Capstone: from audio meeting to minutes

Collect everything in one stream: from the raw audio of the meeting to the finished protocol that can be shared. Pipeline of four links. Entrance - recording: one good microphone, smooth sound and at the beginning the spoken consent of the participants. Transcript - a transcript with a glossary of names and terms, markings by speaker, and markings of uncertain places. Summary - solutions, tasks with responsibilities and deadlines, open questions, with a volume limit are extracted from the transcript. Issue - a short protocol in an understandable form plus a separate list of tasks. The key to the capstone is coherence: the output of each link becomes the meaningful input of the next, and at each junction you set the format explicitly, rather than relying on guessing. Pro trick: run the protocol back through the check - ask the model to check each task and figure with the transcript and mark discrepancies. The second technique: build privacy into the process itself, and not at the end - consent at the beginning of the recording and cleaning up sensitive things before forwarding. A common mistake is to assume that smooth text is true; in important documents there is always a step of human verification of decisions and numbers.