Skip to content

Colophon v2026

How this site is made, and why, from the framework to the tests.

How this site is made, and why, from the framework to the tests.

The framework

The site is built with TanStack Start and TanStack Router, on React 19 and Vite. Most pages are prerendered to plain HTML when the site is built, so they load fast and work without JavaScript. The few that can't be, like Stats with its live numbers and the blog with its filters, are rendered on request and cached.

What I like about TanStack is that everything is typed end to end: routes, links, search parameters and the data each page loads. A typo in a link is a type error before it's ever a 404.

It's hosted on Vercel, which also runs the server functions and stores a little data in Vercel Blob.

The design

The components come from shadcn/ui, built on Base UI, and styled with Tailwind CSS v4. shadcn copies each component's code into the project instead of installing a library, so everything can be changed, and nothing changes behind my back. A few pieces come from community registries: the table of contents from toc-cn and the audio player from audiocn.

Text is set in Inter, with JetBrains Mono for code and the terminal-style headings, both self-hosted through Fontsource. Icons are Lucide, and I've drawn a few more in the same style.

There's a light and a dark theme, following your system unless you pick one, and everything that moves has a still version for people who turn animation off.

The writing

Posts are MDX: Markdown, with GitHub's extras like tables and footnotes, plus a few components of my own for videos, audio, files and star ratings. Code is highlighted by Shiki when the site is built, so no highlighter is sent to your browser.

Most of the 400-odd posts are an archive going back to 2004, first on WordPress, now moved over as MDX files in the same repository as the code. Posts over five years old say so at the top, and a mistyped or long-gone URL gets suggestions for the posts it might have meant.

What a post gets on top of its words, from tables of contents to structured data, is in What a blog post needs in 2026.

Generated, not made by hand

  • Share images for every page, drawn from its title with Satori and turned into PNGs with resvg.
  • An RSS feed, made with feed, with the full text of recent posts.
  • A sitemap and an llms.txt, from the same list of pages and posts the site itself uses.

Live data

  • Stats pulls public numbers from the GitHub and npm APIs.
  • Now playing comes from Spotify's API.
  • Health numbers come from my phone: the Health Data Export app posts them to the site, which keeps the latest day in Blob storage.
  • Comments are replies to a post's thread on Bluesky, read from its public API.

The small things

  • A command menu, built on cmdk: press ⌘ + K, or Ctrl + K.
  • Keyboard shortcuts with TanStack Hotkeys: press ? to see them all.
  • Numbers that count up with Number Flow, and a little confetti if you find the right key sequence.

Making sure it works

  • TypeScript everywhere, linted with oxlint and formatted with oxfmt.
  • Unit tests with Vitest, and browser tests against the real build with Playwright and axe for accessibility.
  • Every component and page state in Storybook, with network requests mocked by MSW, and every story tested in a real browser.
  • Dependencies kept up to date, but nothing installed until it's been out for a day, in case a release turns out to be hijacked.

How all that fits together, and why, is in What a website needs in 2026.

How it was built

I wrote this version of the site in November 2026 with Claude Code, an AI coding agent. It did most of the typing; I decided what to build, read every change, and tested it before it went anywhere. The other posts in this series are about what that was like.

Privacy

There are no cookies and no trackers. Vercel Web Analytics counts page views without cookies or identifying anyone, and that's all.