How to Write AI Prompts for Meeting Notes That Actually Capture Decisions

Most people ask an AI note-taking tool, or a pasted transcript, to "summarize this meeting" and get back something that reads well and helps almost nobody a week later: a tidy paragraph describing what was generally discussed, with no clear record of who agreed to do what, by when, or what was actually decided versus merely floated as an idea. A meeting summary that cannot answer "what do I owe by Friday" has failed at the one job a meeting summary actually has, and the fix is not a longer or more polished paragraph. It is a completely different prompt structure, built around decisions and owners instead of topics and chronology.

Why "summarize this meeting" produces the wrong kind of summary

A model asked to summarize will default to a narrative recap organized by conversation flow: what came up first, what came up next, roughly in the order it was said. That shape works fine for catching up someone who missed the meeting entirely, but it is close to useless for the people who were in the room and need to know what happens next. Decisions, owners, and deadlines are often only implicit in a transcript, mentioned once in passing ("yeah, I can have that by Wednesday") and never restated, so a generic summarization prompt tends to compress that detail into vague language like "the team discussed next steps" instead of surfacing it as an actual commitment attached to a name and a date.

The four things a meeting-notes prompt should ask for by name

A one-line summary at the top: a single sentence stating the overall outcome of the meeting, written for someone who will only ever read that one line and nothing else. Decisions made: what was actually decided, phrased as a completed choice ("the team will ship the redesign in two phases") rather than a description of a discussion that happened. Action items with an owner and a deadline: every task needs a name attached and a date, and if the transcript never states one explicitly, the model should say so rather than quietly leaving it blank or inventing one. Open questions: anything raised but left unresolved, so it does not silently disappear between this meeting and the next.

A prompt structure that gets there

Rather than "summarize this transcript," ask for something closer to: "Read the attached meeting transcript and produce notes in exactly this structure: a one-sentence TLDR at the top, a 'Decisions' section listing only choices that were actually finalized, an 'Action items' section listing each task with an owner and a deadline (write 'unassigned' or 'no date given' if the transcript does not state one), and an 'Open questions' section for anything raised but not resolved. Do not include general discussion that did not lead to a decision or a task." That structure forces the model to sort everything it reads into one of four buckets instead of defaulting to a single blended paragraph, and it makes gaps in the meeting itself visible instead of quietly smoothing them over.

Handling a messy or informal transcript

Real meetings ramble, get interrupted, and have people talking over each other, especially in auto-generated transcripts from Zoom or Google Meet that misattribute speakers or garble names. Tell the model explicitly to work only from what is clearly stated in the transcript and never invent a name, deadline, or decision that was not actually said. If a speaker’s identity is unclear (labeled "Speaker 2" instead of a name), ask the model to preserve that label rather than guessing who it was, since a confident wrong guess attached to a real commitment is worse than an honestly unlabeled one.

Different meetings need different summary shapes

A status update or stand-up: the useful shape is almost entirely action items with owners, with barely any narrative recap needed since everyone already has the context. A brainstorm or strategy session: the useful shape flips toward open questions and options considered, since a brainstorm rarely produces final decisions, and forcing false ones into the summary misrepresents what actually happened in the room. A client or external call: the notes need to separate what your side committed to from what the other side committed to, since those two lists often get forwarded to very different people afterward, and mixing them creates real confusion about who owes what.

A worked example, before and after

Weak prompt: "summarize this meeting transcript." No structure, no instruction about decisions versus discussion, no mention of owners or deadlines, so the model produces a pleasant-sounding paragraph that would take another read of the full transcript to actually act on. Structured prompt: "Summarize the attached product planning call. TLDR in one sentence. Decisions section: list only what was finalized, not what was debated. Action items section: name, task, deadline, marking anything without a stated deadline as 'no date given.' Open questions section: list anything raised but not resolved by the end of the call." The second version gives the model a fixed shape to sort the transcript into, so what comes back can be scanned in twenty seconds by someone deciding what to do next.

Common mistakes

Asking for a summary with no format specified, which almost always produces a narrative paragraph organized by conversation flow rather than by decisions and owners. Letting the model fill in an owner or deadline that was not actually stated in the transcript, which creates a false sense of certainty about who agreed to what. Treating a stand-up summary and a strategy-session summary as the same shape, when one should be almost entirely action items and the other almost entirely open questions. Pasting a long raw transcript without asking the model to flag ambiguous speaker attribution, so a garbled auto-transcript name gets treated as a confirmed source for a real commitment.