Michael Lane

← All work · 2026

This Site

An exhibit about the site you’re on. Very meta.

The page you’re reading is the exhibit. I gave Claude Code the brief and the taste; it wrote every line. Then I rejected twenty-nine of its mockups until the design was right. Every “no” is below. Yes, this exhibit is inside the thing it’s about. If you click “Case study” from here, you will eventually find this sentence again.

An early dark design with warm gold accents.A graphite and lime palette mockup.A poster-style mockup with an enormous headline.Cards with hard black offset shadows and a tilted blue banner.A color-blocked mockup with black and blue bands.A Zork map drawn beside the headline.Zork’s opening text set faintly beside the headline.A large compass rose beside the headline.A grid of 32 small boxes, two of them blue.A working text-adventure prompt beside the headline.The finished home page, with soft blue waves.The original source code as a tilted, softly focused photograph, in the dark version.Unopened Worlds’ stylesheet rendered the same way.ZorkAI’s GameEngine.cs as a tilted, fading panel of real, selectable text.

“Give me something distinct from Unopened Worlds.”

29
mockups rejected
0
lines of code written by me
0
design decisions made by the AI
2
subagents doing QA in parallel
What I did
The brief, the audience, the taste, every pick and every rejection, the reference images that unstuck it, and the copy.
What AI did
All of the implementation, the screenshots, the recorded game sessions, and a QA sweep that caught bugs I’d have shipped.

The brief

A portfolio for someone hiring an engineering leader. It had to be honest that AI writes most of my code now, and it had to prove I still own the judgment. So the site itself had to be built that way.

What “directing” actually looked like

The first version was a résumé in disguise, and I said so. The second was dark, gold and literary, a near-twin of Unopened Worlds. Then came palettes, five typefaces, a poster style with hard shadows and a tilted banner, three more styles, three variants of the one that was “better but still boring”, and five “flourishes” for the empty white space, including a compass, a Zork map and a live text-adventure prompt.

Most of those were good work. They were also wrong, and saying so quickly was the job.

What unstuck it

Words weren’t enough. Two things were: a link to another company’s website, for the soft waves in the background, and three stock photos of code on a screen, which became the source-code panels at the bottom of every case study. When the AI had a picture of what I meant, it nailed it on the first try.

What I learned

The model is fast and tireless, and it will happily polish the wrong idea for hours. It needs someone who knows what good looks like, says no early, and shows rather than tells. That’s not a new skill. It’s the same one I use with engineering teams.