Minutes, transcript or chapters: which one to ask for

Four outputs, one ladder

Meeting Notes shipped able to produce one thing: notes. That is the right default and the wrong only option, because a summary answers a single question well and every other question not at all - the material those answers need is exactly what a summary throws away.

So the four outputs are a fidelity ladder rather than a menu of features. Notes keep the outcome, at a few per cent of the session. Full minutes keep the meeting: every point, argument and figure, in the order it happened, at roughly a third of the transcript's length. The transcript keeps the words, with the speech recognition's damage repaired. Chapters keep the shape - where in the session each subject lives.

Each rung discards less than the one above it. That is the whole distinction, and once you know what the document is for the choice is rarely close.

Which one do I actually want?

The useful test is not what kind of meeting it was. It is what you expect to do with the document after the meeting is cold.

The middle two are the ones people under-use. A colleague who missed a design review does not want your summary of it - a summary tells them where it landed and leaves them unable to reopen anything, because the reasoning that got it there is gone. Minutes are written so someone can read the session instead of listening to it: the points made and by whom where the transcript establishes it, the objections, the figures as given, with decisions and action items marked inline at the moment in the discussion where they happened rather than gathered at the end.

What belongs in a set of notes, section by section, is its own question. This one is about which document to ask for in the first place.

Why the transcript behaves differently from the other three

The transcript is the odd rung, in two ways that surprise people.

It is never translated. Every other output is written about the session, so it follows the language you set for the tab. The transcript is the session. Running it through a second language would quietly replace what people said with a paraphrase of what they said, and a record that has been through two languages is no longer evidence of anything. Each speaker stays in the language they spoke, even where that is two languages in one call.

It is also the one output whose length tracks its input. Everything else compresses, so a long meeting can be handed to the model in big pieces and the answer comes back in a small corner of the window. A stretch of transcript needs as much room to come back as it took to send, which is why the same meeting takes more passes as a transcript than as minutes.

What the clean-up changes is narrow and worth being precise about: punctuation, capitalization and sentence boundaries, which speech recognition supplies almost none of and which are most of what makes a raw transcript unreadable; filler and false starts; obvious mis-hearings where the intended word is unambiguous, names especially. What it never changes is the substance - no turn is summarized, dropped as repetitive or skipped as unimportant - and it never changes the speaker. The instruction is to fix the transcription, not the person: hesitancy and phrasing that are theirs stay theirs. Where the audio is past repair it says "(unclear)" instead of guessing.

Chapters are the output that exists because of the timestamps

Every line in the transcript carries the time it was said. Without that there is nothing for a chapter to point at, and "where did we talk about the contract?" has no answer at all.

A chapter is a subject, not a fixed length: twenty minutes on one thing is one chapter, a rapid run through five items is five short ones. The headings say what a stretch was concretely about - "Pricing for the enterprise tier", not "Discussion". The start and end times are taken from the line markers rather than estimated, and the chapters cover the session end to end with no gaps, opening chat and wrap-up included, briefly.

They are also the cheapest way to make a long recording usable at all: on a two-hour call, the first thing worth asking for, and what tells you which stretch deserves a transcript.

Do I have to decide before I press Stop?

You do not have to, but it is worth doing. The Meeting output setting says what this meeting is for, and stopping the recording produces whatever it is set to. Set a meeting to full minutes and Stop gives you minutes - not notes and then a second click. The button and Stop produce the same thing, so there is one answer to "what do I get from this meeting" rather than two.

Everything else stays one click away on the button's own menu, and choosing from that menu runs immediately rather than only changing a preference: someone opening it wants the document, not a setting. Nothing is lost by picking wrong. Each output lands as its own result in the session, so a meeting can end up holding all four, browsable back and forth, each labelled with what it is, when it was made and whether it covers the whole session.

What each one costs before you ask for it

All four cost model time, which is why none of them runs on its own. A session longer than the model's context window is read part by part, and the menu says how many passes each output would take right now, next to its name, before you commit to it.

That number is not decoration. On the built-in on-device model a nine-pass read is minutes of waiting, and the fix - a model with a larger context window - is one menu away. The pass counts also differ between outputs on the same meeting: the transcript is divided into smaller parts than the others, because of the length-preserving problem above, so a long session takes more of them; notes take one extra pass beyond their parts, because the parts are read and then merged into a single set of notes. Minutes, transcript and chapters are joined rather than merged - each part covers a consecutive stretch and the finished document is chronological anyway - which is also why they degrade gently on a small window. More parts, same document.

If a session is longer than even a part-by-part read can cover, the result says so in its own title: it is labelled as the most recent stretch rather than presented as the whole call. Check that label before forwarding anything from a long meeting.

What if the document is nearly right?

Notes and minutes can be reshaped after the fact - cut to the action items alone, reorganised by topic, turned into a recap email you can send round, or, for a lecture, into key terms or something to revise from.

Those rewrites are deliberately not offered on the transcript or the chapters. Every one of them compresses or reorganises, and a compressed transcript is simply no longer the record of what was said; the chapters have the same problem in miniature, being already the shortest true thing about the session. Both still take typed questions, which is what people actually want from them anyway.

Anything can also be fixed by hand - a misheard name, a line nobody agreed to - without regenerating the nine-tenths that were right. All four export to Markdown with the date, the participants and the model that wrote them, and an exported file says when a human edited it.

The short version

Ask for notes by default. Ask for minutes when someone who was not there needs to follow the argument rather than learn the conclusion. Ask for the transcript when the exact words matter. Ask for chapters when the recording is long enough that the problem is finding things in it.

None of this needs a bot in the call, and none of it leaves your Mac unless you point the tab at a cloud provider - what that changes is set out on the security page, and how a recorded call works from the start is on Meeting Notes. If you are wondering why none of these four appear while the call is still running, that decision has its own reasoning.