Manim Video

Manim Video

Use this skill for Python-based Manim projects that need a clean planning phase, strong visual direction, reusable scene structure, and a deterministic render pipeline.

The output contract is usually a workspace project directory containing plan.md, script.py, concat.txt, draft renders, and optionally a stitched final.mp4.

Creative Standard

This is educational cinema, not animated slides.

• Make it beautiful and professional. The draft should feel intentional, polished, and presentation-ready, not like a rough internal diagram dump.
• Before writing code, define the narrative arc: confusion -> visual hook -> mechanism -> payoff -> recap.
• Geometry before algebra. Show the shape, motion, system flow, or data change before the formal equation.
• Every scene teaches one dominant idea. If a scene has two unrelated claims, split it.
• Direct attention with opacity layering: primary elements at 1.0, contextual elements around 0.35-0.45, and structural elements like axes or grids around 0.12-0.2.
• Give reveals breathing room. Add self.wait() after every important animation and make the main aha moment the longest pause.
• Keep one cohesive visual language across the whole project: shared palette, shared timing constants, shared font rules, and stable concept-color meaning.
• Avoid repetitive scenes. Vary dominant color, layout, and entry animation across scenes while keeping the overall palette consistent.
• First-render quality matters. If the layout is cluttered, the typography is inconsistent, or the pacing feels rushed, the scene is not done.

Modes

Stack

Single Python script per project. No browser, no Node.js, no GPU requirement.

Pipeline

PLAN --> CODE --> RENDER --> STITCH --> AUDIO --> REVIEW

• PLAN: write plan.md with the narrative arc, misconception, aha moment, palette, typography, and scene beats.
• CODE: write script.py with one independently renderable Scene subclass per clip.
• RENDER: draft render all scenes at -ql unless a real render blocker prevents it.
• STITCH: update concat.txt and stitch a draft video when ffmpeg is available.
• AUDIO: add narration or music only when the user asks for it or the project explicitly includes it.
• REVIEW: verify timing, readability, clean exits, and overall visual coherence before claiming the output is done.

Creative Direction

Color Palettes

Start with a HybridClaw palette unless the user explicitly asks for a different visual language.

Animation Speed

Typography Scale

Font Rules

Minimum readable text size is font_size=18.

Text density rules:

• If more than 4 text elements are strongly visible, reduce body text toward 20-24 instead of 26-30.
• Keep title/subtitle stacks to 2 text blocks. If a third block is needed, move it to a side label, caption, or a new scene.
• Clamp all titles before .to_edge(...), for example fit_text(title, max_width=11.0).
• Clamp any potentially long text block to the usable frame width instead of trusting the raw font_size.
• If a width-clamped title still feels oversized, lower it toward 34-36 or split it into 2 lines. Do not keep a full-sentence title at 48.
• Do not write new text on top of old text in the same area. Use ReplacementTransform, FadeOut, or a layout shift first.
• Use buff >= 0.5 for edge-positioned text so draft renders do not clip.

Performance Targets

Default Workflow

1. Run the bundled setup check first:

python3 skills/manim-video/scripts/check_setup.py

Treat the setup check as advisory by default. Missing manim, ffmpeg, or pdflatex should stop render commands, but it must not stop planning or script editing. 2. If the user wants a new project, create the target directory and write plan.md, script.py, and concat.txt directly. Preserve the exact requested output path. 3. Fill in plan.md before heavy coding. Read references/scene-planning.md and references/visual-design.md when the narrative, pacing, or screen layout is still fuzzy. When the user already gave a topic, draft the audience, teaching objective, misconception, aha moment, narrative arc, palette, typography, subtitle strategy, and scene beats yourself instead of asking the user to fill placeholders. 4. Keep reference loading small. For a normal request, start with only these core references:

• planning and scene beats: references/scene-planning.md
• layout, typography, and pacing: references/visual-design.md
• shared constants, title helpers, and scene structure: references/python-patterns.md Read at most 3 references before the first draft of script.py. Do not bulk-read the whole reference folder.

5. Add only one specialty reference when the task actually needs it:
   • symbolic derivations: references/equations.md
   • charts, counters, and algorithm state: references/graphs-and-data.md
   • moving cameras or 3D scenes: references/camera-and-3d.md
   • reactive geometry or continuously changing values: references/updaters-and-trackers.md
6. Read stage-specific references only when you reach that stage:
   • render and stitch commands: references/rendering.md
   • failures or confusing Manim behavior: references/troubleshooting.md
   • final polish before -qh or claiming the draft is done: references/production-quality.md
7. Implement script.py with one Scene subclass per clip. Use shared constants for palette, typography, and timing. Put subtitles on significant animations, keep scene exits clean, and make every scene independently renderable.
8. For ordinary requests like "create a short animation explaining X", do not stop after planning and scripting when rendering is available. Produce a draft render in the same turn: render all scene classes at -ql, update concat.txt, and stitch a draft video if ffmpeg is available. Render is part of the normal pipeline, not an optional next step.
9. Use manim ... as the default render entry point. A global manim CLI from uv tool install manim is fine. Fall back to python3 -m manim ... only when Manim is installed into the host interpreter but the CLI is unavailable.
10. Before production render, run the production checklist in references/production-quality.md. Use -qh only when the user explicitly asks for a final or high-quality export. For the default one-turn path, a stitched draft render is the finish line. Use references/rendering.md for the exact command shapes.

If required tools are missing, the fallback is still useful:

• create plan.md
• write or revise script.py
• explain what still needs to be installed
• stop short of claiming a video was rendered