What this covers
The “paste the transcript and ask for notes” pattern produces something that reads cleanly and quietly invents commitments people didn’t make. Someone said “we could look into Q2” and the action item reads “Alex will deliver by Q2.” This guide is for the version of meeting notes that gets sent to the team — where a phantom action item costs you trust, and a missed real one costs you a deliverable. The fix is a transcript-prep step plus a tightly-constrained extraction prompt.
Who this is for
PMs, team leads, founders, and anyone who runs or attends meetings where decisions and follow-ups need to be written down accurately. If your meeting outputs are “vibes plus a vague summary,” you don’t need a workflow — keep doing what you’re doing. If somebody actually owns the action items afterward, the precision matters.
When to reach for it
- A 30-60 minute meeting where you need decisions, action items, and open questions written up the same day.
- A recurring meeting where the structured notes feed into a project tracker.
- A client or stakeholder meeting where the notes go out as a recap and need to be defensible.
- A multi-speaker discussion where pulling each person’s commitments by hand would take longer than the meeting itself.
Before you start
- Confirm everyone is OK with the meeting being recorded or transcribed. Consent first, transcript second. This is not optional.
- Prep the transcript before pasting: at minimum, run a find-and-replace on speaker labels so “Speaker 1” becomes “Alex” — the model assigns commitments much more reliably with names.
- Decide your output shape up front. The standard four-bucket template is: decisions / action items / open questions / parking lot. Use it consistently across meetings so the team learns where to look.
- For long meetings (>45 min of transcript), plan to chunk. A single prompt with a 90-minute transcript loses fidelity in the middle.
Step by step
-
Get the transcript. Either from a tool that captures it (Zoom, Meet, Otter, Teams) or by pasting your own recording into a transcription service. Confirm speaker labels are present.
-
Clean the transcript before extraction: replace “Speaker N” with real names, add timestamps every few minutes if your tool didn’t, and delete obvious junk (filler hellos, audio-issue interruptions).
-
Run extraction with the four-bucket template:
Extract from this transcript into four sections: 1. Decisions — only items explicitly agreed to. Quote the decision verbatim and tag the speaker. 2. Action items — only commitments made by a named person. Format: "[Owner] will [verb] [object] by [date if stated]". If no date was given, write "by: not stated". 3. Open questions — questions raised but not answered. 4. Parking lot — topics deferred to a future meeting. Do not infer ownership. Do not invent dates. -
Run a separate pass for a narrative summary (3-5 sentences) — keep it apart from the structured extraction so a summarization mistake doesn’t pollute the action items.
-
For long meetings, chunk the transcript by topic or by 15-minute windows, run extraction per chunk, then merge — re-running extraction on the merged output to deduplicate.
-
Send the draft to one attendee for a sanity check before broadcasting to the full group. Action-item attribution errors are easier to catch by someone who was actually in the room.
A prompt that prevents invented commitments
You are extracting action items from a meeting transcript.
Rules:
- An action item only counts if a named person explicitly
said they would do it. Phrases like "we could" or "maybe
someone should" are NOT action items.
- Never assign an owner the transcript doesn't assign.
- Never invent a deadline. If no date was said, write
"by: not stated".
- If a commitment is ambiguous, list it under "ambiguous"
with the quoted line, not under action items.The “ambiguous” bucket is the unlock. Without it the model forces every soft-sounding statement into an action item; with it, the genuinely committed items get cleaner and the gray-area ones surface for human judgment.
Quality check
- Sample 3 action items at random. Find the line in the transcript that supports each one. If you can’t, the model invented it.
- Verify owners. The most common failure is attributing an action to whoever was speaking when the topic came up, not whoever volunteered.
- Check deadlines. “By next week” in the transcript should appear as “by: next week” — anything more specific is an inference.
- Look for missing items. Skim the transcript for “I’ll” / “I can” / “let me” — anywhere the model missed a commitment.
How to reuse this workflow
- Save a
meeting-notes-prompts.mdwith the four-bucket extraction prompt and the narrative-summary prompt as separate sections. - For each recurring meeting type (1:1s, sprint planning, customer reviews), keep a tailored prompt — the action-item phrasing for a 1:1 differs from a planning meeting.
- Build a “transcript prep” checklist: speaker labels replaced, junk removed, timestamps spotty-checked. Run through it every time before extraction.
Recommended workflow
Consent → record/transcribe → clean transcript (speaker labels, junk removal) → four-bucket extraction prompt → separate narrative summary → attendee sanity check → broadcast.
Common mistakes
- Skipping consent. Recording without explicit OK isn’t worth the trust damage when it surfaces.
- Pasting a raw transcript with “Speaker 1, 2, 3” labels. The model will assign ownership erratically — sometimes to the wrong person, sometimes to “the team.”
- Treating “we could” as an action item. The model does this by default. The prompt has to forbid it explicitly.
- Inventing deadlines. The model fills in “by Friday” when nothing was said. The prompt has to forbid this too.
- One giant prompt for a 90-minute meeting. The middle of the transcript loses fidelity. Chunk, then merge.
- Broadcasting without a sanity check. The cost of one wrongly-attributed action item is much higher than the cost of a 10-minute review.
FAQ
- Which transcription tool should I use?: Whatever your meeting platform already provides is usually fine for note-extraction quality. Dedicated tools (Otter, Fireflies, Granola) win on speaker labeling and easier export, which matters more than raw accuracy.
- Can ChatGPT transcribe audio directly?: Voice/audio input handles short clips well; for long meetings, use a transcription tool first and paste the text. You’ll get better speaker separation.
- What about privacy for sensitive meetings?: If the meeting covers personnel, legal, or sensitive customer data, don’t paste the transcript into a consumer-tier account. Use an enterprise plan with the right retention policy, or redact before pasting.
- How do I handle meetings with 8+ participants?: Speaker labeling degrades. Prep the transcript with explicit role tags (PM:, Eng Lead:, Designer:) for the people whose commitments matter, and accept that for peripheral participants the attribution will be noisy.