How to Write Meeting Minutes Using AI That People Actually Read

Why nobody reads your meeting minutes

The problem is not effort. It is that most minutes are optimised for the person who wrote them rather than the person receiving them. Atlassian’s survey of 5,000 knowledge workers across four continents found meetings are rated ineffective roughly 72% of the time, and one of the causes they name is people leaving with no clarity on what happens next. Minutes are supposed to be the fix for that. Usually they repeat the problem in written form.

AI made this worse before it made it better. Ask any assistant to summarise a transcript and it will do exactly that: compress an hour of talking into six paragraphs of talking. Shorter, still unreadable.

Same meeting. The second version is shorter, and a person who was not in the room can act on it. The first one asks them to do the interpretive work themselves, which is precisely the work they will not do.

The 90-Second Minutes Standard

Here is the test we use at PromptPeakAI, and the one this whole article hangs on. The 90-Second Minutes Standard: your minutes are finished only when somebody who missed the meeting can read them in ninety seconds and correctly answer what was decided, what changed for them, and what they owe by when.

It is deliberately a reader test, not a writer test. You cannot pass it by working harder on the document. You pass it by cutting everything that does not answer one of the three questions, which is usually most of what a transcript contains. Every instruction further down this page exists to hit that standard, so when you write meeting minutes using AI you are aiming at a measurable target rather than a vague sense of tidiness.

Most guides stop at “be concise”. Concision is not the point. A one-line summary is concise and equally useless. The point is completeness against those three questions, and ruthlessness about everything else.

What belongs on the page

Good minutes have a fixed shape, and the shape does the work. Once your team recognises the layout, they stop reading top to bottom and start jumping straight to the part that concerns them. That jump is what ninety seconds buys you, and it is the reason the shape has to be decided before you write meeting minutes using AI rather than left for the assistant to choose.

Microsoft’s own guidance on writing meeting minutes in Word puts a typical one-hour meeting at somewhere between half a page and two pages, and draws a useful line between minutes and notes: notes are the scratch record the note-taker keeps for themselves, minutes are the formal account that goes out to everyone else. If your minutes read like notes, that is the whole diagnosis.

The hardest part is the cutting, so here is the rule in table form. When you are unsure about a line, find its row.

One caveat worth naming: formal governance settings play by different rules. Board minutes, works councils, and regulated committee records often have statutory content requirements, and this article is not that. Check your obligations before trimming anything in those rooms.

How to write meeting minutes using AI

The workflow below works with whatever assistant you already have open, because the structure lives in your prompt rather than in the tool. If you are starting from a recorded call, our guide to turning a meeting transcript into actionable minutes covers the extraction mechanics in more depth, and choosing an AI meeting notetaker covers where that transcript comes from in the first place. This section is about turning either input into a document people read.

Three steps, and the first one is the step almost everyone skips.

Almost every bad AI summary traces back to a missing reader. The assistant does not know whether it is writing for the three people who attended or the twelve who did not, so it hedges and includes everything. Tell it who opens the document and what they need from it, and half the bloat disappears without any further instruction.

“Summarise this” is an invitation to compress. You do not want compression, you want extraction into a fixed shape. Specify the four sections and the exact format of an action line, and you get the same document every week rather than a fresh interpretation each time.

You are writing the minutes for a meeting I ran. They will be read by people who were NOT in the room, most of whom will spend 90 seconds on it.

Here are my inputs:
[paste your transcript or rough notes here]

Produce exactly these four sections and nothing else:

1. HEADER — meeting name, date, attendees.

2. DECISIONS — only choices the group actually settled. One line each, written as a statement of what is now true. Name who approved it. If nothing was decided, write "No decisions reached."

3. ACTIONS — a table with columns: Task | Owner | Due date. One row per commitment.

4. OPEN QUESTIONS — anything raised and left unresolved, with who is chasing it.

Rules you must follow:
- Use ONLY what is explicitly in my inputs. Do not infer, extrapolate, or fill gaps.
- If a task has no clearly named owner, put "UNASSIGNED" in the Owner column. Never guess.
- If a due date was not stated out loud, put "NO DATE SET". Never invent one.
- Ignore greetings, small talk, tech check chatter, and any thread that was raised and dropped.
- No preamble, no closing summary, no commentary about the meeting itself.

The four constraint lines at the bottom are doing most of the work here. Explicitly telling a model what not to produce is consistently more effective than describing what you want, which is why negative constraints are the technique we lean on hardest across the site. If the structure of that prompt feels familiar, it is the Role, Goal, Context, Format pattern with the format half made unusually specific.

The first output will usually contain at least one action with UNASSIGNED or NO DATE SET in it. That is the prompt working correctly, not failing. Those gaps are real: they existed in the meeting and everyone walked out pretending otherwise. Surfacing them is the single highest-value thing this workflow does.

Look at the ACTIONS table you just produced.

List every row where the Owner is UNASSIGNED or the Due date is NO DATE SET.

For each one, write me a single short question I can send in chat to close the gap. Address it to the most likely person based ONLY on who spoke about that topic in my inputs, and say explicitly that you are guessing so I can check it.

