When my book needed a website, I didn't open a design tool. I wrote a brief, handed it to an AI coding agent, and watched it build the thing.

A year ago that would have been a party trick. Now it's just a Tuesday. But "the AI built it" hides the most interesting part of the story, which is everything that happened after.

The brief

The site is for The Algorithm Ate My Homework, a satirical book narrated by a fictional, deeply overconfident AI called PROSE-7. So the landing page had to be the joke. I wanted it to look like a slick AI product launch, the kind of page Claude, ChatGPT or Gemini would have, but for an AI that is confidently wrong about everything.

The centrepiece: a HAL 9000-style lens. Glowing in PROSE-7's yellow rather than the classic menacing red, tracking your mouse like an eye, with a crack across it in a shape that echoed the book's branding.

I wrote all of that into a markdown brief (outcome, audience, constraints, references, definition of done) and handed it over.

What it nailed

Honestly? A lot.

Structure and layout. The page came back clean, responsive and properly built. Sections in a sensible order, decent spacing, no layout collapsing on a phone.

The mechanics. The mouse-tracking eye worked first time. That's the sort of fiddly interaction that would have taken me an evening of trial and error. It took the agent minutes.

Speed of iteration. Every change was a sentence away. "Make the lens bigger." "Slow the tracking down." Done.

If your fear is "AI can't build production-quality web pages," I'm sorry to report that it can.

What I still had to fix

Here's where it gets interesting. Almost everything I changed fell into one of three buckets.

1. The joke wasn't landing. The first version looked like a real AI product page. That was the brief, and it was the problem. It was so convincingly earnest that the parody disappeared. Good satire needs one deliberate wobble, a detail slightly off, so the visitor gets that small wait a minute moment. The AI played it perfectly straight because I'd asked it to. Knowing where to put the wobble was a comedy decision, not a code one.

2. Generic defaults crept back in. Wherever my brief was silent, the model filled the gap with the statistical average. Placeholder-feeling copy. A gradient I didn't ask for. Spacing that was fine rather than intentional. None of it was wrong. All of it was forgettable. Each fix was small; together they're the difference between a page and a brand.

3. It couldn't judge its own output. The agent could tell me the page met the acceptance criteria. It couldn't tell me whether the lens felt menacing or just a bit silly, or whether a first-time visitor would understand what they were looking at. That required looking at it like a stranger would. That's the oldest design skill there is.

The pattern

Look at that list again. The AI handled build, mechanics and speed. I handled tone, intent and judgement.

That's not a coincidence, and it's not a temporary gap waiting for the next model to close it. The things I fixed weren't technical failures. They were decisions that depend on knowing why the page exists and who it's for. You can write some of that into a brief, and you should. But the final call ("does this feel right?") still needs someone with taste in the loop.

What this means for designers

If your value is "I can build a landing page", that value has been repriced, sharply.

If your value is "I know what this landing page needs to make someone feel, and I can tell when it's not there yet", you just got a very fast, very cheap production team.

The shift I'd encourage: stop measuring yourself by what you can produce and start measuring yourself by what you can direct and correct. Write better briefs. Get faster at spotting the generic. Get comfortable saying "technically fine, but not right" and explaining why.

(See the finished result, cracked yellow eye and all, at prose7.com.)

The AI built my landing page. I made it mine. Both jobs were real. Only one of them is getting cheaper.