Claudebot skits

Short low-poly 3D comedy skits about coding with an AI, written as a JSON script and rendered to video with locally generated voices.

Try the working modelSource is private
Status
Experiment
My part
Design and all code
Started
28 Oct 2026
Input
A JSON script
Voices
Generated on my machine, no API cost
  1. Scriptskits/ep1.jsonLines, speaker, mood and camera shot for each line.
  2. Voicescripts/voice.mjsKokoro speech on my machine, or my own recorded take cut at the silences.
  3. Timingscripts/transcribe.mjsWord timestamps, so captions and pop-up screens land on the spoken word.
  4. Scenesrc/three/*.tsxLow-poly guy, bot and room in three.js, with a camera rig per shot.
  5. Renderremotion renderRemotion draws every frame in React and encodes the MP4.

Try it

Watch a real episode. The steps beside it are the pipeline that made it.

This is a small model that runs in your browser, not the real app. It follows the same rules.

A developer asks an AI to fix one bug. The AI rewrites everything. That is the whole genre, and I wanted to make short videos of it without opening a video editor.

A skit is a file

Each skit is a JSON file: a title, two voices and a list of lines. Every line names its speaker, the text, a mood and optionally a camera shot.

{
  "speaker": "bot",
  "text": "Done! I rewrote everything.",
  "mood": "happy"
}

One command turns that file into a finished video. Moods are neutral, happy, angry, shocked and tired. Shots default to a close-up of whoever is talking, or can be set to a wide or a two-shot.

How it renders

The scene is a low-poly developer and an orange robot, drawn with three.js. Remotion drives the scene frame by frame and encodes the result, so the animation is ordinary React code and every render of the same script is identical.

Voices are generated on my own machine with the Kokoro speech model, so a new episode costs nothing to produce. Lines that have not changed reuse their audio, which keeps re-renders quick while editing a script.