A meeting minutes AI prompt that cites its sources

"Summarise this meeting" is the most common prompt for meeting notes, and the reason so many AI notes can't be trusted. It asks the model for a fluent paragraph, and a fluent paragraph is exactly where a "maybe" turns into a decision and a task lands on the wrong person.

A better prompt asks for something different: minutes where every line points back to the moment it came from, and where the model says so when it isn't sure. Below is the prompt we recommend, why each part is there, and how to check what comes back. If you are new to formal minutes, start with how to write meeting minutes with AI, which covers what minutes must contain. Not sure whether you need minutes or just notes? How minutes differ from notes and a transcript comes down to who approves them and who relies on them later.

What you need first

A transcript with timestamps and speaker names on each line, like this:

[00:41] Sara: Let's lock it in: the Arabic onboarding ships this Thursday.
[01:02] Sara: Omar, can the pricing deck reach the client before Monday?
[01:20] Omar: Yes. I'll send it Sunday night so they have it first thing.

Without timestamps the model has nothing to cite. Without names it will guess who said what.

The prompt

You are writing formal meeting minutes from the transcript below.

Rules:
1. Use only what is in the transcript. Do not add facts, names, numbers or dates
   that were not said.
2. After every decision and every action item, add the timestamp of the line it
   comes from, in square brackets, for example [01:20]. If it comes from several
   lines, give the first one.
3. Keep hedges. If someone said "probably", "maybe" or "if", the minutes must say
   so. Do not turn a suggestion or a question into a decision.
4. An action item's owner is the person who said they would do it. If nobody
   took it on, write "Owner: not agreed".
5. Write dates exactly as said, and add "(relative to the meeting date)" when a
   date is relative, such as "next Thursday".
6. If something is unclear, list it under "To confirm" instead of guessing.
7. Write in the language the meeting was held in.

Output, in this order:
- Attendees (names that appear in the transcript)
- For each topic: a two-line summary, then "Decision:" with its timestamp
- Action items: action, owner, due date, timestamp
- To confirm

Transcript:
[paste the transcript here]

Why each rule is there

• Rule 1 stops the model filling gaps with what usually happens in meetings like yours.
• Rule 2 is the point of the whole prompt: every line people will act on becomes checkable in seconds.
• Rules 3 and 4 target the two most common errors in AI minutes: softened commitments that come back hard, and tasks assigned to whoever spoke most.
• Rule 5 flags relative dates so you resolve them against the calendar before sending.
• Rule 6 gives the model a place to put doubt. Without it, doubt comes out as confident prose.
• Rule 7 keeps notes close to what was said. Translate afterwards, if you need to.

How to check the result

Open the transcript next to the minutes and, for each decision and action item, jump to its timestamp. Read the line. Does it say what the minutes say, with the same hedges, owner and numbers? Our checklist for AI meeting summary accuracy walks through the eight things to look for.

Two limits to keep in mind. A chat model can still cite the wrong line, so the check is not optional. And very long transcripts may not fit in one prompt; split them by topic or by half hour and merge the minutes afterwards.

Or skip the prompt

MeetNotely does this as part of every meeting: it records from your browser with no bot joining, writes the transcript with speaker names, and links each summary line, decision and action item to the second it was said, so a timestamp is one click from the recording. For turning that into formal minutes, see our guide to writing meeting minutes with AI. Every feature in detail covers Ask, search and share links, and pricing starts with a free plan.