This article provides rich insights into the process of data generation for discourse analysis from three separate studies of the video recordings of a single science classroom in action. The central claim is that multimodal transcription can contribute to developments in discourse theory. A three-stage reflective heuristic is developed and used in the article to support meta-methodological discussion on different researchers' negotiations with the complexity of the video data. The focus is how the different researchers attended to modalities of meaning making (e.g. speech, learning artefacts, whiteboard notes, gestures, bodily actions) and appropriated, adapted and transformed their theoretical framework in order to construct the transcripts for each study. The three-stage heuristic is shown to facilitate transparency in analytic decision-making and is recommended for promoting much needed discussion on processes of data generation for discourse analysis that draws upon video recordings of action in situ.