I Rebuilt My Website with AI. Here’s What I Actually Did.

I’ve had a personal website for years. It sat there. WordPress install, Kadence theme, content I hadn’t touched since whenever I last felt motivated enough to log in. It wasn’t broken. It just wasn’t doing anything.

So when I started working with Claude, I figured my own site was the perfect guinea pig. Low stakes, no client breathing down my neck, and I’d learn something about what AI tools actually do when you aim them at a real project.

What followed was four rounds of work, a few design crises, and one misspelled URL that had been live for two years. Here’s how it went.

Round 1: Content, No Design

I started by writing a site strategy document with Claude’s help. Seven pages, a color palette pulled from the Florida Keys, 16 photos uploaded to the media library. Everything was done through the WordPress REST API using Cowork. No browser, no admin dashboard. Just API calls pushing content into WordPress like someone filling out forms in the dark.

It worked. Technically. The content was solid. But the site looked like a term paper with a nice color scheme. Turns out, telling an AI what words to put where without mentioning how anything should look gets you exactly that: words, in places, looking like nothing.

Round 2: The Honest Assessment

I had Claude pull up the site in a browser and go through every page like a design reviewer. The diagnosis came back fast: the homepage was carrying the entire site on its back. Hero image, bold headline, circular headshot, stats bar. Actual composition. Every other page was just a nav bar followed immediately by body text, like a Word document that accidentally became a website.

The problem was clear. My strategy document was thorough about content and completely silent on design. I’d built a house and forgotten to hire an architect.

Round 3: The Rebuild (I Mostly Watched)

This was the heavy lift, and I want to be honest about my contribution: I pointed at things and said “that doesn’t look right.”

This round moved to Claude Code, which connected to my server via SSH and built a child theme from scratch. Design system with custom color properties, three font families, a spacing scale. I understood maybe 40% of what was happening. The other 60% was Claude fighting CSS selectors while I nodded from the sideline.

The single biggest fix was simple: every interior page got a full-width dark teal banner with the title and a gold accent line. One component. Transformed the entire site.

Then the CSS wars started. The Kadence theme has a wrapper element that silently overrides your background color. The WordPress Customizer, where custom CSS lives, turned out to be inaccessible through the API. The Customizer tab crashed at one point. Good thing Claude knew what it was doing, because I was just sitting there watching two systems argue with each other.

I contributed one useful thing: I flagged that the nav text was hard to see. (I’m colorblind. “Hard to see” for me means it might as well not exist.) Changed all nav links to white. Done.

The footer said “WordPress Theme by Kadence WP.” Claude tried the normal fix. Kadence didn’t cooperate. So Claude hid the original text with CSS and replaced it through a pseudo-element. I don’t know what that means, but the footer says what I want now, so I’m not asking questions.

Round 4: Words and Voice

Back to Cowork for the content pass. This round I was actually useful, because it was about writing, not code.

I’d built a voice rules document that defines how I write. Short sentences. Active voice. Contractions always. A long list of banned phrases: “harness,” “leverage,” “game-changer,” and the fatal “Not X. This is Y.” pattern that AI generates like it’s getting paid per use.

Claude built a scanner that checked every page against those rules and found violations. The About page said “I’m not a tech person.” That’s the banned negation pattern. Gone. I caught “storytelling person” and said make it “storyteller.” The kind of stuff that matters when your name is on the site.

The best catch was the blog post. The original title was “Harnessing AI Tools to Unleash Artistic Creativity: A Journey with Midjourny.” Midjourney was misspelled in the URL. The title used “harnessing,” which is literally on my banned word list. The post read like an AI content mill wrote it, because one basically did. Rewrote the whole thing.

What I Took Away

The process looked like working with a contractor. I described what I wanted. Claude built it. I reviewed and gave notes. Some rounds went fast. The CSS wars did not.

My honest contribution was taste and direction. I knew what I wanted the site to feel like. I knew when something looked wrong or sounded wrong. I didn’t know how to fix any of it. Claude did the fixing. I did the pointing.

The split between tools mattered. Claude Code handled the server work, the SSH connection, the child theme, the real programming. Cowork handled the browser automation, the content, and the design review. Neither alone would have gotten the site where it ended up. That’s the thing nobody mentions about working with AI: knowing which tool to use for which job is half the skill.

The site’s at chadnewman.com. It works. It looks like mine. That’s more than I could say a month ago.