When there is no recording

Plenty of meetings never get recorded. IT blocks the bots, the client says no, or it was a corridor conversation that turned into three commitments. This is the case the other guides skip entirely, and it is more common than the recorded one.

The method barely changes. What changes is your input quality, so the work shifts to the five minutes right after the meeting while you can still remember it.

The second set is not longer by much. It just contains the two things AI cannot manufacture: who committed, and what got settled. Dump that into the same core prompt and it produces the same document a transcript would.

Below are the scrappy notes I typed during a meeting. They are incomplete and written in shorthand.

[paste your rough notes here]

First, list back to me every point where you cannot tell WHO owns something or WHAT was actually decided. Number them as questions.

Do NOT write the minutes yet. Do not guess at any answers. I will fill the gaps first, then ask you to produce the document.

Asking for the gaps before the document is the small move that makes this reliable. It stops the assistant smoothing over your shorthand with plausible invention, and it takes about ninety seconds to answer. Microsoft Copilot handles this well inside Word if that is where your notes already live, which our guide to using Copilot in Word walks through.

Sending minutes people actually open

A perfect document sent badly still goes unread. Delivery is part of the job, and it is the step most people treat as an afterthought once they have learned to write meeting minutes using AI and assume the hard part is over.

• Send within the hour. Interest in a meeting decays fast, and after a day your minutes compete with everything that happened since.
• Put the decisions in the message body, not in an attachment. An attachment is a request to open a second thing, which is a request most people decline.
• Name people in the actions. A reader scanning for their own name will find it in about two seconds. A reader scanning for “the team” finds nothing and stops.
• Use the subject line as the headline. “Notes from Tuesday’s sync” tells nobody anything. “Phase 2 approved, starts October — 3 actions” does.

If your minutes travel as an email rather than a document, the format shifts slightly, and our guide to post-meeting action summaries covers that shorter version. The agenda matters here too: minutes written against a real agenda are dramatically easier to structure, which is one of the quieter arguments for writing a proper meeting agenda in the first place.

Take the minutes you just produced and give me the message that carries them.

Write:
1. A subject line under 12 words that states the single most consequential outcome, not the meeting name.
2. A two-sentence opening that tells a non-attendee whether they need to read further.
3. Then the DECISIONS and ACTIONS sections, unchanged.

Do not add pleasantries, do not thank anyone for their time, and do not summarise the summary at the end.

Before any of that goes out, run the review gate. It takes under two minutes and it is the difference between a document that builds your credibility and one that quietly damages it.

When AI minutes go wrong

Four failures account for nearly everything that goes wrong here, and each has a specific fix rather than a general “review carefully”.

The name key in row four is a thirty-second fix that people rarely think of. Before your transcript, simply write: Speaker 1 is Priya, product. Speaker 2 is Marcus, operations. Everything downstream improves immediately.

When an output has already gone wrong, do not re-run the original prompt from scratch. Audit the draft you have instead — it is faster, and it shows you which rule was missing.

Here are the minutes you produced, and below them my original inputs.

[paste the draft minutes, then the transcript or notes]

Go through the draft line by line and flag anything that does not appear in my inputs. Specifically list:
- Any owner name that was not explicitly attached to that task
- Any date that was not spoken aloud
- Any decision where nobody actually approved it

Show each flagged line and quote the part of my inputs you based it on. If there is nothing to quote, say so plainly.

Invented owners and dates are the failure to watch hardest, because they are confident, plausible, and quietly corrosive to trust. This is ordinary model behaviour rather than a bug, and our explainer on why AI invents things covers the mechanism. When something in the output looks specific but unfamiliar, treat it the way you would any unsourced claim and apply a proper verification pass against your own notes.

What not to paste in

Meeting content is some of the most sensitive material in a business: performance conversations, commercial terms, legal exposure, unannounced changes. Before this becomes a weekly habit, decide what actually goes into the chat box.

One correction worth making, because it trips people up constantly: paying for a personal plan does not change how your inputs are treated for training. On the consumer tiers, the default is the same whether you pay or not, with an opt-out you have to switch on yourself. Contractual guarantees that your content is never used for training come with the business and enterprise products, not with a personal upgrade. Our guide to using AI safely at work goes through the settings in detail.

If your organisation has already licensed something with commercial data protection, use that for meeting content and use it consistently. Most compliance problems here come from mixing tools rather than from choosing the wrong one.

What AI cannot do here

Being straight about the limits is what keeps this workflow trustworthy over months rather than weeks. There are four jobs you cannot hand over, however well you write meeting minutes using AI.

There is also a limit worth naming on the accountability claim itself. Minutes make commitments visible; they do not make people do things. If the same action rolls forward for four weeks, that is a management conversation, and no prompt will substitute for it. Teams that want the follow-through without the meeting should look at restructuring the meeting itself rather than optimising the write-up.

Start with the next meeting on your calendar. Write meeting minutes using AI with the core prompt exactly as it is, send them inside the hour, and see who replies. In practice, the reply rate is the only measure of these things that has ever mattered.