In short
A good prompt for a meeting summary asks for decisions, action items with an owner and a date, and open questions, tells the model to use only the transcript, and to say "none" when nobody committed to anything. Remove anything said off the record before you paste.
A good prompt for a meeting summary does four things: it names what you want back (decisions, action items with an owner and a date, open questions), tells the model to use only the transcript, asks for a short quote beside each item so you can check it, and tells the model to write "none" when nobody committed to anything. That last instruction matters most. Language models tend to produce what meeting notes usually contain, and most meeting notes contain action items.
This guide gives copyable prompts for common kinds of meeting note, and covers what to do before you paste a transcript into a chatbot.
Summarise a transcript with a prompt
- Check you are allowed to paste it. A transcript holds other people's words. Check your organisation's policy on AI tools and the chatbot's own data settings, and make sure the people in the meeting knew it was recorded.
- Remove what should not leave. Delete anything said off the record, health details, passwords, and anything a client asked you to keep out of writing.
- Put names on the speakers. Replace "Speaker 1" and "Speaker 2" with names where you know them. A model cannot give a task to the right person if it does not know who said what. Naming the speakers in a meeting transcript shows how.
- Paste the prompt, then the transcript. Use one of the prompts below. Put the transcript after the instructions, clearly marked.
- Check every item against the transcript. Use the quotes to find each one. Delete anything the transcript does not support.
- Label it before you share it. Mark the result as AI-written and note what you checked. How to label AI-written notes you share has wording.
A general summary
You are summarising a meeting from its transcript, which follows.
Use only what the transcript says. Do not add background knowledge.
Write:
1. Three to five sentences on what the meeting was about and how it ended.
2. Decisions: each decision actually made, with a short quote.
If none were made, write "No decisions were made."
3. Action items: only tasks a named person explicitly agreed to do.
For each: owner, task, due date if one was said, and a short quote.
If nobody committed to anything, write "Nobody committed to anything."
4. Open questions: things raised and not settled.
If something is unclear in the transcript, say so rather than guessing.
TRANSCRIPT:Action items only
From the transcript below, list only the tasks that a specific person
said they would do, or agreed to when asked.
For each: owner | task | due date as said (or "no date") | quote.
Do not include suggestions, ideas or things "we should" do unless
someone took them on. If there are none, write "No action items."
TRANSCRIPT:What makes an action item worth writing down is covered in writing meeting action items that get done.
Decisions and disagreement
From the transcript below:
1. List each decision made, who made or agreed it, and a quote.
2. List each point where people disagreed: what each side said,
and whether it was resolved. Include disagreement that was
dropped rather than settled.
Use only the transcript. If there was no disagreement, say so.
TRANSCRIPT:Summaries tend to smooth over dissent. Asking for it by name is the simplest counter.
A follow-up email
Draft a short follow-up email to [name] after the meeting below.
Confirm what was agreed and by whom, and ask for anything still open.
Only include commitments that appear in the transcript. Do not
invent dates, prices or numbers. Leave [brackets] where a detail
is missing.
TRANSCRIPT:Minutes
Write minutes of the meeting below, from the transcript only.
For each topic, in the order discussed:
- the topic in a few words, and roughly when it started;
- what was said, in two or three sentences, naming who said it;
- any decision, stated exactly, or "no decision".
End with a list of action items (owner, task, date as said).
Do not add anything that is not in the transcript.
TRANSCRIPT:Minutes are a more formal record than a summary. Meeting minutes vs notes explains the difference, and formal minutes of a board or committee need a person to approve them whoever drafted them.
Adjusting a prompt to the meeting
The general prompt works for most meetings. A line or two added for the kind of meeting helps:
| Meeting | Add to the prompt |
|---|---|
| One-to-one | "Keep personal matters out of the summary." |
| Client call | "List what the client asked for, in their words." |
| Interview | "Do not rate or judge the candidate; list what they said about each topic." |
| Planning | "List what was left out of scope, and why." |
| Retrospective | "Group points under what went well, what did not, and what we will change." |
Why these prompts are written this way
"Use only the transcript." Without it, a model may fill gaps with general knowledge: a typical timeline, a common price, the usual next step. That text reads well and was never said.
"Write none." A model asked for action items will usually produce some. Giving it an explicit way to say there were none removes the pressure to invent them.
Quotes. A quote beside each item turns checking into a search rather than a re-read. If the quote is not in the transcript, neither is the item.
Names. "Speaker 2 will send the deck" is only useful if you know who Speaker 2 is, and a model guessing names from context can guess wrong.
Even with a careful prompt, check the output. A prompt reduces the patterns can you trust an AI meeting summary? describes; it does not remove them. Wrong numbers and names misheard by the recogniser pass straight through any prompt, because the model cannot hear the audio.
Long meetings
Models read a limited amount of text at once, and the limit varies. A two-hour transcript may be too long, or the model may pay less attention to the middle. Two approaches work:
- Summarise in parts. Split the transcript into sections of about half an hour, summarise each with the same prompt, then ask for one summary of the summaries.
- Ask narrower questions. "What did we decide about the launch date?" needs less of the transcript to be right than "summarise everything".
Before you paste: privacy
Pasting a transcript into a general-purpose chatbot sends it to that chatbot's provider. What happens next depends on the service, the plan you are on and its settings: how long conversations are kept, whether they are used to improve models, and who at your organisation can see them. Read the provider's data controls and your organisation's policy. Does your AI notetaker train on your meetings? lists the questions, and they apply to chatbots too.
The audio is a separate matter. Pasting text sends text; the recording stays wherever it is.
How Notey fits in
Notey writes a fixed set of notes from a meeting's transcript: a summary with action items, minutes, tough questions, where you differed, and a follow-up email. Two of the behaviours above are built in: the summary says so when nobody committed to anything, and Ask Notey says when the transcript does not cover a question. Everything it writes is labelled AI-generated. There are no custom prompts and no choice of model, so if you want your own prompt, use another tool.
To do that, export the transcript from a meeting as plain text and paste it wherever you like, with the checks above. Only transcript text leaves your Mac either way; the recording stays on your disk. How AI meeting notes are made explains the whole chain.
Frequently asked questions
What is the best ChatGPT prompt for meeting notes?
One that names what you want (decisions, action items with owners and dates, open questions), limits the model to the transcript you pasted, and tells it to write "none" rather than guess. Asking for a short quote beside each item makes checking faster.
Is it safe to paste a meeting transcript into a chatbot?
It depends on the service, your plan and its settings, and on your organisation's rules. A transcript contains other people's words and often personal data. Check the chatbot's data controls and your policy first, and remove anything said off the record.
How long a transcript can I paste?
Models have a limit on how much text they read at once, and it varies by model. For a long meeting, summarise it in parts and then ask for one summary of the parts, or use a tool built for meetings.
Why does the summary include an action item nobody agreed to?
Because the model writes what is likely, and meetings usually end with action items. Tell it to list only tasks someone explicitly agreed to, with a quote, and to say "none" otherwise. Then check each item against the transcript.
Should I ask the AI to judge a candidate or a colleague from a transcript?
No. Ask it to list what was said, and make judgements yourself. A summary that rates people from a transcript can be unfair and, in hiring and employment, can raise legal questions.
Do these prompts work with any AI chatbot?
They are plain instructions, so they work with most general-purpose chatbots. Results vary between models, so check the output the same way whichever you use.