Atelier

A generative animation toolkit for Claude Code — with the parameters actually verified, and the ability to watch what it built.

Claude can write a flow field. What it can't normally do is see one, know that physarum fuses into trunks outside a narrow occupancy band, or know that the attractor seed it just picked collapses to a single point. Atelier supplies the knowledge that isn't in the tutorials, machine-checks the numbers it quotes, and ships a capture script so the work gets judged by watching it.

Flow field, masked line-by-line reveal, type and palette — all from these skills, recorded with the capture script. Source: hero.html.

---

Pieces

→ Browse the generative gallery

Mycelium — 1,048,576 physarum agents on the GPU, four shader passes. The network is a consequence of a million agents agreeing.	Clifford — deterministic chaos, 420,000 points a frame. Brightness is visit density, log-scaled.
Murmuration — three local rules, no leader. Brightness is local flock density.	Strata — noise fed into the coordinates of noise, twice. Only the inner layer evolves, so it breathes.
Perihelion — bodies falling around three slow masses; they brighten at perihelion and cool at apoapsis.	Ember — a flow field where oxblood is calm and gold is fast.

The rule every piece follows: colour is earned. It maps to a property the system actually has — velocity, density, visit count, age — never to a rainbow of hue over time. Colour that encodes nothing is the loudest signal that no decision was made.

The generative-motion skill

Reference	What's in it
systems.md	Nine systems with verified parameter ranges and each failure mode: physarum, strange attractors, boids, reaction–diffusion, differential growth, n-body, metaballs, interference, cellular automata
noise-and-fields.md	Which noise and why, fbm, domain warping, curl vs angle fields, and turning a field into a composition rather than wallpaper
gpu-and-shaders.md	When to go GPU (architectural, not an optimisation), WebGL2 GPGPU with ping-pong state textures, attribute-less drawing, the silent-failure extension minefield, debugging in the order that finds it fastest
three-dimensional.md	Raymarched SDFs, smooth-minimum blending, free infinite domains, instancing at scale, and camera discipline
density-and-tone.md	Mapping an unbounded accumulation buffer onto a display — the most common way a correct piece looks wrong, and how to find the floor without guessing
post-processing.md	Bloom as a mip chain, grain weighted to the midtones, aberration, feedback, gradient-map grading, and the order these must run in
rendering.md	Accumulation trails, blending modes, envelopes, batching and scale
art-direction.md	The eight tells of tutorial work, the five decisions that make it a piece, and the tuning order
diagnostic.md	Black frame, saturating, dying, exploding, clumping, 4fps, beautiful-then-grey-mush — worked by symptom

The numbers are checked, not remembered

A parameter just outside the viable band doesn't produce a worse image — it produces nothing, and you can't tell by looking at the value. So the quoted values are verified by scripts that ship with the skill, CPU-only, in seconds:

skills/generative-motion/validate/attractors.py     # Lyapunov + occupancy per seed
skills/generative-motion/validate/gray_scott.py     # runs each (f,k), classifies the result

This caught a real error in this repo's own documentation. The reference used to advise animating an attractor by interpolating between two good seeds — measuring it showed 8 of 13 steps along that path are non-chaotic, collapsing the render mid-animation. The advice is now to orbit a single seed, and --path will check any route you want to take.

Both scripts also search: attractors.py --search clifford 10 finds new viable seeds and reports how many candidates it rejected on the way; gray_scott.py --scan maps the viable band so you can pick your own pairs.

Watch your own work

skills/motion/scripts/capture-motion.py sketch.html --out piece.gif --duration 2400

Drives headless Chrome over the DevTools protocol and records real composited frames, so it works for canvas, WebGL, CSS and JS animation libraries alike. Needs only Chrome and ffmpeg — no pip installs; the WebSocket client is stdlib.

Also in the box

Interface motion, for when the animation serves a UI rather than being the artwork.

Skill	Role
motion	Production recipes with tuned values and failure modes — modals, drawers, toasts, FLIP reordering, view transitions, drag, springs, game feel — plus an 11-point diagnostic for motion that feels dead. 20 running demos.
performance-craft	Compositor discipline on the web; ProMotion timing, Metal overdraw and particle budgets on Apple silicon; engine budgets; and a degradation ladder ordered by visual damage.
critique	Capture the real thing, score an 8-dimension rubric with evidence, output ranked changes with exact values. Accessibility findings outrank aesthetics.
design-direction · color · typography-and-layout	Aesthetic point of view, OKLCH palettes and contrast discipline, type and grid.

Install

/plugin marketplace add jangles-byte/atelier

/plugin install atelier@atelier

Or as plain skills:

git clone https://github.com/jangles-byte/atelier.git
cp -r atelier/skills/* ~/.claude/skills/

Skills auto-trigger on relevant work — "make something generative", "a flow field", "why is my shader black", "animate this". Each SKILL.md is a short router; depth lives in references/ that load only when needed, so context stays light.

Contributing & license

PRs welcome — see CONTRIBUTING.md. Principle first, exact values, verified numbers, no stack lock-in. MIT.