The prompt was never the hard part
Mar 2026 / 3 min
When I rebuilt my site I didn't type one prompt and accept whatever came back. I'd browse reference sites I liked, screenshot the specific things that worked, the layout, the motion, the type, and feed those screenshots to one model to get a careful description back. Then I'd take that description into a separate session to generate the actual build prompts. A pipeline of chained steps, each one narrowing the gap between a reference I could see and a prompt a model could act on.
I built a first prototype a while ago and recently went back to push it toward the tier of site that gets featured rather than one that merely passes as a portfolio, running the same screenshot-to-description-to-prompt loop with a much higher bar for what I'd accept.
The honest part is where the work actually lives. Models don't do what you ask a good fraction of the time, and what separates a decent result from a broken one is clicking through what came out and catching the things that are subtly wrong, far more than the wording of the prompt. That's the one opinion I'll defend here: writing the prompt is the cheap step, and verifying the output is the job. The stack was React with TypeScript, Vite and Framer Motion, but the stack is the least interesting thing about how the site got made.