Most short videos about code are a few text scenes and a diagram. So I wanted to write those as data, not drag them around in an editor.
With yamlreel you write a short YAML file with the scenes, and it renders the video:
pip install "yamlreel[ffmpeg]"
yamlreel init system-design my-video.yaml
yamlreel render my-video.yaml --preview
Each render gives you a folder per format (landscape, vertical, square) with the video, a poster frame for the thumbnail, a contact sheet of stills to check before you publish, and a share-copy.txt with a title, description and hashtags.
It draws animated diagrams and typed code, scans in tables and arrays, adds word-by-word subtitles and can add a voiceover. The background music is generated on your machine, so there's nothing to license.
The demo video on the release page was made by yamlreel from one YAML file, with no video editor. It's MIT licensed and needs Python 3.10+ and ffmpeg.
If you try it, tell me what's awkward in the YAML format. That's the part I most want feedback on.

Top comments (2)
Since you asked about awkward bits in the YAML: the thing that usually hurts in "video as data" formats is hand-written durations. If a scene's length can default to "auto" (voiceover length, or typing speed times the code length) with an optional override, people stop fighting the timing. YAML anchors for shared styles, validation errors that point to the exact line, and a way to render a single scene (something like --only scene3) would also make iterating much faster. Disclosure: I build Vidmoat, a hosted AI video editor where you (or Claude/Cursor over MCP) edit a real timeline by chat, so I think about this same trade-off a lot. Writing it as data wins for repeatable explainers like yours, and the contact sheet before publishing is a great touch.
Thanks @fredabila, this is useful. Anchors with merge keys already work for repeated scenes, but you’re right there’s no good home for shared defaults, and validation errors name the path (scenes[1].secnds) but not the line. Rendering a single scene is also missing, and I could have used it today. Per-scene durations already stretch to the narration, but an explicit seconds: auto with typing-speed defaults for code scenes is a good idea. I’ll look at these. Glad the contact sheet helps.