Record, and get it back readable
Hit record and leave it. Transcription starts while the meeting is still running rather than waiting for you to stop, so the transcript is usually there the same day instead of the next morning. It is not a live caption and does not pretend to be one: the recorder bar tells you plainly which state you are in, whether that is recording, transcribing, drafting, or waiting because the machine that does the work is not switched on. Recording never depends on that machine being up, so audio is never lost while you wait for it.
A big model, because these meetings are hard
A student engineering meeting is close to the worst case for speech recognition: several people, one microphone in a room that echoes, mixed first languages, and vocabulary no general model has any prior for. Nobody outside the team says ullage, or throat diameter, or reads out a specification key. The small fast models people benchmark to prove speech recognition is cheap mangle all four, and a summary of a mangled transcript is worse than no summary, because it reads as confident. So this runs a full-size model, and the notes are drafted by a model large enough not to write fluent sentences the transcript does not support.
The notes come out structured
Not a wall of summary. An agenda in the order the meeting actually went, the decisions, and the actions, where every action carries a kind and whatever name the transcript said. The kind matters because it is what lets an action become the real thing later: a task, a specification revision, a component change, a follow-up meeting. Every one of them is a proposal until somebody accepts it, so nothing is created behind your back. And they land as ordinary editable notes on the project rather than a locked output you copy somewhere useful.
Several recordings, one summary
Somebody stops the recording early, a tab crashes, the room moves. You end up with two or three takes of what was really one conversation. Prototype treats those as parts of a single meeting and produces one transcript and one summary from all of them, rather than a summary per recording that you then have to reconcile. If you have edited a summary by hand, a later take will not overwrite your edits.
The audio stays on your hardware
Transcription and summarisation run on a machine your team runs, not on a third-party transcription service. The recording of your propulsion review does not become somebody else's training data. That machine is scoped to your organisation and to the job it was given, so it cannot read the rest of your project data, and the box doing the transcribing is not a way into everything else.
Transcripts inherit the meeting's access list
A transcript is only visible to the people the meeting was visible to. If a meeting is restricted to a role group or a named set of people, the transcript and the summary are restricted the same way, and they do not turn up in search for anybody else. This matters more than it sounds: a recording is a far more complete record than notes, so a meeting that was safe to hold becomes unsafe to record if the transcript reaches further than the room did.
Pasting in a transcript you already have
Anything captured elsewhere can be dropped in and summarised the same way, which matters for a team part-way through a semester with a term of meeting notes already sitting in another tool.