How to Prompt AI to Turn a Voice Memo Into Meeting Notes

Talk into your phone for two minutes right after a meeting, then paste the transcript into this prompt: “This is a rambling voice memo I recorded right after a meeting. Turn it into structured notes with three sections: decisions made, action items with an owner for each, and open questions. Do not add anything I did not say, and flag anything ambiguous instead of guessing.” That single instruction fixes the two ways this task usually goes wrong.

Why a voice memo is a different input than a transcript

A meeting transcript is already structured by turn-taking and has multiple speakers to disambiguate. A voice memo is one person thinking out loud, often out of order, correcting themselves mid-sentence, and using “I think we said” instead of a clean decision statement. If you have an actual meeting transcript instead, our guide to prompting AI to summarize a long meeting transcript is the more direct match. A voice memo needs a prompt that expects messiness and disorder, not clean turn-by-turn dialogue.

The three failure modes this prompt prevents

Left unguided, a model turning a rambling recording into notes tends to fail in one of three ways: it smooths over genuine ambiguity into a false-confident decision, it invents an owner for an action item nobody actually assigned, or it drops something you mentioned in passing that mattered. Naming decisions, action items, and open questions as three separate required sections forces the model to sort ambiguous statements into “open questions” instead of silently resolving them for you.

The full prompt template

1. State the input clearly: “This is a rambling voice memo I recorded right after a meeting, not a clean transcript.”

2. Ask for the three sections by name: decisions made, action items with an owner, open questions.

3. Add the constraint that matters most: “Do not add anything I did not say, and flag anything ambiguous instead of guessing.”

4. If names get slurred or mumbled in the recording, add: “If a name is unclear, write [unclear name] rather than guessing which teammate it was.”

5. Ask for a one-line summary at the top before the three sections, so the notes are skimmable without reading everything.

A short before and after

Raw voice memo transcript, unedited: “okay so, um, I think we're going with the Tuesday launch, unless, no wait we did agree on that, and someone needs to, I think Priya said she'd handle the email but I'm not sure if that included the social post too, and we still don't know if legal signed off.”

With the prompt above, that becomes: a one-line summary (“Launch confirmed for Tuesday; email ownership needs clarifying; legal sign-off still open”), a decisions section listing the Tuesday launch date, an action items section listing “Priya: send launch email (scope unclear, confirm if social post included)” rather than a falsely confident single task, and an open questions section flagging the legal sign-off. Nothing is invented, and the genuine ambiguity around the social post is preserved instead of erased.

Once you have notes, turn them into tasks

Structured notes are the middle step, not the end goal. Once you have decisions, owners, and open questions cleanly separated, turning meeting notes into action items is the natural next prompt, and it works far better on notes that are already this organized than on a raw transcript.

If you need formal minutes instead

Notes for your own reference and formal meeting minutes for a record are different documents with different tolerances for informality. If what you actually need is the latter, prompting AI to draft meeting minutes covers the more formal format and what it needs to include.

This is one small pattern out of many. Our prompt engineering hub rounds up the rest of the templates worth keeping on hand.

FAQ

Do I need a perfect transcript first?

No. Most voice-to-text tools produce a rough transcript full of small errors, and that is fine input for this prompt. The instruction to flag ambiguity rather than guess covers transcription errors as well as your own rambling.

What if I mention several unrelated topics in one memo?

Add a line asking the model to split the output by topic before applying the three-section structure to each. Otherwise action items from unrelated topics can get merged into one confusing list.

Can this work directly from an audio file?

Most current AI assistants need a text transcript, not raw audio, though some tools now transcribe and summarize in one step. Either way, the prompt structure above applies once you have text, whether you typed it or a transcription tool produced it.