Building Better Websites with AI

There are two stories going around about AI and websites, and both are wrong.

The first is that you can describe a site to a model and get something good. You cannot, at least not yet, and not for anything you would want your name on as a designer. The second is that AI has not really changed design work, that it is a toy for prototypes and the real work is untouched. That one is wrong too. The truth sits in an awkward middle that is more interesting than either headline, and it has quietly changed how I build sites for a living.

Here is the short version. We should be doing better design than ever right now, focusing on fewer pages, and pouring most of the craft into a small foundation that everything else gets built from. Get that foundation right and a whole team can extend it for months without a designer in the room. Get it wrong, or skip it, and you end up with exactly the slop everyone is worried about.

The reason AI sites look like AI sites is that most of them were never designed. Someone prompted a model, accepted roughly what it gave back, and shipped it. The model has no opinion about your company. Left alone, it reaches for the same soft gradient, the same rounded cards, the same diagram of inputs flowing into an engine and out into outputs that you have seen on every other platform this year. The first concepts a model hands you look like AI-generated concepts, because that is what they are. To get past that, a person with taste has to actually design the thing. That means making opinionated decisions, which is exactly where AI models are lacking for design.

So the majority of work moves to the front. Instead of designing everything at a normal level, you design a few key pages and assets at a level you would be genuinely proud of, and you treat those pages as the source of truth for everything that follows. The homepage especially. Get the type, the color, the spacing, the way you show the product, and the one or two ownable brand elements exactly right on that one page, and you have set a bar that the rest of the site can be measured against.

On the pages that matter, you end up doing more design than the old approach ever asked for. I have spent two weeks on a single site that, in the old world, I would have given three days. Your effort just has to show up in different ways than before. People can feel when effort has been applied to something, and that effort is just as important when AI is involved. A design that has ten or twenty percent more polish and intention than the thing next to it, and that gap is easy to feel and almost impossible to describe.

A foundation is more than a pretty homepage. It is a working style guide, built in code, that defines the real rules: the type styles, the spacing, the colors, the components, and the system for making product imagery. The imagery part matters more than people expect, because plain AI sites skip it. They drop in a stock graphic and move on. A real foundation says, here is how we show the product, here is the workflow illustration style, here is what a chart looks like, and it builds those as actual components rather than flat images.

Building it in code, rather than inside a visual platform, is a deliberate choice, and it is the choice that makes the rest of this possible. When the styles live in code, tied to real tokens, a person can prompt a new page into existence and the model has something true to build from. It reads the homepage, sees the styles you already committed to, and keeps going in the same language. A founder I work with put it plainly: doing it custom makes it so much easier to vibe-code everything that comes after. The homepage and all that comes with it also becomes a set of rules.

Once that baseline is set, you can bring in more people, and they do not have to be designers. This is the part that still surprises clients.

One client's marketing lead had never opened a terminal in his life. We showed him how to install the tools, and within a week he was pushing his own changes to the site, generating new pages off the homepage, in the same style, because the rules were already there for the model to follow. Another founder rebuilt an entire section of his product site himself and started, half-joking, calling himself a developer. A third had twenty-five reports he wanted as feature pages. Because the imagery was built as code off the baseline, he could spin those pages up quickly, and at one point generated a full second-language version of the site on his own, with all the imagery translating along with it, because none of it was locked inside a flat file.

For a lot of what a growing company needs, this is enough. It is a great way to test whether you even like a new message before you commit to it. You can put a variation in front of customers, see what lands, and learn the hard way, as one founder did, that nobody actually understood the term he had built his whole homepage around. That lesson is worth far more when it costs an afternoon of prompting than when it costs a designer a week.

For the routine stuff, the designer's job ends when the foundation is done. I mean that as the goal, not a concession. I genuinely want a client to be able to say they do not need me anymore. We set the bar once, and you do not pay us to help you adhere to the bar. Editing the hero copy, adding another tool page that mostly reuses existing sections, spinning up a quick variation to test, the team should just do that themselves.

The moment that changes is when something starts to matter. When a page shows real commitment, when it carries weight in the funnel, when you want it to feel finished rather than functional, that is when a designer should come back and put intention into it. On one project we drew the line cleanly: the long tail of tool pages reused components and only had to clear a low bar, enough to signal breadth, while the two or three pages that actually drove the business got real, custom design effort. Knowing which pages are which is most of the skill.

The danger is forgetting that the bar still applies. The same system that lets a team ship a good page also lets them ship an ugly one, and an ugly page does not just sit there harmlessly. It actively works against the brand you spent real money to establish. I have watched sites drift after handoff, getting worse with every well-meaning edit, until I wanted to quietly take the company off our own customer list. The capability cuts both ways, and the only protection is a team that keeps caring about design even when generating something careless is one prompt away.

None of this happens as one big handoff. It runs as a loop, and the loop goes roughly like this.

Design the foundation by hand. Extend it with AI to build out the rest. Test what works, ideally in front of real customers. Refine the parts that matter by hand. Then extend again from the improved base. Each pass makes the system more capable, because you are feeding it more context and teaching it how to handle the edge cases you only discover by running into them.

Order matters here, and it is easy to get wrong. In a perfect world you lock the homepage as the base style guide before you ever ask the model to generate eighty pages. I have been messy about that myself, generated a pile of pages first and then gone back to make everything reference the same styles, and it works, but it is more painful than doing it in sequence. The clean version is foundation first, then breadth.

And you will hit walls. I once asked a model to build out a large batch of pages overnight, had no idea if it would work, and woke up to find it had quietly broken something that was fine the day before. The fix was to revert the homepage to yesterday and try again. No matter how refined a build is, you launch and then notice something a week later. It is the curse of the demo. You cannot account for every edge case in advance, and you should stop expecting to.

If there is one thing I would tell anyone serious about this, it is that the quality you want does not come from a clever prompt. It comes from volume.

I've tracked how many prompts a site takes. One homepage I was genuinely happy with took around twelve hundred prompts over multiple days of back and forth. Another was closer to two thousand. That is the opposite of set-and-forget while you go out to walk your dog. You sit with the thing, describe what you want, wait, describe what you want to change, and repeat that loop until it is right. People hate the AI design tools partly because they show up expecting one shot, and one shot gives you the generic version every time.

If you care even a little, anything decent is at least ten or twenty prompts. The work I would actually put my name behind is hundreds, sometimes thousands. You can get a model to the polish you want. It just comes down to how long you are willing to keep going.

It would be easy to read all of this as a story about design getting cheaper and worse. I read it the opposite way.

The boring parts compress. The hours that used to go into wiring up pages now go into the decisions that were always the real work, the spacing that has a rhythm, the imagery that actually explains the product, the small fights that give a brand an opinion. A teammate of mine framed it well: the first eighty percent is covered now, so we get to spend the last twenty percent purely on polish, and the polish is where people feel the difference. Less time on plumbing, more time on the part that matters.

It also resets the bar. Generic sites, the kind a local service business needs, will get fully automated, and that is fine. Intentional brand work moves in the other direction. When everyone can look passable, passable stops being worth anything, and the only way to stand out is to be genuinely good on the pages that count. The future I would bet on is a designer building ten foundations so good that a model, and a whole team, can extend them for a year without the work ever feeling like slop, instead of one person cranking out fifty mediocre ones.

That is the whole approach. Do the foundation like it is the only thing that matters, because for a while it is. Let AI carry the long tail. And come back by hand the moment a page earns it. Repeat.