sweety-handdraw-video: an agent skill for narrated hand-drawn explainers

npx skills add threerocks/sweety-handdraw-video --skill sweety-handdraw-video

Getting a narrated video in which the drawing keeps pace with the voice is hard to do from one prompt: the voiceover runs long, the pictures drift, and the drawing itself is a black box. sweety-handdraw-video is an MIT-licensed Agent Skill from threerocks that gives Codex or Claude Code a fixed pipeline for the job: script, timed narration, storyboard, line-by-line drawing and local rendering. The default delivery is a 1280×1920, 30 fps MP4 with a cover, subtitles, the script and a check record.

How it works

The skill fixes the audience, the problem, the mechanism and the call to action first, then generates the narration one paragraph at a time: each under about 100 Chinese characters, cut only at sentence ends. Paragraphs that fail a hum, noise or clipping check are dropped, and the rest are joined with measured pauses and timestamped automatically. The storyboard then works backwards from those seconds and gives every drawing a stroke budget. The author measured the default method, a light block-in followed by marker color, at roughly 3.9 strokes per second, so pictures are designed to fit the time rather than the time squeezed to fit the pictures.

Each illustration starts as a pure black line drawing and must pass a budget check before color is added. The pen blocks in the shape, draws object by object, erases the blocking lines, and only then colors from light to dark with a marker.

What it constrains

• One subtitle track: burned-in captions, subtitle events, MP4 subtitle tracks and web tracks must not stack, and delivery checks for extra tracks and sidecar files.
• Stroke order for Chinese: characters in GB2312 follow standard stroke order and other glyphs are traced from their shapes; a missing glyph fails the build.
• No image-to-video step: the film is composed and drawn locally, and generated pictures are only source material.
• AI labeling stays: the output carries an implicit AI-generated mark, and publishers must still declare AI generation on the platform.

Who it’s for

It suits solo creators and small teams making explainers, opinion pieces or fables where every stroke should land on the narration. A single illustration or a script alone is outside its scope, as its own description says.

The prerequisites are real. The host must be able to generate images: Codex has that built in, while Claude Code needs an image tool or MCP, or illustrations you supply. Without one, the skill stops rather than delivering a text-only video. You also need FFmpeg, Python and Chinese fonts. The setup script installs FFmpeg and a separate Python environment, and on Linux the Noto CJK fonts; Chromium (about 150 MB) and the VoxCPM2 model (about 3.2 GB) download only if you add a flag. For voice, use a local VoxCPM2 model on Apple Silicon, or a VoiceStudio service you run yourself. macOS and Linux are supported; Windows is not promised.

Handwriting fonts are yours to supply. On macOS the docs suggest Hannotate SC and fall back to Hiragino Sans GB, and they say a fallback cannot count as an accepted handwriting style. On Linux you install LXGW WenKai yourself.

Licensing needs a look. The repository is MIT, but the five vendored upstream guides keep their own licenses, and a cartoon character that its upstream author restricts to demonstration use is not shipped. The two realistic voices exist only as short clips in the package; their source recordings stay on the owner’s machine, and cloning someone else’s voice is out of scope. Some style reference images are fetched from jsDelivr on first use and checked.

The two sample films in the README (49 seconds and 1 minute 52 seconds) were made with v1.2 and keep its preview watermark, so they do not show the 3.0 line-first flow. This entry is based on v3.0.2’s README and SKILL.md; the editorial team has not yet run a full video end to end on its own machine.

Try a short topic first. Six choices come up before work starts: style, hand, voice, pen-text advert, watermark and subtitles. Anything left unchosen after 60 seconds falls back to a plan drafted from the content, and neither advert nor watermark is added by default. In Claude Code, ask to use sweety-handdraw-video, or type /sweety-handdraw-video; in Codex, use $sweety-handdraw-video.

For a sibling, story-to-handdrawn-video turns a Chinese story into a silent 3:4 flipbook, while this skill keeps narration, drawing and subtitles in one pipeline. For a presenter-led alternative, see lanshu.