Building this site with an AI in the room

I work in life science, not software. That's worth saying up front, because it explains almost every decision behind this site — including the decision to build it at all.

The brief was small on purpose

I didn't want a CMS, a database, or an admin panel to fight with. I wanted to write in plain markdown, drop the file in a folder, and have it show up on the web looking the way I wanted it to look. Everything else was negotiable.

What we ended up with reflects that: posts and projects live as .md files with a small frontmatter block on top — title, date, type, tags. A build step reads them, turns them into structured content, and bundles the result straight into a Cloudflare Worker. No server to patch, no database to back up, nothing running that I have to think about between writes. The whole site is, at rest, just files in a folder.

Where the AI actually did the work

I used Claude Code for essentially all of the implementation — the build script that parses the markdown, the design system (a light "paper" aesthetic with a life-science-leaning palette), the routing, the validation step that stops a post with broken frontmatter from ever reaching the live site. I described what I wanted in plain language and reviewed what came back, the same way I'd review a colleague's draft: does this say what I meant, does it hold up, what's missing.

That's a genuinely different way of building something than what I was trained for. I know how to think about a problem clearly — that's most of my day job. I did not, before this, know how to turn "I want a sticky header and a right-hand margin for asides" into working code. The AI closed that gap, but it didn't remove me from the decisions. Every design choice — the palette, the typography, what a "type: journal" post should look like versus a "whitepaper" — was still mine to make. The AI just meant the distance between deciding something and seeing it on the page shrank from weeks to minutes.

What surprised me

The part I didn't expect was how much the process resembled writing itself. A first draft of a page was never right, and it didn't need to be — I could just say what was wrong with it and get a revised version back, the way you'd give notes on a manuscript. The iteration loop that normally makes software slow to build (write code, run it, notice it's wrong, go back and change the code) collapsed into something closer to editing.

I also learned to be specific. Vague requests got competent but generic results. The site got better exactly in proportion to how precisely I could say what I wanted — which, again, is not so different from writing.

What's still mine

None of this makes the AI the author of the site. The judgment — about what belongs here, what the palette should feel like, what's worth publishing at all — stays with me. The AI is closer to a very capable collaborator who never gets tired of a fourth revision. The site is still mine; I just had help building the thing it lives in.