What do you think about it?
My various experiences:
1) just let the agent come up with ui from scratch. 5/10. Gemini 3.1 pro is the best at 1 shotting ui.
1.5) use ai canvas design tools. I’m not that impressed. (1 w Gemini) and a good prompt still feels better.
2) start with something and work with agent to edit it. 3/10. Opus seems to take direction fairly well if you give it screenshots or point to a specific dom element.
3) skills that aim to embed a designers process, tricks, mindset. 2/10. Yes it cleans it up but it does not feel considered or smart.
4) iterate on static prototypes first with all of the above. Make your own example html pages. Make them interactive. Maybe a component library too. 7/10. Best outcomes. Still does not feel agentic or smart.
I apologize if I missed this from the ride up… were the optimization generalizable to the other tasks?
I found when doing similar improvement loops that the agents start to cheat and add improvements that only work for this very specific task.
I’m just wondering if you baked in rules prevent this? Or if you tested the improvements from one task on another task? Or maybe this is not relevant at all because you were trying to optimize for specific domains?
@fuho Prev version no. Not even me. But would be easy to do. The site periodically checks repos for big changes and auto updates.
.explainer is vey interesting. Thank you.
Will need a way to offset token cost if growth continues the way it has.