The old version of this site was a pile of Markdown files glued together with a templating helper. It worked until I wanted project devlogs, tags, and search. Each one meant more glue, so in April I started over.
What I wanted
- Write in
.norgand publish the file as it is. - Ship one Go binary: no database, no Node at runtime.
- Give each project a folder, with its devlogs next to it.
What runs the site today
| Layer | Choice |
|---|---|
| HTTP | Go net/http and its stdlib router |
| Templates | templ, compiled to Go |
| Content | a Neorg parser I wrote, about 1,000 lines |
| Code blocks | Chroma for highlighting |
| Styles | Tailwind v3 standalone binary plus handwritten CSS |
| Logs | zerolog |
| Deploy | a distroless Docker image on Railway |
The app parses every .norg file once at startup and keeps the results in memory. A request reads from a slice and renders a template. I have no cache to invalidate, because nothing changes until the next deploy.
All of it comes to about 2,400 lines of Go, with roughly 2,200 more lines of tests.
How it came together
April: an empty shell
My first commit fixed a go:embed build that broke when the CSS file did not exist yet. A humble start.
May: the parser and the projects
The Neorg parser landed on May 2. A week later it handled tables, task states, definition lists, spoilers, and inline math. On May 19 projects became folders: index.norg describes the project, and every other file in the folder turns into a devlog at /projects/{slug}/{post}.
July: hardening and a visual pass
I ran govulncheck, and it flagged a reachable vulnerability in the crypto/tls package of Go 1.26.4. I moved to 1.26.5 and kept pulling the thread:
- The parser validates every link it renders, since its HTML goes out through
templ.Raw. - Search answers
400to any query over 256 bytes. - The build pins the Tailwind download by SHA-256 and pins GitHub Actions to commits.
The next day I self-hosted Source Serif 4 and Source Sans 3 and moved the palette to warm paper and espresso. Two weeks after that, I split the parser out of a single 1,015-line file into its own package of five files.
October: a front door
Until this week, / redirected to /blog. Now it opens on a menu with three doors: Blog, Projects, and About. Each door shows a photo from my coffee routine: writing over a cup, a pour-over at home, a takeaway latte from the café. Hover one and its frame grows from 96 to 192 pixels, a trick I took from Fernando Borretti's site. Click it and a CSS view transition slides the site name into the header of the next page, with no JavaScript.
What I would keep
Parsing at startup was the best call. A broken .norg file stops the app at boot, where I see it, instead of failing in front of a reader.
Writing the parser myself comes second. It took longer than pulling in a Markdown library, but I can read every line that turns my notes into HTML. That made the security pass a short afternoon instead of an audit of someone else's renderer.
What is still rough
Search scores each word on its own. A search for bento grid returns anything that mentions bento or grid, and rebuild does not match rebuilt. Phrase matching and stemming come next.
The devlogs on the Blog Engine project page have the details.