From Meeting Transcript to Action Items in Under Five Minutes
The meeting ends, everyone agrees it was productive, and then… nothing happens. No one wrote down who owns what. A week later half the decisions are re-litigated because nobody can remember what was actually decided.
This is the most reliable, lowest-risk AI task at work, because the source material is complete and factual — you’re not asking the model to invent anything, just to organize what was already said. Here’s the whole workflow.
Step 1: Get a transcript
Most video tools (Zoom, Teams, Meet) can produce a transcript
A text version of everything that was said, usually with speaker names and
timestamps. Most meeting tools generate one automatically if you turn it
on; if not, you can record the audio and drop it into a transcription tool
first. The AI works from this text, not the video itself.
Step 2: Paste the transcript, not a summary
Drop the full transcript into your AI tool. Yes, the whole thing, even if
it’s long — the model can hold a surprising amount of text in its context window
The amount of text a model can consider at once — your message, any
documents you paste, and its own replies. Modern tools have windows large
enough to fit an hour-long meeting transcript with room to spare, so you
rarely need to trim it down first.
Step 3: Use this prompt
Copy this, adjust the last line to your team’s reality:
“This is a transcript of our team meeting. Produce three sections:
1. Decisions made — only things that were actually decided, quoted or closely paraphrased. If something was debated but not resolved, put it under Open Questions instead.
2. Action items — a table with the task, the owner, and any due date mentioned. Only list an owner if a specific person clearly took it on; otherwise mark it ‘unassigned.’
3. Open questions — anything raised but not resolved.
Do not invent owners, dates, or decisions. If it wasn’t in the transcript, it doesn’t go in the notes.“
That final instruction is doing real work. It draws a hard line against a hallucination
When a model fills a gap with a confident-sounding guess instead of
admitting it doesn’t know. In meeting notes this shows up as a plausible
but fake owner or deadline — exactly the kind of error that causes real
problems — so you tell it up front to leave gaps as gaps.
Step 4: Sanity-check the owners and dates
Read the action-items table with one question in mind: is every owner and date actually something a human said? This takes thirty seconds and is the only real review the output needs. Everything else — decisions, open questions — is low-stakes if slightly off. Owners and dates are the part worth verifying, because that’s the part people will act on.
Step 5: Send it while it’s warm
Fix anything wrong, then send it to the room the same day, with a line like “Here’s what I captured — reply if I got anything wrong.” You’ll get corrections while memories are fresh, and you’ve created a written record that ends the “wait, what did we decide?” cycle.
Why this one is worth starting with
If you’re nervous about using AI at work, this is the task to build confidence on. The stakes are low, the material is factual, the review is quick, and the payoff is immediate and visible to your whole team. Once you trust it here, you’ll start spotting the next ten tasks shaped just like it.
The takeaway
You’re not asking the AI to have opinions or make judgment calls — you’re asking it to be the diligent note-taker nobody wants to be. Give it the full transcript, tell it not to invent anything, verify the owners and dates, and send. Five minutes, and the meeting actually leads somewhere.
Related: Build a Reusable AI Assistant — turn this prompt into a saved tool so you never paste it again.