Caveman Papers

Many great research papers never reach the people who would actually enjoy them. The format asks for too much attention before you know whether you care. Caveman Papers is an animation agent pipeline I built to change that: give it a PDF, and it produces a short vertical reel — something you can watch in 60 seconds and decide whether to go deeper.

The pipeline treats a paper as raw narrative material. An LLM reads the PDF, extracts the central argument and key result, then rewrites it as a compact visual story with a beginning, tension, and payoff. A scene generator breaks that story into timestamped segments and matches each one to a visual style. A narration agent writes voiceover copy calibrated to the reading level and pacing of short-form video. Everything gets assembled — background visuals, animated text, narration — and rendered into a portrait-format short.

The pipeline generates visuals in whatever style fits the paper — flat cartoon, chibi, 3D render — so each reel has its own character.

Subscribe and watch the channel here: youtube.com/@CavemanPapers

Below are two complete reels generated by the pipeline.

The design goal was not to summarize papers better. It was to solve a different problem: science communication as an interface problem. Most people’s first contact with a paper is either the abstract (too dense) or a Twitter thread (too lossy). A well-made reel can sit in between — cinematic enough to be watchable, accurate enough to be trustworthy, short enough to complete. The goal is to make the first step easier and earn the viewer’s attention before asking for it.