A 4-screen RFP marketplace, designed from a timed brief that allowed AI-assisted execution on one condition: it couldn't look like AI had done it.
Everything below this line is the scrutiny layer: reasoning, wrong turns, and the system. Skip it if the scan told you what you needed.
The brief was short, but three lines in it did all the real work.
Nothing else in the brief needed real invention. Login, a detail page, a dashboard shell are well-understood patterns on their own. Where I spent design time tracked directly to these decisions, not to the screen count.
The system underneath the four screens, shown together.
What I'd tell you if you asked twice.
Three hours was never the real budget. Getting from a generic AI-assisted first pass to a system that holds together (consistent shell, real icons, real interaction states, plus a layout bug where two screens had silently doubled in width) took several more passes after the timer ran out.
That's the single biggest structural change in the project, and it came from rereading the brief, not from user research I don't actually have on a timed exercise.
It's triaging correctly what can stay rough and what can't. I spent the least effort on login, the most on the listing card, because that's where the brief's actual difficulty lived.
Organizations creating and managing RFPs are real in the brief's premise but out of scope for the four required deliverables. Worth knowing if you're wondering why there's no admin view here.