Skip to content

smykla-skalski/explainer-video

v1.0.2MIT

Make 3Blue1Brown-style explainer videos: Manim animations on black with CMU Serif text, narrated with ElevenLabs text-to-speech, from a topic to a checked 1080p60 MP4 with the audio synced, transcribed and checked for clicks and gaps.

GitHubOpen Website

explainer-video

Make 3Blue1Brown-style explainer videos: Manim animations on black with CMU Serif text, narrated with ElevenLabs text-to-speech. The agent writes the narration, voices it, animates each beat on the words the voice says, and builds one 1080p60 MP4. Before it reports, every clip is trimmed, transcribed and checked for clicks and dead air, the soundtrack is checked against the picture, and every frame is reviewed on contact sheets.

The plugin is a portable Agent Plugin with one Agent Skill, so the same package works in Claude Code, Codex and other compatible agents. It runs on macOS and Linux.

Installation

Claude Code:

claude plugin marketplace add smykla-skalski/sai
claude plugin install explainer-video@sai

Codex:

codex plugin marketplace add smykla-skalski/sai
codex plugin add explainer-video@sai

Other agents that read Agent Skills can use skills/explainer-video/ directly, for example by linking it into ~/.agents/skills/explainer-video.

Requirements

  • ffmpeg and ffprobe, uv and Python 3. uv installs Manim (on Python 3.13), numpy and faster-whisper on first use; LaTeX is not needed.
  • An ElevenLabs API key in ELEVEN_API_KEY, or tell the agent where to read it (for example a 1Password reference). Narration costs about 1,000 ElevenLabs characters per minute of video, plus retakes.
  • Network access to ElevenLabs, PyPI, CTAN (the CMU fonts) and Hugging Face (the Whisper model). In a sandboxed agent, allow these commands to run outside the sandbox.

Usage

Ask in plain words, for example "make a 5-minute 3b1b-style video on how RAG works", "a short narrated explainer about bloom filters" or "fix the click at 1:31 in my video". In Claude Code, /explainer-video works too, and in Codex $explainer-video.

A full video takes about an hour of agent time; a one-minute video less. The agent cannot listen to audio, so it reports what it measured and asks you to name a timestamp for anything that sounds or looks wrong.

How it works

skills/explainer-video/scripts/new_project.py copies the pipeline into a project folder: tts.py voices script.py and trims each clip, video.py holds the Manim scenes built on the shared helpers in explainer.py, build.py renders the scenes in parallel and builds the soundtrack from the clips at the times the scenes logged, and narration.py check compares the final file with that reference. transcribe.py and sheets.py drive the review. scripts/selftest.py builds the template project with synthetic tones, at no ElevenLabs cost, then checks that a project made by an older version still renders after a refresh.

To bring a project made by an older version of the skill up to date, run new_project.py on it: it refreshes the shared code, removes the old build.sh and sheets.sh, and adds the Manim import that an old video.py needs.

When a video shows a defect in the skill itself, the agent fixes the project copy and writes a GitHub issue draft for this repository instead of editing the installed skill, which plugin updates replace.

License

MIT