Introduction

Mathematical animation has emerged as a powerful medium for communicating complex ideas, pioneered by Grant Sanderson's 3Blue1Brown channel. His open-source animation engine, Manim, enables programmatic creation of publication-quality mathematical visualizations. However, writing Manim code requires deep familiarity with its API, common failure modes, and visual design principles that make the difference between a mediocre animation and a compelling one.

We present 3brown1blue, a system that bridges this gap by encoding expert knowledge about Manim animation into a structured skill that AI coding assistants can consume. The system operates at two levels:

1. A Claude Code skill (21 rule files, 3 templates, 2 utility scripts) that teaches the AI how to write correct, visually compelling Manim code
2. A pip-installable CLI (pip install 3brown1blue) that provides generate, install, status, update, and uninstall commands

Methodology

Knowledge Extraction

We analyzed 422 frames from three 3Blue1Brown videos (Neural Networks Ch2, Transformer Ch6, MLP Ch8) and extracted 22 implementable visual patterns including probability sidebars, skip arcs, grid fills, interactive sliders, and heatmaps.

We identified 12 visual design principles from the works of Edward Tufte, Bret Victor, and Grant Sanderson:

1. Geometry before algebra
2. Opacity layering (primary 100%, context 40%, grid 15%)
3. Persistent context
4. Linked dual representations
5. Parameter manipulation via ValueTracker
6. Continuous morphing
7. Question frames
8. Annotations ON objects
9. Color as semantic data
10. Concrete values over placeholders
11. Progressive complexity
12. Emotional anchoring

Through iterative production testing, we added 4 more principles learned from analyzing competitor content:

13. Live values in diagrams -- numbers flow through diagrams and cascade on parameter changes
14. Density ramp -- visual complexity mirrors conceptual complexity (start with 2-3 elements, end with 15+)
15. Per-scene skeleton -- one anchor diagram per scene that never disappears
16. Caption zone -- bottom 20% of frame reserved for narration text

Crash Prevention

We documented 12 common Manim failure modes (6 crashes, 6 silent bugs) and created safe_manim.py, a drop-in wrapper library that prevents the 6 crash cases:

• Create(Text()) tracing outlines instead of writing
• Arrow(color=interpolate_color()) crashing
• Brace.get_text(font_size=) crashing
• MathTex with dollar signs crashing
• LaggedStartMap(Write, group) crashing on Text
• Updaters freezing during wait()

Production Quality Rules

Through iterative testing, we established production rules including:

• $\text{DIM_OPACITY} = 0.1$ on dark backgrounds (not 0.3)
• Manual text wrapping (Manim has no auto-wrap)
• Container boundary enforcement
• Safe coordinate bounds: $x \in [-6.5, 6.5]$, $y \in [-3.5, 3.5]$
• Minimum data visualization thresholds (bar width $\geq 0.3$, dot radius $\geq 0.06$)

Results

Video Generation

Using the skill, we autonomously generated four complete mathematical explainer videos in a single Claude Code session:

Each video followed the full pipeline: web research (2000+ words) $\to$ curriculum design (10-12 scenes) $\to$ scene code $\to$ low-quality test render $\to$ 1080p final render $\to$ ffmpeg stitching.

Multi-Editor Support

The skill installs into 5 AI coding assistants:

• Claude Code (~/.claude/skills/manim/)
• Cursor (.cursor/skills/)
• Windsurf (.windsurf/skills/)
• Codex
• GitHub Copilot

Package Distribution

The tool is distributed via PyPI as 3brown1blue (v0.1.2) with zero required dependencies beyond click>=8.0. Optional LLM provider dependencies are available for standalone generation.

Discussion

The key insight is that structured skill engineering -- encoding domain expertise as rule files, templates, and crash-prevention wrappers -- dramatically improves AI code generation quality in specialized domains. Without the skill, Claude writes Manim code that hits approximately 40% of common failure modes. With the skill, the crash rate drops to near zero.

The visual design principles (especially principles 13-16, learned from competitive analysis of Insightforge's backpropagation short) represent a shift from "make it work" to "make it compelling." The density ramp principle alone -- starting sparse and ending dense -- transforms flat, monotonous animations into narratives with rising tension.

Limitations

1. Videos lack narration -- scene transitions feel abrupt without voiceover
2. Text overflow still occurs in dense scenes despite manual wrapping rules
3. The winding machine visualization (wrapping number line around circle) was confusing without narration, suggesting some concepts resist pure visual explanation
4. Video durations are 4-5 minutes rather than the target 10 minutes

Conclusion

We demonstrate that AI agents can autonomously produce multi-scene mathematical animations at production quality when equipped with structured domain knowledge. The 3brown1blue skill, available at github.com/AmitSubhash/3brown1blue and pypi.org/project/3brown1blue, makes this capability accessible to any AI coding assistant.

Acknowledgments

This work builds on Grant Sanderson's 3Blue1Brown and the Manim Community Edition. The visual design principles, pacing instincts, and geometry-before-algebra philosophy are Grant's contributions to mathematical communication. We studied 422 frames of his videos to extract the patterns encoded in this skill.