For a new screen, the usual way with a coding agent is to describe it and let it build. You find out what it looks like when it runs, and fix it from there.
The other way is to have the agent draw the screen first and only then build it. In Styx that's a design task: the agent produces ordinary HTML files in the task's branch, at desktop, tablet and phone widths (1280, 834 and 390 pixels). You look at them on a canvas, click what's wrong and say what to change.
What I like about it is that the design is cheap to change. Moving a button on a drawn screen is one message, and a build task that already exists gets told when the design changes. What I don't like is the extra step. For a small change it's slower than just building it.
So I use it when a screen is new or has several states, and skip it for tweaks.
Do you design first, or let the agent build and fix it afterwards? And where does your cut-off sit?
