I’ve rebuilt this site more times than I want to admit, usually by opening wp-admin, poking at things for twenty minutes, and closing the tab feeling worse than when I started. This time was different. This time I did the whole rebuild — theme, color system, page templates, the works — working alongside Claude instead of alone. I want to write down what that was actually like, including the parts that didn’t go smoothly, because “AI redesigned my website” makes for a nice headline and a dishonest one if I stop there.
Why I did it this way
Pile of Tech has always been two things bolted together: a personal blog about my family and being terrible at video games, and a professional presence for my IT career. I’d let a fantasy world-building project (RIP Kimbrasen) sprawl across the site, my theme was years out of date, and I didn’t have a plan so much as a pile of good intentions. I wanted a second brain for this — something that could hold the whole plan in view while I made decisions in small pieces, and that could actually reach into WordPress and do the work instead of just telling me what to click.
That second part matters. This wasn’t “ask a chatbot for advice.” Claude was working directly in the browser, logged into my actual wp-admin, making REST API calls, running JavaScript against the block editor’s internal data store. That’s a very different level of access than I’d normally hand to anything, so I want to be upfront that this only worked because I stayed in the loop on anything destructive or hard to undo.
What actually went well
Having the whole thing structured as phases — discovery, strategy, platform decisions, visual design, then the build — kept me from the thing I always do, which is redesigning the homepage for the fifth time instead of finishing anything. When I said “let’s lock the palette” or “the professional links are LinkedIn and YouTube, nothing else,” that stuck. I didn’t have to re-argue settled decisions three sessions later.
The style guide work paid off too. Having exact type scale and color values (down to the specific WordPress color slugs, which auto-prefix themselves in ways I never would have caught on my own) meant the build phase was mostly execution, not more decision-making.
Where it got messy
My own caching plugin worked against me. WP Super Cache serves a different version of the site to anonymous visitors than what I see logged in. So a check that looked clean from my session could still be showing stale content to everyone else. More than once we thought something was broken when really we were just looking at the wrong cache layer.
The Rank Math setup wizard flat-out doesn’t work on my install. The “skip this step” button loops back to step one, forever. Small thing, but it ate real time before we gave up on the wizard entirely and just went and configured the settings pages directly.
wp-admin fought the tooling. The visual browser tools — screenshots, page reading — would hang indefinitely inside wp-admin specifically, even though the page had clearly finished loading. Public-facing pages were fine. It took a while to figure out the pattern was wp-admin’s own background polling, not a broken connection, and to switch to driving things through the REST API and JavaScript directly instead of banging my head against a stuck screenshot.
Removing one plugin quietly broke something unrelated. I pulled Jetpack because I didn’t need most of what it did anymore. Turned out my About page photo was being served through Jetpack’s own image CDN, so the second Jetpack was gone, so was my photo. Nothing about that was documented anywhere obvious — we only found it because the image broke and someone went looking.
None of these were catastrophic. All of them were the kind of thing that, on a solo late-night WordPress session, would have had me quietly rage-quitting and blaming myself for missing something obvious.
What I actually learned
The technical lessons are real — verify saves independently, know your caching layer, don’t trust a green checkmark just because it’s green. But the bigger thing I took from this is about the working relationship itself.
Having a structured plan turned “redesign my website” from an overwhelming blob into a sequence of small, checkable decisions. And having something that could hold the entire plan in its head — every earlier decision, every open item, every weird bug we’d already diagnosed once — meant I never had to re-explain myself or worry about contradicting a choice I’d made three weeks earlier and forgotten about. That’s not something I’ve gotten from any tool I’ve used solo.
But I also learned that “the AI did it” undersells what actually happened. What happened is I had a collaborator that was relentlessly methodical in a way I am not, at 11pm, on my own site. It caught its own mistakes because it went back and checked. It flagged things it wasn’t sure about instead of guessing. It asked before deleting anything. The value wasn’t that it knew more WordPress than I do — it’s that it was more disciplined about verifying its own work than I’ve ever managed to be on this site.
I also learned the value of using Projects and Project Instructions as a way to leverage multiple chats for the same purpose, but use less of my Claude usage. I had learned this from helping my son’s cub scout troop build a new website. The different this time was that I requested tackling this site in “phases”, used a new chat for each phase, and leveraged the Instructions to keep the history moving between chats. I also had each phase summarize a new Instruction writeup to maintain the essence of one larger chat. The result is a more streamlined approach to leveraging AI tools for the work, at a lower cost.
The site’s not fully done. There’s still content cleanup, SEO work, a handful of small bugs to chase down. But for the first time in years, I actually know what’s left, in what order, and why — instead of just having a vague sense that the site “needs work” and no idea where to start.