3d-character
Skill75 starsChanged 40 days agoApache-2.0View on GitHub
Style for edit-video: a cute 3D character reel with no camera and no Blender. A chibi kid (a boy or a girl, colours and hair yours to set) talks to camera on a soft pastel stage, lip-synced to a locked voice, with camera cuts on spoken words, a floating card that types in, and the line word-lit on top. Built in Three.js and rendered by HyperFrames on your machine.
In this repository15 other files it gives its agents.
- AGENTS.mdAGENTS.md
- .claude/CLAUDE.mdCLAUDE.md
- find-ideasSkill
- write-scriptSkill
- edit-videoSkill
- animated-explainerSkill
- film-stylesSkill
- product-motion-adsSkill
- cartoon-characterSkill
- adaptive-glassSkill
- breakout-cardSkill
- motion-effectsSkill
- paper-gridSkill
- text-over-footageSkill
- publish-and-analyzeSkill
The file
Files
3D character
A style on top of edit-video, which owns the voice, sound mix and review. The whole
reel is a small 3D scene drawn in code: a chibi character built from simple shapes,
one pastel colour for the floor and the sky, soft light, and the camera cutting on
the words. three.js (pinned 0.168.0) comes from the workspace runtime that edit-video's
scripts/init-workspace.mjs --install prepares, like GSAP; the builder copies it into each
reel, and HyperFrames renders it like any other composition.
The look: simple and cute
- One character, chibi proportions: a big round head, glossy dot eyes, blush, a single smooth fringe, a small body and stubby arms. It talks with the voice, blinks, nods on accent words, points, hops and waves; on a happy beat its eyes curve shut.
- One colour, soft light. The floor and the sky are the same pastel, so the character floats in a clean space with a soft shadow under it. No desks, no busy props, no dark moody sets: they read as clutter at phone size.
- At most one prop: a floating card that pops in, types two short lines and fills a progress bar, then leaves before the next wide shot.
- The spoken line on top, big and word-lit, with the accent word in a serif italic in the accent colour.
The character
CHARACTER in the spec sets it:
| Option | Values |
look | 'boy' (side-swept hair, hoodie, shorts) or 'girl' (two buns, top, skirt) |
skin, hair, top, bottom, shoes | Any colour, as '#rrggbb' |
Pair it with a stage colour that flatters it: peach #FFE9D6 with a blue accent for
the boy, sky #E6F4FF with a pink accent for the girl.
A new character is new shapes in assets/characters.js: spheres and capsules for
the body, a mouth that scales with the voice, eyes that blink. Keep the same
update(state) contract and the stage drives it. Keep it simple: a few big shapes,
big eyes, one colour per part.
Workflow
Work in a run folder from edit-video (scripts/new-edit.mjs). Each reel is a
folder work/<reel>/ with its spec.py and voice/. Start from
the worked example.
- Script. One idea, 10–40 s, short lines. Mark the word to stress with
*stars*; the character nods on it. - Voice.
edit-video'sscripts/voice.py work/<reel> --engine files|hyperframeswritesvoice/vo.wavand every word's time. Use a voice that suits the character;edit-video's voice options cover free local voices and cloning. - Spec.
build(T)returns the shots (full,close,side) and when they cut, the acts (wave,point,hop,happy) and when they start and end, the card, and one sound cue per visible event. The contract is in the docstring ofscripts/build.py. - Build:
python3 <skill>/scripts/build.py work/<reel>. It copies the stage, the character and three.js into the reel and writesindex.htmlandcues.json. - Look before rendering.
hyperframes snapshot work/<reel> --describe false --at …at the hook, each cut, the card and the last second. Or openindex.htmlfrom a local web server and firehf-seekevents in the browser console. - Render:
hyperframes render work/<reel> --fps 30 --quality delivery --workers 2 --output previews/<reel>-v01.mp4. It draws at twice the size and scales down, so allow about 10 s per second of video. WebGL renders on the GPU. - Sound and review with
edit-video:sfx-sync.mjs mix work/<reel>/cues.json --voice work/<reel>/voice/vo.wav …, then the usual review of the encoded cut.
Shots and acting
- Open full: the whole character under the hook, so frame 1 reads at a glance.
- Close on the line that matters, side shot when the card comes in, back to full for the ending.
- Point at the card while it's on screen, then wave and go happy on the call to action.
- Cut on spoken words, about every 2–3 s. Hold a shot longer only while the character is doing something.
Pitfalls already paid for
- A detailed mascot with realistic eyes, fur and a moody set looked off-putting. Simple shapes and a pastel stage read as cute.
- Framing by height alone crops a wide character. The stage frames by height and
width (
widthin the character). - A card left in the last wide shot gets cut off at the frame edge. Give it an
outtime. - The composition has no GSAP timeline; the root carries
data-no-timelineso HyperFrames' lint accepts it, and time comes from thehf-seekevent.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

