I Built This Entire Website With Claude Code. Here's What Happened.

There is a section of the internet that has been very excited about vibe coding for the past year. The concept is simple: describe software to an AI, have it write the code, iterate until something working emerges, ship it. I read about it with my usual mix of genuine curiosity and mild suspicion. Then I decided to stop reading and find out.

The thing I decided to build was this website. The one you are currently reading. So you are looking at the direct output of an experiment I ran on myself. I want to be upfront about my starting point: I had no real technical background. I understood HTML in the same approximate way I understand how planes stay in the air. Not mechanically. I knew what React was because people kept mentioning it. That was about it.

I used Claude Code because I had been impressed by Claude's reasoning and wanted to stay in the same ecosystem I was already using for everything else. My brief to myself was simple: a personal site, clean design, no WordPress, real code, something I could own and extend later. I started a session. Four hours later, I had a working Next.js site with a navbar, a hero section, a contact form, and three pages. I was genuinely floored. Not because it was beautiful, because it really was not yet, but because there was running code on my laptop, in a framework I had never touched, that I could open in a browser and see. Something about that felt improbable in a way that actually doing it had made real.

What followed over the next several weeks was a masterclass in iteration. I would describe a change, the hero section feels too cramped, I want more breathing room, and Claude would make it. Sometimes exactly right. Sometimes it introduced a bug somewhere else. Sometimes it solved the problem I described but not the problem I actually had. That last category was the most instructive: the moments that forced me to stop and figure out what I actually wanted, not just what I had said.

I made the same architectural mistake twice, which is an achievement of sorts. Because the first version worked, I kept building on top of it instead of stepping back to ask whether the underlying structure was right for what I actually needed. AI code moves very fast. Bad architectural decisions made quickly are still bad architectural decisions. The speed does not fix that.

The things Claude could not do were as instructive as the things it could. It could not tell me what I actually wanted the site to look like. Taste is mine to supply. It could not debug a visual problem it could not see. When something broke in a way that only showed up in the real browser, I had to take a screenshot and describe it. It could not write the words on these pages. Every sentence was either written or heavily edited by me. The code is Claude's. The voice is mine.

"The cost of changing your mind in a software project is usually real. With AI pair programming, it drops dramatically. That is liberating and slightly dangerous."

The site took six weeks of intermittent sessions. A developer could have built it in a week to a spec. But I needed to discover the spec by building, which is a different kind of process and probably the only kind that was going to work for me. I now understand this site in a way I would not if someone else had built it. I can change it, fix it, extend it. Sort of. Mostly.