I wanted a place to write that felt like mine - not another Medium page or a LinkedIn article that disappears in the feed after a day. I also didn't have weeks to spend on it. So this site is mostly the result of describing what I wanted to coding agents, reviewing what came back, and saying "no, not like that" a few times.

Here's roughly how it came together.


Day one: a terminal with a few moods

The first version went up on September 8th, in a single day. The brief was basically "an engineering notebook that feels like an old terminal." Out of that came:

  • Three color themes - phosphor green, amber, and paper - that you can flip between from the header.
  • A five-second boot sequence on the homepage with typed diagnostics. You can skip it with Enter or Escape, and it remembers if you'd rather never see it again 😅.
  • Vim motions, because of course. j / k to move, gg and G for top and bottom, h / l for back and forward, / to search, ? for the cheat sheet, and :reboot if you want the boot sequence again.
  • A bit of "code dust" that trails the cursor, and a GitHub panel that pulls in my public activity, releases, and contribution graph at deploy time.
The homepage artwork in the green, amber, and paper themes
Same artwork, three themes

The walk-on: fal.ai and about fifty cents

The artwork on the homepage is a little Vault Boy-style version of me at a terminal. I wanted it to move - walk up, type a few keys, give a thumbs-up.

That ended up being one request to Kling 2.1 Pro through fal.ai. The setup was a generated opening keyframe (me mid-stride, hands off the keyboard), the existing artwork as the ending frame, and a written motion prompt. A piece of it:

The bearded man in the W baseball cap takes two short, natural steps to the right toward the stationary retro terminal... He stops at the keyboard, taps a few keys, then turns his head toward the viewer, smiles, and raises his free hand in a clear thumbs-up.

One paid render, about $0.49 at the advertised rate, and it was good enough on the first try. There's a little facial drift mid-walk if you go frame by frame - I can live with it. The amber and paper versions weren't separate renders either - they were recolored locally with FFmpeg from the approved green clip.

Ten frames sampled from the Kling render, showing the walk, typing, and thumbs-up
The render, two frames per second

I didn't expect the agent to write the brief like a spec. Inputs, render targets, a negative prompt, and a checklist for reviewing the result (watch the hands, check the cap lettering, make sure he actually walks instead of sliding). That's the same thing I'd want from a person doing the work.

Markdown, with guardrails

A week later the posts moved out of a hardcoded JavaScript file and into plain Markdown, rendered by a small renderer package I maintain. The build is strict on purpose - it refuses a post that repeats its title as a heading, links to remote images, or references a file outside its own folder. Drafts never make it into the published site.

That strictness is what makes "a few prompts" workable. There are over a hundred tests now, and an agent can't quietly break the site without one of them complaining.

The contact card

Cursor's background agents built the digital contact card - a vCard download, a themed QR code page for conferences, and a CRT-filtered headshot - across a handful of small pull requests, each with screenshots attached for review.

Getting ready for ICC, with Claude Code

For the ICC debrief, Claude Code:

  • Took about 40 phone photos down to 11, resized them, and stripped the GPS data before anything got committed.
  • Added the "Book a call" button to the header and fixed the divider styling.
  • Rewrote j / k on article pages to actually scroll - and then chased down why holding the key acted like a toggle. Best guess: one of my many browser extensions was eating the key-up event 🤓.
  • Moved the whole site onto my own Cloudflare account and this domain, with a redirect so every old link still lands on the right page.

What I'd tell someone trying this

Describe the outcome, not the steps. Put the boring guardrails in early - tests, a strict build, a check before anything deploys - because they're what let you trust the next prompt. And review everything. The agents did the typing. I still had to decide what "good" looked like.

Cheers 🥃

// END OF NOTE
Download .md