Turn Raw Meeting Notes into Minutes and Action Items

Workflow area: Meetings · Tier: Foundation · Hands-on time: ~25 minutes

The value of a meeting evaporates if nobody writes down what was decided and who owes what. But formatting minutes is exactly the kind of transformation work models do well — if you demand structure instead of prose, and if you treat the output as a draft that you, the person who was in the room, must correct. The model was not in the room. You were.

When to use — and when not to

• Use it for working meetingsproject syncs, planning sessions, vendor calls, anything with decisions and follow-ups that need to circulate.
• Use it with messy inputfragmented notes, a raw transcript from Teams/Meet/Otter, or both pasted together. Cleaning mess is the whole point.
• Use it same-dayyour memory is the error-correction layer, and it decays fast. Twenty minutes after the meeting is the golden window.
• Skip it for sensitive meetingsdisciplinary discussions, performance reviews, legal or M&A topics. Those minutes need a careful human author and a controlled tool, not a paste into consumer AI.
• Skip the AI entirely for two-person catch-upsa three-line follow-up email is faster than any pipeline.

The recipe

1. 1.

   Capture ugly, capture everything. During the meeting, type fragments — names, numbers, "PRIYA disagrees re: timeline", "DECIDED: ship Sep 5". Do not tidy while listening. If an AI notetaker ran, grab the transcript too; notes plus transcript beats either alone.

2. 2.

   Prompt for the three-part structure. Minutes, decisions, actions — explicitly separated:

   Copy

   Below are my raw notes from a meeting. Turn them into:
   
   1. MINUTES — 5-8 bullet summary of what was discussed, neutral tone.
   2. DECISION LOG — table: Decision | Made by | Context. Only include
      things actually decided, not proposals or "we should probably".
   3. ACTION ITEMS — table: Action | Owner | Due date. If an owner or
      date is missing in my notes, write "UNASSIGNED" or "NO DATE" —
      do NOT guess or assign anyone yourself.
   
   Meeting: [NAME, DATE, ATTENDEES]
   My shorthand: [e.g. "RK = Rahul, infra = the migration project"]
   
   Notes:
   [PASTE RAW NOTES AND/OR TRANSCRIPT]
   
   Use only what is in the notes. Flag anything ambiguous with
   [CHECK] rather than resolving it yourself.
   

3. 3.

   Review as the person who was in the room. Good output looks like:

   Copy

   DECISION LOG
   | Decision                          | Made by | Context             |
   | Launch moves to Sep 5             | Anita   | QA needs 2 more wks |
   | Vendor demo before contract       | Rahul   | [CHECK] — msg unclear|
   
   ACTION ITEMS
   | Action                        | Owner      | Due    |
   | Send revised timeline to CS   | Priya      | Aug 8  |
   | Book vendor demo              | UNASSIGNED | Aug 15 |
   

   The UNASSIGNED rows are the deliverable, not a defect — they tell you exactly what to resolve before circulating. The classic failure is the model helpfully assigning an owner nobody agreed to.

4. 4.

   Fix the traps by hand. Three things models reliably get wrong from meeting notes: reversed decisions (the group changed its mind late — record the final position), sarcasm and floated ideas recorded as commitments, and misheard names in transcripts. Check each decision row against your memory before it becomes the official record.

5. 5.

   Generate the follow-up email from the corrected version. One more prompt: "Turn the corrected minutes above into a short follow-up email to attendees: thanks, decisions in bold, action table, ask for corrections within 24 hours." That correction window is your safety net — circulated minutes that nobody objects to become the record.

Make it yours

• M365 Copilot in Teams: for transcribed Teams meetings, Recap already produces notes and suggested actions — but run the decision-log prompt over the transcript yourself anyway; Copilot's recap optimizes for coverage, and your decision log optimizes for accuracy. Verify owners before pasting into Planner.
• Google Docs with Gemini: keep raw notes in a Doc, use Help me write with the step-2 prompt, and keep minutes in the same Doc under a divider — one searchable artifact per recurring meeting, which feeds pb-08's agenda builder.
• ChatGPT or Claude: best for the messiest input — photos of a whiteboard (describe or transcribe them), two people's notes pasted together, or a 45-minute transcript. Claude's longer context handles full transcripts without trimming.
• The recurring-meeting loop: save the prompt with your shorthand glossary filled in. Same meeting, same glossary, every week — the marginal cost drops to five minutes.

Quality check

• Every decision is realmade in the room, final position recorded, no floated ideas promoted to commitments.
• No invented owners or datesUNASSIGNED and NO DATE resolved by you, not the model.
• Names spelled righttranscripts mangle names; the minutes must not.
• Numbers verifiedbudgets, dates and targets checked against your notes, not the model's rendering of them.
• A correction window existsminutes circulate with "corrections by [date]" so errors surface before the record hardens.