Two hours, three kinds of prompts

I built my site in two hours. Somewhere around prompt ten, I noticed I'd been writing three completely different things and calling all of them "a prompt."

That distinction mattered more than anything else in the build.

The first kind: single prompts. Disposable, one task, use once. "Make the logo bigger." "Rename this label." Fine for anything low-stakes. I learned the failure mode the hard way: I described a whole section in one lazy sentence and got back something that technically matched my words and missed everything I hadn't bothered to say. A single prompt for a structural decision is a coin flip, not a shortcut.

The second kind: master prompts. A spec, not a request. Scope, constraints, exact data shapes, hard rules, a defined deliverable, written once, in full. The prompt that built my design system named the exact type scale, the spacing values, which components existed and what they were for. Real test for whether something needs this treatment: could you hand it to someone else, or to yourself tomorrow, and get the same result back. If the answer's no because too much was implied, that's a master prompt problem wearing a single prompt's clothes.

The third: workflow. Master prompts run in a deliberate order, each one building on what the last left behind. Mine went: content and copy first, decided before I touched a prompt. Then scaffold, structure only, no visual decisions yet. Then the design system, locked once. Then the site, section by section, using both. No single master prompt, however well written, gets you a finished site alone. Build a piece before what it depends on exists, and you get a worse result with identical instructions. Order isn't a nice-to-have. It's doing real work.

Single, master, workflow, in the order they actually matter.

Here's where I got it wrong, because skipping this would make the whole post a lie. The hero section's photo got marked ready early on and never actually made it into the content file. It rendered with no image for several prompts straight. The AI didn't catch it. Neither did I, not until I went back and checked the real file against what I'd claimed was done. A separate prompt made a folder assumption that didn't match where things had actually landed. Same story.

Neither gap got caught by trusting a status update. Both got caught by re-reading the real thing. Master prompts don't audit themselves. Neither does a workflow. That check is a step you do on purpose, or it doesn't happen.

The working version, short enough to remember: single for anything disposable. Master for anything repeatable. Workflow for the order those specs run in when one depends on the last.

Made it this far? Here's the reward: every prompt from this build, in the exact order I ran them, free. Grab them here: the prompts library