Last week I handed Claude a prompt and a GitHub token and told it to write. Not one post. Not a draft to clean up. I mean I gave it the keys and watched what happened.
The result was 85 posts in a single session: 75 essays arguing that buildings are graph problems and capital is the actual design tool, plus 10 interactive canvas applets — wall sections, room configurators, egress diagrams, dimension strings — all rendering in the browser with sliders, all generated from scratch.
I want to be straight about what that means and what it doesn’t.
What Claude actually did
The 75 thesis posts are a structured argument. Not random takes — a deliberate sequence organized into four claims: that architecture is downstream of capital, that buildings reduce to a graph of rooms, that construction is a set of microservices, and that design tools should behave like software. Each post is a short essay with a title engineered to land like a slap. Claude wrote all 75, organized them into sections, built the index page, and pushed every file to GitHub.
The 10 applets are different. Those required Claude to write working canvas code — actual JavaScript drawing geometry, responding to slider input, rendering immediately on load with no blank state. Wall Section I draws a layered assembly and lets you adjust stud depth. Portal V starts with one exit and a failing egress path so you understand the problem before you touch anything. These aren’t screenshots of UI. They run.
I reviewed all of it. Some titles got edited. A few applets needed a second pass on the drawing logic. But the authorship is Claude’s, and I’m not pretending otherwise.
Why I did it
This blog has always been about the idea that buildings are underspecified software problems — that if you modeled a floor plan the way a compiler models a program, you’d catch errors before you poured concrete. That’s what TestFit is built on. I’ve been writing variations of that argument since 2018.
What I wanted to know is whether an AI trained on everything I’ve written, plus all of architecture and real estate and software, could synthesize that argument faster and more completely than I could alone. The answer is yes, with caveats. Claude can generate the shape of the argument at scale. It can’t replace the judgment about which shapes matter.
The 75 theses are a map. The interactive posts — the ones I’ve been writing by hand since 2024 — are the territory. Both are on this blog now, clearly labeled.