How to Prompt AI to Summarize a Meeting Transcript

A generic "summarize this meeting" prompt returns a paragraph that reads smoothly and buries the one decision that mattered inside a sentence about who joined late and what everyone had for lunch. The fix isn't a longer summary, it's a structure that forces decisions, action items, and open questions into separate, scannable sections, plus a follow-up check that catches what the first pass dropped.

What a generic summary gets wrong

Ask for a plain summary of a forty-minute transcript and you typically get something like this:

"The team discussed the Q3 roadmap and various options for the pricing update. There was general agreement to move forward with next steps, and a few open items were raised for follow-up."

This is accurate in the sense that nothing in it is false, and useless in the sense that it doesn't tell you what was decided, who's doing what, or what's still unresolved. "General agreement to move forward" is the sentence a generic summary writes when the model compressed a real decision into vagueness rather than losing information outright, because you never told it decisions were the thing to preserve.

The three-bucket structure

Ask for exactly these three sections instead of a narrative summary:

• Decisions made. Each one stated as a decision, not a discussion topic, with who made the call if the transcript says so.

• Action items. Task, owner, and deadline. If the transcript doesn't state an owner or deadline for something that sounds like a task, flag it as unassigned rather than silently guessing or dropping it.

• Open questions. Anything raised but not resolved. This is the section a narrative summary loses first, because unresolved threads don't compress into a tidy sentence.

The prompt

Read the meeting transcript below and produce exactly three sections: Decisions Made, Action Items, Open Questions.

Decisions Made: state each as a completed decision, not a topic discussed. Include who made the call if the transcript indicates it.

Action Items: format as Task, Owner, Deadline. If owner or deadline is missing from the transcript, write "not assigned" rather than guessing.

Open Questions: anything raised but not resolved by the end of the meeting.

Do not write a narrative summary. If a section has nothing to report, write "none" rather than inventing content to fill it.

The verification step

First drafts drop things, usually numbers and dates, because they compress less cleanly into bullet points than decisions do. Before trusting the output, run one more prompt against the same transcript:

Scan the original transcript again. List any specific numbers, dates, dollar amounts, or commitments mentioned that do not appear anywhere in the three sections above.

This catches the class of error where a number gets mentioned once in passing, in a sentence that didn't sound decision-shaped enough to make the first pass, and would otherwise silently vanish from the record.

A short worked example

Transcript excerpt: "...so I think we're good to launch the annual plan at $240 a year, that's down from what we floated last time. Sarah, can you get the pricing page updated by Thursday? ...still not sure if we're grandfathering existing monthly subscribers in at the old rate or not, need to loop in finance on that."

Decisions Made: Annual plan priced at $240/year. Action Items: Update pricing page, Owner: Sarah, Deadline: Thursday. Open Questions: Whether existing monthly subscribers are grandfathered at the old rate, pending finance input. Each piece landed in the section built for it, instead of dissolving into "pricing was discussed and next steps were assigned."

Handling messy real-world transcripts

Auto-generated transcripts rarely label speakers cleanly, and crosstalk gets mangled into run-on fragments. Tell the model explicitly to work with what it has: "speaker labels may be missing or unreliable, attribute statements to a role or description if a name isn't clear, and note where the transcript is garbled rather than guessing at the content." This produces a more honest output than one that confidently attributes a decision to the wrong person because the transcript's automatic labels drifted partway through. For a related technique on getting a clean, structured document out of an unstructured source, see our guide to prompting AI to write a design doc.

Frequently asked questions

What if the transcript doesn't label speakers at all?

Ask the model to use "Speaker A / Speaker B" or role-based labels ("the person who raised pricing") rather than guessing names. A wrong name attached to a decision is worse than no name, since it can send credit or follow-up to the wrong person.

How long can the transcript be before this approach stops working well?

Quality degrades gradually with length rather than failing at a hard cutoff, but past roughly an hour of dense discussion, split the transcript into logical segments (by agenda item, if there is one) and run the three-bucket prompt on each segment separately, then combine. This catches details a single pass over a very long transcript tends to compress away.

Should I use this for recorded calls or just text notes?

Either works as long as you have a text transcript to start from; the technique operates on the text, not the audio. Auto-generated call transcripts tend to be messier (see the speaker-labeling note above), so the verification step matters more there than on clean typed notes.

Can this replace someone taking notes live?

For the record-keeping part, largely yes. For the parts that need a human in the room, no: catching tone, sensing when someone's uncomfortable disagreeing out loud, and steering the discussion in real time still need a person. Use this for what happened afterward, not as a substitute for someone paying attention during the meeting.

For more on getting structured, useful output instead of generic filler, see our prompt engineering fundamentals.

This same distill-the-record instinct applies to other after-the-fact documents. See our guides to prompting AI to write a postmortem and prompting AI to write a runbook.