Case Study · Episode 18 · Brand Motion · Linear

EIGHT LOOPS FROM A STYLE FILE: a brand's motion, read from its CSS

I stopped describing brands to Claude in adjectives and started handing it their stylesheets. One extraction of Linear, one style file, eight seamless loops — every colour, curve and duration measured from the live site rather than chosen. The whole method is on GitHub, and it runs on any brand.

Output8 loops · 5s seamless · 1080×1350
Extracted421 properties · 101 keyframes
Built withClaude Code · Playwright · ffmpeg
Studiedlinear.app · unofficial
All eight, running at once — one clock drives every tile, so they stay in sync.
The Premise

Adjectives are lossy

Ask for something “confident, modern, premium” and you get a confident, modern, premium average of everything the model has ever seen.

Those four words could describe two hundred companies. But most brands publish their design system every day, in public, in a form that isn't lossy at all — it's called CSS. So I skipped the describing: headless browser at linear.app, read the live stylesheets and computed styles, and built eight looping motion graphics that were only allowed to use what came back.

The Method

Five steps, and one rule that does most of the work

  1. Read the real CSS, not a screenshotPlaywright and headless Chrome across three pages — custom properties, keyframes, and the computed styles of about four thousand elements. Frequency matters as much as presence.
  2. Write one style fileEvery measured value as a token, plus a human-readable version where each claim about the brand cites the code that produced it.
  3. Write one motion briefCanvas, loop length, easing law. All motion driven from a single clock, so frames are deterministic and recordable.
  4. One line per animationOne object, one change. No hex codes in the prompt — it refers to brand values by role, so the same sentence works on the next brand's style file.
  5. Capture and encodePlaywright seeks each frame and screenshots it; ffmpeg does the rest. Take forty is identical to take one.
421Properties read
101Keyframes parsed
8Loops built
9×Uses of the accent
The Finding

The number a mood board can't give you

I assumed Linear's indigo was a hero colour. The extraction said it appears nine times in four thousand elements.

It isn't a hero colour — it's a pointer, and it marks one thing at a time. Every loop in the set now carries exactly one accent element, and that single constraint is most of why they read as a family. The absences were just as useful: Linear publishes eighteen easing curves and not one of them overshoots. No spring, no bounce anywhere. You can't see an absence on a mood board.

What didn't work is worth saying too. The gradient piece was nearly a failure — restrained, technically correct, and indistinguishable from every AI gradient loop on the internet. It only became this brand's gradient when I put a still hairline grid in front of it. Extraction gives you a palette, not a composition.

Run it on your own brand

The extractor, the style file, the motion brief and all eight prompts — written so you can point them at a different site and get that brand's version instead.

Get it on GitHub →

Unofficial design study. Not affiliated with Linear. No logo, wordmark or trademarked imagery is used anywhere in the work — colours, type feel, shapes, radii, depth and easing only. Inter and JetBrains Mono substituted from Google Fonts for Linear’s licensed Inter Variable and Berkeley Mono.