Extreme Motion Design

A motion design skill for coding agents (Claude, Claude Code, Codex and anything that reads SKILL.md). The agent writes the video as code with Remotion and React, with voice, licensed music and synced sound effects, and it is not allowed to deliver until the frames have been measured.

Leia em português

The film above is the 2.0 presentation, made with the skill. Every number on screen was measured and the imagery is public domain or CC0 (NASA, USGS, NPS and The Met), fetched by the skill with the licence checked per file.

What is new in 2.0

• Reference mode. Hand it a video and reference.py breaks it into measured structure (shots, camera moves, palette, where the visual mass sits, audio), writes a script skeleton with the same timings, and scores the new piece from 0 to 100 against the reference. Calibration: the piece against itself scores 100, using the score boundaries as cuts 86, a reel made from the same request without the reference 38, unrelated videos 23 and 24. It copies structure, rhythm and technique. It never copies frames, assets, text or sound.
• A design meter. design.py counts the signals of the frame a model produces with no direction: near black with one accent, a giant condensed face on the left margin, decorative mono, everything in caps, the same composition in every scene. It flagged my own 1.3 reel (4 signals) and the template example (3). The 2.0 film scores 0.
• A design reference. references/design.md covers composition, type, colour, texture, real imagery and launch film structure, each rule with its source, plus six named art directions to propose from.
• Real imagery with the licence checked. stock.py fetches photo and video from Wikimedia Commons (CC0 and public domain only), NASA, Openverse and The Met, and records the evidence read from the API for every file.
• Face and hand tracking in track.py, for graphics pinned to a body.

Why it exists

An agent writing video code cannot see what it renders. Left alone it ships text cut by the frame edge, captions over a face, the same card sliding into every scene, and audio that clips. I built the skill around that blind spot:

• The audio is the clock. One roteiro.json produces the voice phrase by phrase, the score and the named cues. No scene holds a hand typed frame number.
• The direction is a contract. Before any scene exists the piece declares its concept, its visual law, its motif and the layout and entrance of every scene. The lint rejects a project without it, and rejects one where every scene is laid out the same way.
• The text is measured, not eyeballed. A probe reads every text box from the DOM of each still and the background from the pixels. Cut off, overlapping, too small, low contrast or outside the safe area fails.
• Someone else reviews. The contact sheet goes to a reviewer with a clean context that never saw the code.
• The final file passes a numeric gate. Codec, loudness, true peak, flashes per second, one frame pops, loop seam, contrast against the stills.

What the gates caught while making this reel and a second one built on phone footage

These are real lines from the session, not examples I wrote for the README.

Every one of them was fixed before delivery, and four of them became fixes in the skill itself.

The same request without the skill

I ran the same request through the same model with the skill forbidden, plain Remotion only, and no paid voice. It is the honest comparison, so here it is with what I measured.

What I take from it: a good model already makes a good looking frame without help. The skill does not buy taste. It buys sound, a clock, measured text, true numbers and a second pair of eyes, and it costs time.

Gallery

Install

Claude Code (plugin)

/plugin marketplace add dev-isaacmello/extreme-motion-design
/plugin install extreme-motion-design@extreme-motion-design

Claude Code, Codex, Cursor and others (skills CLI)

npx skills add dev-isaacmello/extreme-motion-design

Claude (claude.ai and the desktop app)

git clone https://github.com/dev-isaacmello/extreme-motion-design
cd extreme-motion-design/skills
zip -r extreme-motion-design.zip extreme-motion-design/

Upload the zip in Customize > Skills. The zip must have the extreme-motion-design/ folder at the top.

Updating

• Plugin: claude plugin update extreme-motion-design@extreme-motion-design and restart.
• skills CLI: run npx skills add again.
• claude.ai: remove the old skill and upload the new zip.

