For most of my career, the constraint was building. You had an idea, and then you spent weeks or months making it real. The gap between the idea and the working thing was where almost all the effort went.
That gap has mostly closed. With AI, I can build a prototype of almost any idea in a day or two. That sounds like pure upside, and it mostly is, but it quietly moves the hard part somewhere else. When you can build anything, the real question becomes what to build. Finding direction. Figuring out what people actually need, and designing something that's genuinely useful. That turns out to be much harder than the code ever was.
I've spent the last few weeks living inside that question, building something I've been calling reef.
I lost count of how many times I started it over. The first version drifted into an agent-first static site generator, then into a local AI wiki builder, which wasn't what I wanted, so I scrapped it. The next was skills-based, built on top of existing publishing platforms. The one after that I deliberately limited to WordPress only: a small agent runtime that takes markdown in and outputs to WordPress. Far more constrained than the versions before it, and that constraint was where things started to come into focus.
That's the strange thing AI does to the work. Starting over costs almost nothing now, so you can afford to throw away three versions of an idea just to understand it. Each version I threw away taught me something about what the thing actually wanted to be.
What kept pulling me back was a small set of primitives. Markdown files. Git. AI. RSS. The URL. None of them new. All of them durable. I kept asking myself what the primitives of publishing on the web really are today, and that short list was the answer I kept landing on.
Underneath all of it sat one conviction I couldn't shake: the site should be just files, and you should own it.
That's the version I believe in now, and I'm calling it basepage. You point an agent at a folder and ask it to build and design a site from your local files. You serve it locally, and basepage injects edit buttons and full revision history straight into the page. Then you publish it anywhere you like. In three prompts the other day, I asked Claude to spin up a new site, import the posts from my old blog at simonbc.com, copy the design, and preview it locally. It worked because the site is just files, plus a few skills that teach the agent the commands and the structure.
It isn't as slick as something like Lovable. But it's something better than slick: the sites aren't spit out by an AI and trapped in a black box. They're real sites you can open, edit, fully own, and carry anywhere.
Building is cheap now. That's exactly why the things worth making are the ones you can still hold onto. Plain files, on your own machine, that you can take with you and publish wherever you want. That's the part AI doesn't change, and it's the part I most want to protect.