I made this video using AI agents, which built the graphics and the editing pipeline from my script, my recordings and the paper's tables. It was a process of collecting information on AI editing, prompting, making editing choices, and iterating on the final product.
The graphics were made with HyperFrames (open source, Apache-2.0, github.com/heygen-com/hyperframes), which comes with skills that teach a coding agent such as Claude Code how to build and check videos. Some skills that I went through when learing how to do this are listed below, in case they are useful for your own research video. The skills are not required to use HyperFrames, but they can help you make a video faster and with fewer mistakes:
hyperframes: the core skill. A video is written as an HTML page with GSAP animation, so charts, maps and equations can be built from your own data and timed to your voice.hyperframes-cli: checks every frame for text running off the screen or overlapping, takes snapshots, and renders the MP4.media-use: transcripts, captions, music and colour grading. I transcribed locally with faster-whisper and used my own recording instead.hyperframes-registry: about 450 ready-made blocks (titles, captions, charts, map overlays), a very quick start if you don't want to design from scratch.slideshow: turns a talk or deck into an animated slideshow, handy for conference presentations.style-sampling: ready-made explainer styles, including an animated science explainer.figma: brings figures and designs from Figma into a video.It is worth emphasizing that human taste and judgment were essential in the process. The script benefited from advice on my Job Market Spiel from academics, especially Jaime Millán-Quijano, and the graphics were improved following comments from Jenifer Ruiz-Valenzuela. Hong Zhumi contributed his taste of music, and Lin Nan, a video editor and designer, offered professional feedback.
The paper