Learning the rules of a board game can be a challenging and sometimes confusing process. While players often develop the strongest understanding through hands-on experience, verbal explanations alone frequently lead to misunderstandings or overlooked details. In contrast, observing an actual game in progress can significantly improve comprehension by providing concrete contextual cues. This thesis investigates how recordings of board game sessions can be transformed into data videos—concise, visually enriched summaries that emphasize key moments, player interactions, and strategic decisions. These videos serve two purposes: they function both as engaging tutorial material for novice players and as analytical tools for post-game reflection and strategic analysis. Furthermore, this work explores the potential and limitations of modern Large Language Models (LLMs) in supporting the semi-automated generation of such data videos. To evaluate their effectiveness, we conducted quantitative experiments to compare the predicted events that are to be shown in the video with predefined ground truth as well as a user study examining how data videos influence players’ understanding of the game and their overall engagement. We were able to achieve good recall and precision in event prediction for the two games Coup and Poker which we used as reference recordings. We identified several limitations, most notably that the quality and usefulness of the event prediction strongly depend on the quality of the generated transcripts. In addition, the formalization of visual cues must be carefully designed to match the respective game situations as accurately as possible. The generated data videos were helpful even without visual annotations and helped in understanding the course of the game.