AI-generated websites - shit in, shit out

Why agencies can calmly surf the AI wave without fearing it, as long as they don’t do one thing: hand over the thinking.

I’ve been designing for over ten years, and I’m one of those designers who want their idea exactly on point. Not roughly, not 90 percent, exactly the way it looks in my head. That’s the standard I brought to my first AI website, which this time I wanted to build entirely with Claude.

It was a test and a learning experience at once.

I started with Opus 4.6 and I’m on 4.8 by now. The first 80% came quickly, then it got sticky.

Direct control instead of a waiting loop

What convinced me most was the direct control. Normally you hand a job to development, wait, and hope the briefing was “on point” so that the programmer implements your idea 1:1. You wait, get the result, go into revisions. Not here. I love simply experimenting & seeing the result immediately is worth gold to me.

Design freedom, no site-builder limits.

We’re not limited by rigid templates and boring site builders here, as a web designer you win back freedom & this is where I see the real benefit, at least for the designer. You can work with maximum freedom on a small budget.

Visual effects

Many topics that looked “simple” to me took ages, while complex ones were implemented with playful ease. One thing runs by itself, the other acts up. You could watch that contrast throughout the whole project.

No cookies, pure HTML

Another pleasant thing: I don’t have to set any cookies. Pure HTML, done. That’s not just relaxed to handle, it’s also relatively secure. Where there’s no CMS and no login, there’s simply not much to attack.

Productive on the side: multitasking and less mental load

A side effect I hadn’t expected: while Claude works on a task, short waiting moments keep coming up, and they were easy to use. Instead of staring at a progress bar, I could get other things done in parallel. The website build basically ran on the side while I carried on with other work.

Even more valuable than the time gained was the mental relief: when a complex task is taken over, I no longer have to hold all of it in my head. That saves energy, and makes multitasking noticeably easier, because I’m not juggling everything at once.

Where it gets stuck

Rebuilding 1:1 is like rolling dice

The web is full of beautiful effects, and in theory you can just “throw” one into the AI. In theory. Because it won’t rebuild them 1:1. Something different comes out every time, it was basically like rolling dice. In the end I simply took the best hit and refined it.

One thing gets fixed, another breaks.

Then came the individual pages, and this is where it got sticky. The same game over and over: something looks good, I send a prompt for a completely different corner, and the beautiful thing that was so lovingly prompted before is broken again. I must have requested the colour switch twenty thousand times. Every time the AI put something underneath it, it stopped working again. And the popovers in the menu? Didn’t work for ages. I had to touch things twice and three times over. For tasks that look really simple to me, the AI sometimes calculated forever.

The trick with the single chat

The first chat was an experience; I could pack an insane amount into it and was amazed at what Claude can do. But once three or four pages were in there, the struggling began and suddenly some things just weren’t processed any more. So I switched to several chats, all in the same project, each page type in its own area. The result: header and navigation suffered from version differences. Only when I packed everything into one single chat and orchestrated from there did it run smoothly. At handover, by the way, Claude itself recommended the ZIP. Uploaded, edited again, downloaded again, done.

Controllability and maintenance

As much as the direct control helps me during the build, I’m critical of it when handing over to companies. Once I give a site like this to a company: in the end, anyone who edits the site through the chat can break a whole lot.

It doesn’t work without HTML knowledge

And then there’s something that shouldn’t be talked away: you need to know HTML. Anyone who is technically completely “lost” won’t get anything polished out of even the best AI. This is not a magic machine that hands everyone a finished, technically flawless website.

My tips if you try it yourself

The most important thing first, and it’s really the whole core: put in as much information as possible. Not just “build me X”, but the background, the purpose, the audience, the story behind it. With AI it’s no different from the rest of my work. I first have to understand what the story is about before I can give it a shape. What you don’t put in won’t be built, it’s that simple.

And a few very practical things on top:

• Do as much as possible in one single chat. Spread it across several, and your header and navigation will keep tearing apart.
• Treat effects as exploration, not as a 1:1 order. Have it show you variants, take the best one and then polish.
• Save your state after every bigger step, ideally as a ZIP. A change in one place likes to shoot down the next.
• Accessibility, for example, is not something the AI builds in on its own, and if you don’t know it, you won’t ask for it.

The Holy Grail?

Many in the industry are treating AI like the holy grail right now, and that’s actually the good news: you don’t have to fear this wave, you can simply surf it. It will only roll over you in one single case: when you need the AI in order to become better. The knowledge it takes has to be there & that’s where many clients will fail when they try to build their website themselves.

Benjamin Trenkle, senior developer Joomla! and website specialist, puts it in a nutshell:

“AI is a tool, and you have to know how to use it. Not the solution to every problem.”

— Benjamin Trenkle, senior developer Joomla! & website specialist (wicked-software.de)

My verdict stands: shit in, shit out. What you don’t know, you can’t put in. What you don’t put in won’t be built. And that’s exactly why we’ll probably see a lot of half-finished websites. You get to the first 80 percent quickly. But the last 20, especially as a designer who wants it one hundred percent right, they pinch, and they’re exhausting to implement, partly because the AI doesn’t always bring enough capacity for them.

Time saved? For me, honestly, zero. Sure, I reached the 80 percent quickly. But because I want it one hundred percent, I ended up sitting at these eleven pages longer than if I’d simply “rocked” them in WordPress. I’m now into the third week of working on it. If you need something small and fast that looks really beautiful, you’re in the best hands here. But the final fifth that makes your project “complete”, you still end up doing that yourself. And what if the client wants all buttons “rounded”? You can’t promise it before you’ve tested it.

What made the difference in the end

One more honest observation: at the beginning it ran sluggishly, Claude got stuck and didn’t deliver what I needed. In the end, the decisive factor wasn’t a “smarter” model but more headroom. With a higher usage tier I stopped running into limits constantly, had continuous access to the strongest model and could push through big tasks in one go.

From then on, the tool reliably did what it was supposed to. The insight: it’s less about magic than about headroom, and that’s a question of the plan you choose. A pity this connection isn’t really communicated anywhere.

My look ahead

I see the real future of AI in web design somewhere else entirely: building templates for existing content management systems. That’s exactly where my next experiments are headed: I stay inside the CMS I learned from the ground up and have templates built there that are compatible with my actual theme. Because there, I’m fairly sure, AI can do really good work, and that way we don’t lose all the clients who are sitting on CMS systems.

Complete AI websites for companies, on the other hand, I currently view critically.

And one last honest word: the code that came out of it was, for long stretches, absolute rubbish. The only reason I could straighten it out at all is that I know the subject & how to work with Visual Studio Code. Without that knowledge, the whole thing would have failed.