Use

Ask in plain language. The skill instructions are written in Portuguese and the agent follows them in any language.

a 20 second 16:9 video with a calm pt-BR narration presenting our annual plan launch

a 9:16 reel from this phone video, with the caption following my hand

animate this logo in 4 seconds and loop it seamlessly

How a video is made

Techniques covered: keyframing, easing and springs, camera moves and parallax, kinetic typography, morphing, seamless loops, UI micro interactions, anticipation, overshoot, squash and stretch, smear, speed ramps, match cuts, shape and liquid transitions, mattes, frame by frame, cel animation, stop motion, rigging with IK, lip sync, 3D, compositing, point, plane, face and hand tracking, match moving and rotoscoping.

Requirements

• Node 18 or newer.
• Python 3.10 or newer. Scripts declare their dependencies inline (PEP 723) and run with uv run.
• ffmpeg on the PATH.

Remotion licence

The skill is MIT. Remotion, which renders the video, is source available and not open source. Read this before adopting the skill at work:

• Free for individuals, for companies with up to 3 employees, for non profits and for evaluation.
• From 4 people on, a Company License is required. At the time I checked (2026-10-13) a seat cost US$ 25 per month and a seat is needed for each person who writes Remotion code, including through a coding agent.
• A freelancer or agency that hands over only the final file counts its own headcount. If the client operates the project, the headcounts add up.
• The draft terms for Remotion 5.0 say contractors will count towards team size. 5.0 was not released when I checked.

Sources: LICENSE.md, licence FAQ, pricing. The skill warns the user when the company has more than 3 people.

What the skill runs and accesses

• npm install in the video project (Remotion, React, Three from the npm registry).
• uv run on the scripts in scripts/, which installs the Python dependencies each file declares.
• voice.py reads threads, RAM, disk and GPU (nvidia-smi). With --check-key it makes one read request to api.elevenlabs.io.
• score.py and tts.py send the script to ElevenLabs only when that voice was chosen, using the key in the project .env. score.py --dry prints the character count first and the paid run needs --yes.
• tts.py downloads the Kokoro voice model (about 330 MB) and align.py downloads Whisper small (about 460 MB) on first use.
• track.py --face downloads YuNet (230 KB) and --hand downloads the MediaPipe hand model (7.5 MB). roto.py --backend rembg downloads its segmentation model.
• music.py queries Incompetech and Openverse.
• stock.py queries Wikimedia Commons, the NASA image library, Openverse and The Met, and Pixabay only when you put a key in .env. It downloads the file you pick and strips the audio track of videos.
• reference.py only reads the video file you give it. It does not download from any platform.
• brand.py --url fetches the page you name and the stylesheets, fonts and images that page references, and takes two screenshots with a local headless Chrome. --repo only reads files.
• Nothing else is contacted.

Limits

• The agent reviews still frames and numbers. Rhythm and the feel of the motion still need a human eye before publishing.
• The layout probe only sees DOM text. Text drawn on canvas, in WebGL or inside an image is not measured, and text covered by another opaque layer is measured as if it were visible.
• Scene similarity compares 24x24 grey thumbnails. It was calibrated on six videos (medians from -0.08 to 0.71) and I had no repetitive video to calibrate the warning threshold against, so it is conservative.
• The fidelity score measures structure, not quality or content, and its tempo reading is unreliable when a voice sits over the music.
• The design meter thresholds come from practice and were calibrated on a handful of videos. Its verdict is a warning, not a gate.
• A free licence covers copyright only. Recognisable people, trademarks and recent artworks still need permission, and stock.py cannot detect them.
• brand.py picks the accent colour by heuristic. On a monochrome brand it can be wrong or empty, and it says so.
• Face tracking fails on profiles and on faces looking down. Hand tracking loses fast gestures.
• render.cjs --scene only knows what you tell it changed. After editing shared code, render everything.
• The rhythm warnings of qa.py and lint.py are heuristics tuned on a few videos.
• Only Kokoro is integrated as a local voice. The other models in assets/tts-models.json are recommendations by licence and memory.
• The skill instructions are in Portuguese.

Licence

MIT