"Make it look professional" isn't an instruction — it's a hope. The assistant runs for a long time and self-corrects, which means the highest-leverage thing you can give it is a finish line it can check itself against:
"You're ready when: the page loads with zero console errors, every section has a scroll animation, the CTA is visible at every scroll position on a phone-sized screen, and everything works when I double-click the file."
Vague adjectives give you a first draft. Verifiable criteria give you a finished product—because the model keeps working until the checklist is true, instead of stopping when the output seems plausible.