Taste is still doing the work
I wrote something on andresmax.com today about taste being the new competitive moat. The argument came from watching six different Lovable prototypes that technically worked but felt like nobody made them. The features were there. The judgment wasn’t.
That’s the AM take. The builder take is more specific.
Last week I did a full pass on this site. Accessibility audit. Bolder typography. Scroll reveals. Parallax on the hero. Rebuilt the contact page. None of those are big features. Each one is a small decision. But stacked together, they’re the difference between a site that looks like something and a site that looks like it was generated.
That’s what taste actually is. Not a single call. Hundreds of them.
I use AI heavily across everything I build, tini.bio, gratu, lst.so, this site. Claude Code handles the execution. But every decision about what to execute is still mine. Should this animate in or just appear? Does this section need a divider or does the whitespace do it? Is this copy honest or does it sound like a landing page template?
You can’t prompt your way to those answers. You can’t describe taste in a system prompt and expect it to stick.
What you can do is pull real data into the conversation. MCPs and APIs mean I’m not building blind. In a working session I can pull Google Analytics, Search Console data, real user behavior, and feed that directly into the decisions I’m making with Claude. The build loop isn’t “guess, ship, check later.” It’s “pull the data, decide, build, ship.” All in the same session. That changes the quality of every decision because you’re optimizing against reality, not assumptions.
The uncomfortable truth for solo builders right now is that AI removed the last excuse. You used to be able to say “I’m a solo dev, I don’t have a designer.” That’s gone. Anyone can get to a working UI in an afternoon. The question is whether it feels like someone actually made it.
Three seconds. That’s how fast users trained on great software form an opinion. Not “this looks unprofessional.” Just a feeling that something’s off. Or that it’s right.
I think about that every time I’m reviewing something Claude built. Is there a decision buried in here that I’m just accepting because it’s close enough? Close enough is how you end up with six prototypes that all kind of look like each other.
The things I build at mx.works are small. tini.bio, gratu, lst.so. None of them are going to out-feature some well-funded competitor. But they can feel considered. They can feel like a person built them on purpose. And now, with the right tools plugged in, those decisions are grounded in data, not vibes.
That’s the moat I’m actually trying to build.