Short films that make a design decision visible.
Three films, 16 to 24 seconds each, every frame drawn in code. With Claude Code I built the system that makes them, and I direct each one. In the film above, a single line sweeps a dashboard while Ghost, my design-system scanner, counts seven deviations on prepared demo data.
A decision does not show up in a screenshot.
It lives in a token name, a diff rule, a threshold. A still shows the result. A short film can show the decision being made: the options, the rule that picks one, what it changes.
In Decide once, four prompts each pick their own blue and three fail contrast. The one that passes is named for its job, so one edit restyles every screen.
Decide once, 24 seconds: one named token restyles four screens.
Every frame is a function of time.
Each film is one HTML page, and everything on it is computed from one number, the time t. Below is Decide once's own page, running live: drag to set t.
Decide once's source page. Any t draws one exact frame.
A script renders each frame in headless Chromium with subframe motion blur, at 60 frames a second, and the score is written in code from the same cue times.
My verdicts became the system's rules.
I set each film's brief and judge every cut. When a cut fails, the reason is written into the system, so the next film starts past it.
In films about my own work, every number that makes the argument is measured: contrast from the hex values, test counts from the repo, changed pixels from two drawings of one screen.
Which rule broke went public before its review came back. The review found no link from a rule to the element it changed, so the rework pins each rule in place.
First cut, as posted. Rework, each rule pinned in place.
Now a separate grader, a Claude instance that did not build the film, scores every cut before I pick one.
Live on LinkedIn and in the Ghost case study.
Two films ran as LinkedIn posts, on September 29 and October 1.
No screen recordings. Every frame of every film is drawn by its page.
The rework of Which rule broke now runs in the Ghost case study.
Seven deviations on prepared demo data, each named by rule.