An agent skill that draws your story in 20 crayon styles

What it does

The project is a Remotion-based renderer paired with an agent skill-package/. Feed it Chinese story text or a stack of ordered images, and it generates a 3:4 vertical silent MP4 using a hand-drawn visual system—think colored-pencil diary, ink wash, or zine collage. An agent runtime such as Codex, Claude Code, or Kimi Code handles the natural-language prompting, sentence splitting, storyboarding, and image generation, while the renderer composites the layers and spits out a clean picture track meant for post-production dubbing.

The interesting bit

The real craft is in the 20 tightly defined style recipes—each with fixed example images and a JSON contact sheet—that lock the look so generated frames stay consistent across scenes. The default colored-pencil-diary style is deliberately naive, but you can switch to anything from kid-crayon chaos to linocut-editorial without retraining a model. The renderer also builds a local black-and-white underlayer that aligns with the final color frame, enabling a left-to-right text → sketch → color reveal.

Key highlights

• 20 built-in hand-drawn styles (colored pencil, ink wash, gouache, zine riso, whiteboard, etc.) selectable by name, ID, or alias.
• Agent-driven workflow: natural language commands trigger sentence splitting, storyboarding, image generation, and rendering.
• Silent H.264 output (1080×1440 or 720×960 preview) designed as a picture track for later voiceover and music.
• Accepts either raw Chinese story text or ordered image uploads, preserving original wording and composition.
• Optional bottom-right page-curl transition with a ghost of the previous frame on the paper back.

Caveats

• Requires a full local toolchain (Node.js 20+, Python 3.10+, FFmpeg, Chrome) plus a compatible agent runtime; this is not a lightweight web app.
• Image generation defaults to Codex Image2 and only falls back to OpenAI if explicitly requested, so your results depend on whichever image provider the agent uses.
• The README warns that time jumps, ambiguous references, medical scenes, or age-sensitive characters should be reviewed manually via a visual plan before rendering.

Verdict

Content creators and developers building automated short-form video pipelines will find the style discipline and agent integration useful. If you just need a one-off slideshow, the toolchain is probably overkill.