I Let an AI Agent Rebuild My Personal Site in a Weekend

My old personal site was a monument to good intentions. A static site generator I'd half-forgotten how to use, a theme I'd forked and then "improved" until nothing upgraded cleanly, and a build that only worked on one specific laptop — which I no longer own.

So one rainy weekend in a guesthouse, I tried something I'd been putting off: I gave an AI coding agent the repo, a short brief, and told it to rebuild the whole thing.

The brief

I kept it deliberately plain. No 40-page spec, just what I actually wanted:

• Move everything to a modern framework with markdown posts
• Keep every existing URL working
• Fast, accessible, readable on a phone in bright sunlight
• Dark mode, an RSS feed, and a sitemap
• No tracking scripts, no cookie banner

That last point matters to me. A personal site should not need a legal disclaimer to tell you what it's doing.

What went surprisingly well

The boring migration work. Converting 60-odd posts from one front-matter format to another is exactly the kind of task I'd put off for months. The agent wrote a little script, ran it, spotted three posts with broken dates, and fixed them. Total time: about four minutes.

URL preservation. I expected this to be painful. Instead it read the old config, worked out the permalink pattern, and generated redirects for the handful of posts that didn't fit. I spot-checked thirty links and every one resolved.

Accessibility basics. Alt text placeholders, proper heading order, focus states, sensible contrast. Not perfect, but a much better starting point than my old theme, which used light grey text on a slightly lighter grey background because I once thought that looked "minimal."

Where it went off the rails

Over-engineering. At one point it proposed a plugin system for my blog. My blog has one author. I don't need a plugin system; I need a folder of markdown files.

Confident wrongness. It told me a configuration option existed that... did not exist. When the build failed, it apologized and fixed it, but this is the part people underestimate: you still need to read what it did. An agent that is wrong 5% of the time and confident 100% of the time needs a human who checks.

Design taste. Its first pass looked like every SaaS landing page from 2023. Gradient blobs, a "Get started" button on a blog. I had to be quite specific: "This is a personal blog. It should feel like a well-designed notebook, not a product."

What I'd do differently

1. Start with constraints, not features. "No JavaScript unless necessary" saved me more cleanup than any feature request.
2. Work in small, reviewable steps. When I asked for everything at once, I got a 3,000-line diff. When I asked for one thing at a time, I understood every change.
3. Keep a running notes file. I had the agent update a DECISIONS.md as it went. Future me will be grateful.

The verdict

The site is faster, cleaner, and — crucially — I understand how it works again. I'd estimate the agent saved me two full weekends of tedious work and cost me maybe three hours of review and course-correction.

That's the honest pitch for AI coding tools right now. Not "it builds things for you," but "it does the tedious 80% while you steer the interesting 20%." For a personal project, that trade is absolutely worth it.

The site you're reading now is the third iteration of that experiment. More on how it evolved in future posts.