paper2video

A Claude Code skill that turns a research paper or project into a polished explainer / promo video, in the style of 3Blue1Brown, built with Remotion.

Give Claude this skill and a link to the work: a GitHub repo, an arXiv paper, a project page, or a working directory it can read. It then produces the whole video:

• Story and script, written for viewers with no context and checked claim by claim against the paper.
• Narration in several forms:
  • ElevenLabs (optional key);
  • free edge-tts voices;
  • your own recording;
  • no voice at all.
• Animations built from real data, with word-timed captions, music with ducking and subtle sound design.
• Two languages: English and Chinese cuts, plus an optional vertical (9:16) cut.
• Publishing kit: covers, titles, descriptions and chapters for YouTube and Bilibili, and on request any other platform.
• Editor handoff: stems, SRT captions and EDL markers if you want to finish in your own editing app.

Showcase

Two videos made with this workflow:

Hook, task and failure. The paper's hero visual, built from real data, opens the film. A primer then explains the task, and the models' actual accuracy curves show where they fail.

Mechanism and results. A layer cursor sweeps a heatmap of real read-out data. Every chart has a source line, and caveats are shown rather than hidden.

Real model outputs and trajectories. Steered text is typed out with its key phrases highlighted, the 3D trajectories are the paper's own, and a strength dial runs over the robustness curves.

Covers and vertical cuts.

Install

With the skills CLI:

npx skills add Lunamos/paper2video

Or manually:

git clone https://github.com/Lunamos/paper2video
cp -r paper2video/skills/paper2video ~/.claude/skills/

You don't need to set up Remotion yourself. The skill tells Claude to:

• create the Remotion project;
• install the official Remotion agent skills;
• copy in the template;
• install whatever it needs (Node packages, and uv for the Python helpers).

You only need Node.js 18+, ffmpeg and Python 3.10+ on the machine.

Optional: set ELEVENLABS_API_KEY in the environment or in the project's .env to get the most expressive narration and a generated music bed. Without a key, the skill uses free edge-tts voices, or your own recording, or no voice. It never prints or commits the key.

Use

Use the paper2video skill to make an explainer video for https://github.com/<you>/<repo>
(paper: https://arxiv.org/abs/xxxx.xxxxx). English for YouTube and Chinese for Bilibili,
third person, keep it under 4 minutes.

You can also point it at a local folder or a server it can read. It asks only what it can't sensibly default, then works end to end. Afterwards, give feedback the way you would to an editor, e.g. "at 1:42 the chart jitters" or "the opening sounds abrupt". It maps timestamps to scenes and re-renders only what changed.

To collaborate, drop logos, screen recordings, photos, a music track or your own narration into the project. To finish in Premiere, Final Cut, DaVinci or 剪映, ask for the handoff export.

What's inside

skills/paper2video/
  SKILL.md              the workflow: intake → understanding → script → data → voice → scenes → fact-check → render → deliverables
  reference/            rigor, voice, visual style, platforms (YouTube + Bilibili), third-party papers from an
                        arXiv link (finding data, digitising figures, CPU illustrations), materials & editor handoff
  scripts/              voice-over pipeline (ElevenLabs / edge-tts / recorded / none) with ASR QA, music, SFX,
                        captions, mastering, vertical cut, handoff export, contact sheets, figure-data
                        extraction (vector PDFs, colormaps, digitising), final-file checks
  template/             Remotion starter: theme, captions, anchor-driven timeline, scene shell, vertical layout, covers

Principles

• Truth over hype. Every number traces to the paper. Illustrations are labelled, and so are selected examples.
• Real data early. The strongest real visual goes in the first seconds, followed by a primer so anyone can follow.
• Motion explains. Strokes draw on and beats are timed to the voice. There are no count-up numbers and no full-frame shake.
• The work is the pitch. The tools that made the video get a credit line, nothing more.

License

MIT for the skill, scripts and template. The showcase videos and images belong to their authors.