How to get something useful out of a meeting transcript
Working with documentsLast checked
Short answer
Do not ask for a summary. Ask for decisions, commitments with owners, and open questions. A meeting summary describes what was discussed, which nobody needs a week later. What people actually want is what was decided and who owes what, and those are different requests.
First, get a transcript
Audio is not among the file types OpenAI lists for uploads, so the recording itself is not the route.
Most meeting tools now produce transcripts automatically. Failing that, any transcription tool handles an hour of audio in a few minutes. What you want out of it is a text file, ideally with speaker labels and timestamps.
Keep the speaker labels
This is the single biggest quality difference. Without labels you get "the team agreed to look at pricing". With them you get "Sam agreed to look at pricing before Friday". One is a note, the other is an action with a name on it.
Ask for the right four things
Instead of a summary:
From this meeting transcript, give me four sections.
Decisions: what was actually decided, with who decided it. Commitments: who agreed to do what, by when if a date was mentioned. Say "no date given" where there is not one. Open questions: what was raised and left unresolved. Disagreements: where people did not agree, and what each position was.
Quote the line each item came from. If something is not present, say so rather than inferring it.
That structure works because it maps to what a meeting is for. The "no date given" instruction is worth keeping: an undated commitment is a finding, and without permission to say so it tends to get a plausible date attached.
The questions worth asking after that
What did everyone avoid?
What was raised and then dropped without resolution?
Meetings are full of these and nobody writes them down, because they are absences.
Who actually committed to something?
List every person who took on an action, and everyone present who took on nothing.
Uncomfortable and useful in equal measure.
Was anything decided twice?
Does this transcript revisit any decision more than once, and did the answer change?
What would someone who missed this need to know?
Good framing, because it forces relevance over completeness.
Handling a long transcript
An hour of speech is roughly ten thousand words, which goes in comfortably. A full day, or a series of meetings, can pass 2 million tokens and get truncated with no warning.
The check is the same as for any long document. Ask what the last thing discussed was. If it cannot say, it did not reach the end, and any list of commitments is missing whatever was agreed in the last twenty minutes, which is usually where commitments happen.
Across several meetings
Process each one separately into the same four-section structure, then bring the structured notes together:
Here are structured notes from six weekly meetings. Which commitments were made and never mentioned again? Which topics keep returning without resolution?
That second question is the valuable one. A topic appearing in five consecutive meetings without a decision is a real signal about the project, and it is invisible from inside any single meeting.
What transcripts lose
Worth knowing before you rely on one.
Tone, hesitation, and who sounded unconvinced. A transcript records that someone said yes. It does not record that they said it flatly after a long pause, which everyone in the room understood perfectly.
For content and commitments, a transcript is better than the recording: searchable, quotable, and checkable. For reading the room, it is not, and no prompt recovers what was never written down.
Common questions
Can I upload the meeting recording directly?
No. Audio is not among the file types OpenAI lists for uploads. Get a transcript first, which most meeting tools now produce automatically, and upload that as text.
Why are the summaries so useless?
Because "summarise this meeting" asks for a description of what was discussed, which nobody needs. Ask for decisions taken, commitments made and open questions instead, and you get something you can act on.
Should I keep the speaker labels?
Yes, always. They turn "the team agreed to review pricing" into "Sam agreed to review pricing by Friday", which is the difference between a note and an action.
A week of meetings is too long to upload. What now?
Process each meeting separately into the same structure, then bring those structured notes together for the cross-meeting questions. Uploading a week of raw transcripts at once blends them and loses which meeting anything came from.
Keep reading
Can you upload an MP3 or WAV file to ChatGPT?
No. OpenAI lists 4 upload families: text, spreadsheets, presentations, documents. Audio is in none of them. Transcribe it, and keep the speaker labels.
How to summarise long documents with ChatGPT
Anything past 2 million tokens is cut off silently, so run the last section test before you read the summary. Then the prompts that stop it coming back bland.
How to pull specific data out of a document with ChatGPT
Name every field, name the format, demand the source: "Return a CSV with columns date, party, amount". The 3 rules that stop ChatGPT filling gaps for you.
Can ChatGPT read SRT and VTT subtitle files?
Yes. SRT and VTT are plain text and upload like any other file. Strip the timestamps on anything feature length, since they take more room than the dialogue.
How to summarise a YouTube video with ChatGPT in 4 steps
ChatGPT cannot open a YouTube link or the video file, so use the transcript. 4 steps to pull it out and paste it in, plus what a text only view misses.