Each summary starts from a recorded lecture. AI agents built the video and, for some talks, a learning page. Dominik Lukeš set the goal, reviewed the storyboards and the finished videos, and sent corrections. Most of the rules described below are those corrections.
Recording and transcript
The lecture is recorded and transcribed. The slides are kept with the transcript.
Script
An agent writes a new script from the transcript, the slides and the lecture notes. It is not a cut of the recording. It is about a thousand words, five to seven minutes of narration, one connected account of the lecture. A plan comes first: who the viewer is, what they already know, the line of argument, and what is left out, such as housekeeping, asides and answers to questions from the room. Views that Dominik gave aloud and that are not on a slide are attributed to him ("Dominik says").
The agent then checks the script against the slides. Where the recording is garbled or unclear, the slides decide. A number on screen matches the chart it sits on. A revenue claim in the summary of the 20 October lecture was cut because the chart it rests on did not support it.
Voice
The narrator is a synthetic voice, Breeze TTS 2, the same in every video. Dominik chose it after listening to a set of auditions. A speech recogniser (Parakeet) transcribes each take to get word times, so that each picture arrives on the word that names it and the captions follow the speech.
Pictures and animation
Each sentence of the script gets an illustrated screen. The screens are described as data and drawn by code: line-art figures, product logos, and real screenshots and charts from the slides with rings and labels drawn over them. HyperFrames turns the result into video. An agent looks at every still before the video is rendered, and any picture that shows names, project names, file paths, booking references or faces is redrawn or masked. The storyboard, with every screen beside its narration, goes to Dominik before any voice or render.
Learning page
For some talks a learning page is built from the same outline, pictures and timings, so it matches the video. It has the video with chapters, the lecture's argument as a series of questions with a short answer to each, a step through the pictures, key terms and cards on applying the lecture. The application steps are marked as a translation of the lecture into practice, which the lecture does not itself state.
Review
The first attempt, a cut of Dominik's own recording over slide bullets, was rejected: "no white screen while someone is talking". His reviews since then have added rules:
- No slide titles on screen. Each video opens on a title page that works as its thumbnail: a short hook and one real picture from the lecture.
- Something changes on screen every two to three seconds, and never four seconds pass without a change. Narration runs at about 150 words a minute.
- Pictures show relationships. When a sentence says that one thing makes, changes or has another, the picture draws that with an arrow or tags, and items that belong together stay on one screen. This rule came from a screen on which Canvas sat on Amazon's view and its label alone on the next.
- Hedges stay in the picture. "Virtually no", "in theory" and "in many ways" appear as captions or arrow labels and do not become flat claims.
The agents wrote the scripts, drew the pictures, made the voice, rendered the videos, built the learning pages and ran the technical checks, including the checks of the narration against the slides. Dominik reviewed the storyboards and the videos.