is better at it. GPT-4o is proper multimodal, while Claude is a mix. I'm not sure about GPT 4.1, but it seems to gave better results than 4o.
You should be able to get good enough results with just a screenshot.
It stumbles on newer things like Next.js though. I'm always amused by the people who think it's taking jobs. It's always going to be mid; it has to be trained on someone's code.
It makes some good guesses, and often gets the color or font weights close. One trick is to have it go over your design system and summarize everything into ui.md and such.
You should be able to get good enough results with just a screenshot.
It stumbles on newer things like Next.js though. I'm always amused by the people who think it's taking jobs. It's always going to be mid; it has to be trained on someone's code.
It makes some good guesses, and often gets the color or font weights close. One trick is to have it go over your design system and summarize everything into ui.md and such.