How I Built a Routine for AI-Assisted Meeting Notes
I got tired of wasting 30 minutes every meeting just writing down notes, so I put together a routine.
I run the recording through Whisper to get a transcript, then feed it into an LLM and have it organize everything into three buckets: decisions, action items, and open issues. The prompt looks roughly like this.
"Take the meeting transcript below and organize it into 1) confirmed decisions, 2) action items with owners and deadlines, and 3) agenda items to revisit at the next meeting. Do not speculate, and do not include anything that isn't in the text."
The most important part was that last sentence. Without it, the model would make up plausible-sounding details that weren't there, which actually made review take longer.
Now I get a summary within five minutes after a meeting ends. It's definitely faster than having a person summarize, and it misses fewer things.
3 answers
Oh, I didn’t know this—I should save this prompt.
Adding a hallucination-suppression line is genuinely the key. I went further and added “mark anything unverifiable as [unverified],” and my review time dropped dramatically. It still often pulls the wrong person-in-charge name, so a human does have to skim it once in the end—but five minutes is enough. Agreed.
Well, that’s a bit… meeting minutes are ultimately a document that establishes accountability, so I’m curious how you handle storing the original recording files and privacy issues. Many companies’ internal policies don’t even allow uploading meeting audio to external APIs.