Paper2video — Design skill for Claude Code

paper2video

[](https://github.com/edwardyen724-g/paper2video/actions/workflows/ci.yml) [](LICENSE) [](pyproject.toml)

**Turn any technical article or paper into a 2–5 minute narrated explainer video, in a 3Blue1Brown-ish style. Open-source end to end.**


*Full sample: [`examples/karpathy_llm_wiki.mp4`](examples/karpathy_llm_wiki.mp4) — a 3-minute explainer of Andrej Karpathy's [LLM Wiki gist](https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f), generated end-to-end by paper2video in under 5 minutes of wall-clock time and ~$0.05 of Claude API credit.*

URL or PDF  →  ingest  →  research  →  script  →  Manim render  →  Kokoro TTS  →  ffmpeg mux  →  video.mp4

Built in an afternoon by an LLM and one human who got tired of bouncing off research papers without ever actually understanding them. Inspired by an Andrej Karpathy gist about LLMs maintaining persistent wikis — which the author learned about *by generating and watching the explainer this tool produced*.

Why this exists

There are way too many good ideas being published every day, and most of us never get to them because reading a paper is a 30-minute commitment with a 30% comprehension rate. paper2video tries to give you the *same understanding* in 3 minutes of audiovisual time. It's not a replacement for reading, but it's a much lower-friction first contact.

What it does

1. Ingests any URL or PDF (HTML extraction via trafilatura, PDF via pymupdf)
2. Researches the topic — extracts key concepts, optionally pulls supporting context from DuckDuckGo
3. Writes a script as a sequence of animated scenes with narration
4. Generates Manim code for each scene by asking an LLM to write Python animation code
5. **Validates