AGENT SKILL
MediaLet Claude Code produce 3Blue1Brown-style explainers with Manim
Yusuke710/manim-skill
156 stars, MIT, site manimate.ai. Four steps: Plan (storyboard, optionally in plan mode), Code (write Manim scenes in Python), Render (keep going until every scene renders successfully) and Iterate (open the video viewer in a browser and paste your notes back). The project layout is fixed as plan.md, script.py, concat.txt and final.mp4.
Our takeStep three is the point: the agent owns the render-until-success loop instead of handing errors back to you. That is exactly where Manim costs humans the most, since one wrong coordinate means another ten-second re-render. The repo also writes 'avoid generic AI slop aesthetics' into its goals, treating taste as a requirement rather than luck, and keeps plan.md as a separate file so iteration does not require re-explaining context.
/plugin marketplace add Yusuke710/manim-skillMaking video the way you write code: Plan → Code → Render → Iterate
Yusuke710/manim-skill (156 stars / 9 forks, MIT, site manimate.ai) lets Claude Code autonomously produce 3Blue1Brown-style explainer video. Manim is the math animation engine open-sourced by 3Blue1Brown's creator and scenes are described in Python — so "make a video" becomes "write code and run a render", which is exactly where an agent is strongest.
The README's opening line is the pitch: "Create videos like you write code with Claude Code." Four steps:
| Step | What happens |
|---|---|
| Plan | A better plan makes a better video. Use Claude Code's plan mode to storyboard the way you would plan before coding; without it, Claude designs the scene structure automatically |
| Code | Claude writes the Manim code in Python |
| Render | Claude runs Manim itself until every scene renders successfully — it does not hand back code and walk away, it fixes errors until there is a film |
| Iterate | Claude opens a video viewer in your browser; you paste your notes back into Claude Code and it refines |
Step three is the point. Making the agent own the "run until it succeeds" loop instead of throwing render errors back at a human addresses the most expensive part of Manim work — one wrong coordinate costs another ten seconds of re-render.
Taste is an explicit goal
The repo calls out avoiding generic AI slop aesthetics — the cheap look that is obviously machine-generated. The project structure is fixed to support that:
plan.md # storyboard and narrative plan
script.py # Manim scene code
concat.txt # clip concatenation list
final.mp4 # final cutKeeping plan.md as its own file matters: iteration edits a plan and code, so the context never has to be re-explained.
Dependencies and install
# System deps (macOS)
brew install cairo pkg-config ffmpeg
# Manim Community Edition
uv tool install manim
# Install the skill
/plugin marketplace add Yusuke710/manim-skill
/plugin install manim-skill/manim-skill
# Optional: Manim CE best practices
npx skills add adithya-s-k/manim_skill/skills/manimce-best-practicescairo/pkg-config/ffmpeg are hard requirements; on Linux use the equivalent packages (libcairo2-dev libpango1.0-dev ffmpeg pkg-config). Rendering is CPU-bound, so budget time for longer videos.
Where it fits and where it does not
It is best for concept explanation: math derivations, algorithm visualisation, geometric transforms, physical processes — Manim's native ground, and things a slide-screen-recording cannot do. It is not for live-action editing, complex character animation, or brand films needing precise art direction.
At 156 stars this is still early. But the reasoning is complete, dependencies are clear, and the iteration loop is well designed; as a reference implementation of agent-driven video production its value is not in the star count. Our own AI Games and demo-video line borrows the Plan→Render→Iterate loop directly.