Playbook, meetings

Turn messy meeting notes into clean action items

A five-minute routine that pulls owners, dates, and decisions out of a transcript, and catches the model when it guesses.

The Curio Team · 3 min readWritten with AI from the sources below and checked against them before publishing.

Bad meeting notes cost you twice. Once in the meeting, and again a week later when nobody remembers who owns what. We turn transcripts into action items in about five minutes now. The trick is not a fancy tool. It is a tight prompt plus a verify pass that catches the model when it guesses. Here is the routine we run, from paste to send.¹

Language models are strong here because summarizing and extracting structure from text sits close to what they were trained to do. But they will invent an owner or a due date if you let them. So we ask for the same fields every time. We also make the model separate what was decided from what is still open.^{1 2}

Any capable chat model does this well. Claude, ChatGPT, and Gemini all read a long transcript and return structured output on demand. The workflow matters more than the brand. What changes your result is the prompt and the check, not the logo on the tab.

The prompt below does the heavy lifting. We paste it once, then drop the notes underneath. You can save it as a snippet and reuse it for every meeting, which is what turns a one-time trick into a habit.

Run these five steps in order. Each one takes under a minute.

1. Paste the raw notes. Drop the full transcript or your messy notes into the chat without cleaning them first. Messy is fine. The model handles filler and false starts better than you expect, and hand-editing first risks dropping a real commitment.
2. Ask for owners and dates. Tell it to pull every action item into a table with three columns: task, owner, due date. When the notes name no owner or no date, it must write Unassigned or No date rather than guess. That one rule stops most fabrication.
3. Split decisions from open questions. Ask for two more lists. One holds decisions the group actually made. The other holds open questions nobody resolved. This split is where the real value lives, because open questions are the things that quietly die between meetings.
4. Verify against the source. Read the output next to the transcript and check every owner and date against a real sentence. If the model cannot point to a line, cut it. We catch one or two invented details almost every single time we do this.
5. Send it fast. Paste the clean table into email or your tracker within the hour, while the context is fresh in everyone's head. Ask people to reply only if their task or date is wrong. Silence means agreement.

One warning. If your transcript is very long, the model may summarize instead of extracting, and quiet commitments get lost. Split it into two halves, run each, then merge. Boring, but it works.

The model writes the draft. We stay the editor.

The verify step is the one people skip, and it is the one that matters. A clean table nobody checked is worse than messy notes, because it looks official while carrying a wrong owner. So we treat the model as a fast first draft and ourselves as the editor. That division of labor is the whole game.²