Communitygithub.com

WhiteRabbitLAB/baitu-video

A complete method for creating explainer videos in any topic (science, AI usage, history, reviews, etc.) using code, with built-in format, scene, and style libraries.

What is baitu-video?

This skill provides a comprehensive, code-driven workflow for producing high-quality explainer videos across any topic. It is designed for AI agent users (e.g., Claude Code, Cursor, Codex) to automate the entire production pipeline: from topic selection and fact-checking, to script writing and voiceover generation, to scene composition and rendering. The system is structured around four independent libraries: formats (narrative skeleton), scenes (visual world), styles (artistic rendering), and channel profiles (branding, subtitles, platform rules). Users interact via four guiding questions (purpose, industry, desired style, target platforms) and receive a full video output in the correct aspect ratio (horizontal, vertical, or both). The skill is highly extensible—each library supports adding new entries via templates—and includes a shared design library for fonts, color palettes, components, animations, and camera transitions. It currently supports two example channels (nutrition science and AI usage) and is tested with bilingual voiceovers (Chinese/English) and cross-platform delivery.

Works with✓Claude Code~Codex CLI~Cursor
npx skills add WhiteRabbitLAB/baitu-video

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

Explainer Video Production

English: SKILL.en.md

Topic-agnostic. Currently distilled from the author's two channels: nutrition science (10 episodes) and AI usage (6 episodes + one 17-minute overview long film). The formats, scenes, and styles have been written in a topic-independent way, so examples from both can be used interchangeably.

Know Before You Use: The Skill Is a Floor, Not a Ceiling

Rules come in three levels of weight: hard rules (facts, safety, platform, decision-maker decisions – must follow), capability boundaries (set only because the model couldn't do it at the time, dated – retest when models improve), and defaults (parameters, templates, skeletons – a starting point, not an upper limit; do better if you can, leave traces for comparison). See the beginning of principles.md for details. Do not abandon a better approach just because the documentation specifies a parameter, and do not skip something just because it's not documented.

User Entry: Four Questions

  1. What do you want to do? (Science explainer / Story documentary / Tutorial / Review & comparison / Opinion / Listicle / News briefing)
  2. What industry? (Health & nutrition / Tech & AI / Humanities, history & art / Natural sciences / Social science, psychology & law / Finance & business / Education & learning / Career skills / Lifestyle)
  3. What feels right? (Choose from style thumbnails)
  4. Which platforms? (Horizontal platforms: YouTube, Bilibili, WeChat video, etc.; Vertical platforms: Douyin, WeChat vertical shorts, Xiaohongshu, etc.) → Output the corresponding aspect ratio only: only horizontal platforms → only horizontal format; only vertical platforms → only vertical format; both → produce both (decision-maker, 2026-10-07). Vertical formats are re-laid out for portrait orientation, not cropped from horizontal.

→ Check recommended combinations in catalog/. Each combination is labeled: ✅ ready to make (finished episode exists), 🧪 needs testing (doable but no sample episode yet), ⛔ temporarily not possible (requires live host, real footage, or licensed material). For popular reference works that can be converted into our templates: catalog/references.md. Styles will be added continuously: styles/ROADMAP.md.

Internal Structure: Four Layers + Design Library

The user chooses by 'purpose × industry × feel'; the library is stored in the four layers below, independent of each other, so any layer can be replaced individually (customizable).

LayerWhat it governsExampleLibrary
FormatWhat to say, how to say it (narrative skeleton)Real experiment documentary, story + method + test, scene walkthroughformats/
SceneWhat world the visuals show; what script sources, components, and recurring elements existArchive desk, strategy game, tech terminal, gallery wallscenes/
StyleHow to draw: lines, surfaces, light, materials, animation feel, character designPaper flat vector, flat geometric, dark tech UIstyles/
Channel profileSettings unique to a channel: mascot, subtitle specs, voiceover, platform rules, decision-maker preferences, toolchain commandsFilled exampleprofiles/

Scenes and styles share a design library (design/): font directory (by script source), color palettes, components, animations & materials, camera motion & transition library (each parameter annotated with source: measured / first-hand / ours / pending). Text in frames is categorized by 'text role' (text-roles.md): when changing scenes or styles, the font, carrier (dialog bubble, note, panel), and entry animation all follow along; subtitles and chapter progress bars belong to channel configuration and do not change. When character performance is needed, add the character module (characters/): use image generation as art, code as the engine.

Engine

Run-ready example period (bilingual voiceovers CN/EN, both horizontal and vertical formats): examples/sky-demo. Commands for making episodes, shared visual components, and inspection tools are in engine/. First time using the project, run self-check from the project root: node .claude/skills/explainer-video/engine/cli.mjs doctor.

What to Read First

TaskRead
First time using in a projectFollow the 'Start' 5 steps in engine/README.md: explainer.json → copy the entire folder profiles/_template/ into your project's profiles/<your_channel>/ (do not put it inside the skill directory, it will be overwritten on upgrade), fill in the three files: channel / voice / preferences; see profiles/example/ for reference (Rabbit and preferences are private to the author, not licensed)
Start an episode from scratchpipeline.md + principles.md + your channel profile
Choose format / scene / styleFirst go through the four questions in catalog/; then check the INDEX of the three libraries
How to define on-screen texttext-roles.md → scene file's script source table → design/fonts.md
How to transition between shots and posestransitions.md + camera motion & transition library design/camera-transitions.md (all parameters have reference sources)
Character performancecharacters/act-pipeline.md
Before deliveryacceptance.md

Room to Grow: Each Library Can Expand

To addProcedure
FormatCopy formats/_template.md; write narrative skeleton, essential shots, transition preferences; register after sample verification
SceneCopy scenes/_template.md; script sources → fill R1–R10 completely, recurring elements, components; register after sample verification
StyleCopy styles/_template.md; must write how characters are drawn in this style; generate samples using existing scenes and characters, then register
Font / Color palette / Component / AnimationAdd a line to the corresponding file in design/ (fonts must be commercially usable)
CharacterFinalize according to characters/act-pipeline.md, then register
ChannelCopy the folder profiles/_template/ (three files: channel / voice / preferences)

After finishing each episode, write any newly discovered pitfalls back into the relevant files.

Status & Unverified

To-do list (numbered by suggested order): TODO.md (author's internal record, not included in the open-source version; public style roadmap is in styles/ROADMAP.md).

  • Scenes and styles are separate in the library; past episodes are 'one big HTML per episode', combining both. Common components have been extracted into style-swappable shared code engine/vc/ (components only recognize text roles; appearance is provided by styles.js; one set per 10 styles; thumbnails in styles/INDEX.md); one finished episode (N10) is fully built with it.
  • Unverified: swapping styles in the same scene (real shoot); mixed editing of multiple styles; two or more characters in a single frame; character lip-syncing to voiceover; character performance integrated into the final episode.
  • Formats (story-method-test, howto-guide, product-review, timeline-epic) and scenes (terminal-tech, field-notebook, gallery-wall), derived from the AI usage channel's scripts, handover documents, and font tables, have not yet been used in new episodes; the gallery-wall scene's visual pages have been deleted, and its entire font table is inferred.

Related Skills