Motion

Two things I want to do eventually both turn out to be about motion. I have articles that could become videos, which is one way to be seen at all, and I build iPhone apps like Harmony Pulse that would feel better with some life in them. These are my notes on four videos that cover both, plus two widgets for the ideas that only make sense when they move.

1 · A video is a function of time

Motion-graphics explainers can be made with nothing but Claude Code and JavaScript. The idea underneath is one sentence: Claude writes a function that takes a time t and returns one picture. Ask for 60 of them a second, play them in order, and that’s a video. A headless browser (Playwright) draws each frame, FFmpeg stitches them into an MP4, and the soundtrack can be code too.

Live · a video is a function of time

Turn on JavaScript to draw frames.

Everything on the canvas is computed from t alone, so frame 75 is always frame 75, and any frame can be redrawn or checked on its own. The contact sheet is the check the videos describe: eight frames side by side, which is how a model that can't watch video still looks at one.

Because the same t always gives the same picture, any frame can be checked on its own. Claude can’t watch a video, but it can look at frames. The loop that makes the results good is a contact sheet: pull a grid of frames, look for anything overlapping or cut off, fix the function, render again.

One creator ran one prompt at three effort levels: about 20 minutes on low, 30 on high, and eight hours on the highest. All three were solid. The lesson they draw is that bad results are usually a prompt problem. A good prompt states:

  • the length and kind of video, and the aspect ratio;
  • that it’s built in JavaScript and rendered with Playwright and FFmpeg;
  • the soundtrack, and whether visuals are timed to a narration you supply;
  • a storyboard first, with frames for each beat, before any real rendering;
  • and the contact-sheet loop before calling it done.

References help most after that: an image for the look, or another video for the composition with your own content poured in.

2 · A studio in a folder

One video was made, it says, entirely out of one folder, and walks through that folder from inside itself. Its structure comes from how Disney made cartoons around 1930: voices recorded first, every syllable charted on an exposure sheet, story sketches filmed against the soundtrack as a story reel, rough animation screened as pencil tests in a room the animators called the sweatbox. Only after all that did a drawing reach ink and paint.

StageMakesA person checks
1 · VoiceA brief, then a script read by a cloned voice, in chunksThe whole take, start to finish
2 · Words (the exposure sheet)Every word with its exact start and end timeNames the transcription got wrong
3 · Storyboard (the story reel)Plain-text beats, each keyed to a phrase, timed into a tableThe beats, before any animation
4 · Animation (pencil test and sweatbox)The scene as code, from a reusable kit; a contact sheetThe whole thing, with the voice playing
5 · Render (ink and paint)The video, a phone copy, thumbnails, chaptersThe title and thumbnail stay human

The folder itself is simple. A top-level file tells any AI what lives where. A second file holds the pipeline as one table. Each stage has its own folder with a short contract saying what it reads, what it makes, and who checks it, and each video gets a folder of its own.

The part worth stealing has nothing to do with video. Take any work you repeat, give each step a folder and a contract, and put a person’s check between steps. Fixing a beat in a text file is far cheaper than fixing a rendered scene, and the check is what stops one bad take from becoming a bad video.

3 · Motion in apps

This section comes from John, a staff engineer at Meta who has built animations for Instagram Reels and Threads. His argument is that AI-built apps all drift toward the same lifeless middle, because a model has no way to validate what feels good. Ask for “an animation” and you get something generic. Name the technique and you can steer it.

Live · same start, same end, different feel

Turn on JavaScript to compare curves.

The top track moves the dot; the bottom one shows where it was every 50 milliseconds, so bunched marks mean slow and spread marks mean fast. The red line is the end point; springs overshoot it and come back. The spring presets are the gentle, snappy and bouncy kinds the app-animation video demonstrates; the exact constants are mine.

His vocabulary, compressed:

TechniqueReach for it when
KeyframesA choreographed move from one state to another: position, size, color, opacity
Easing and springsAnything that should feel physical, like a button press or a card settling
GesturesSomething should follow a finger, then spring or fling when released
PhysicsA world that should react on its own: bounce, gravity, tilt
SVG paths and morphsMotion along a curve, or one shape becoming another
ShadersEffects that change pixels directly, like rounded masks or refraction
Layout transitionsCards reordering or a list becoming a grid; the quickest way to feel premium
ParticlesConfetti and sparkles on a success
Sprites and videoHand-drawn frames, or a pre-rendered clip
Rigging, RiveCharacters with expressions and many small moving parts

Two habits from his own puzzle app stuck with me. Every character animation starts and ends at a resting state, the way game characters are built, so any animation can follow any other. And he reviews every animation by eye, because the clipped corner or cut-off frame is exactly what an AI misses. He even had Codex drive the Rive editor through computer use to add new animations.

4 · Skills that check their own work

AI Labs tried a pile of the video skills people have published for Claude and kept eight that each do a different job. Every one is an install command and a slash command.

RepoWhat it does
huashu-motionLaunch films built from a site’s own colors, fonts and words, with the music written first and every cut on its beat
rossark-motionRules for animating drawings, mascots and logos, plus a notes file of what each character can and can’t do
take-oneOne line to a finished short film, planned by a strategist, a creative director and a storyboard artist
reanimateCopies an animation from a site, video or GIF by recording it frame by frame and rebuilding it until the two match
argoClicks through your app in a browser and turns the recording into a narrated demo
explainrooExplainers with a voice-over, every scene timed to the word it belongs to
video-directorDirector rules, with one checker for the plan and another for the finished video
earshotTurns sound into notes Claude can read, lined up against the frames

What they share matters more than any one of them. The good ones give Claude a sense it lacks: frames to look at, a recording to listen back to, a cue sheet for sound, a side-by-side against the original. Several hand the finished cut to a reviewer that never saw the build, and send it back if the score is low. And their rules are the same few: a hook in the first three seconds, text big enough to read on a phone, a pause before the biggest moment. One more lesson came free. Explainroo’s first cut looked like a template, and it only got good once it was rebuilt around a real Claude Code session.

The plan

  • Turn one article into a short video the folder way: script, timed words, storyboard, scene as code, contact sheet, render.
  • Keep a person’s check between every stage, starting with the storyboard.
  • Give Claude a way to check what it can’t sense (frames for the picture, a cue sheet for the sound), and have a reviewer that never saw the build score the cut.
  • Use real screens over template boxes wherever the product has them.
  • In the apps, name the motion I want (an ease-out on arrival, a snappy spring on press) instead of asking for “an animation.”
  • Give every animated thing a resting state, and review the result by eye.

Sources

Four videos, watched in October and November 2026. Everything above is my synthesis of their arguments, and the details (effort timings, the Disney history, the job history, what each repo does) are the narrators’ own. AI Labs says huashu-motion is an open clone of a motion tool Anthropic offers only on Team and Enterprise plans; I haven’t checked that. I couldn’t find reanimate’s repository from its name alone, so it has no link. The captions garble one animation tool’s name, so I’ve left it out. Three of the four end by pointing to a paid community or vault, and one has a sponsor read.