Don't summarize your meetings

Every AI note-taker does the same thing. It turns an hour of talking into a page of bullets nobody reads.

I don’t want a summary. I want the one decision we made, findable in March.

So I record my meetings. All of them, with consent. The team knows. Right now I have 286 transcripts.

And the agent that goes through them has one job: throw things away.

Most meetings give me a couple of follow-ups. Some give me a decision worth keeping. Some give me nothing. That’s fine. Nothing is a valid answer. Pretending a status call was deep is not.

What runs underneath

MacWhisper records. That’s all it does for me. Of everything I tried, it’s the recorder I trust most.

Then two models take over. Soniox (stt-async-v5) writes the text, plugged in behind MacWhisper’s Deepgram option. NVIDIA’s Nemotron 3 Diarization works out who said what. It runs locally on my Mac through NeMo-Speech.cpp, about 3.7 seconds per audio minute on an M2 Pro, for free. If it fails, pyannoteAI in the cloud takes over.

One gotcha. Nemotron’s offline preset merged two of four people in a 108-minute meeting. The streaming preset matched pyannote on about 98% of the audio. Don’t trust defaults. Test on your own recordings.

Names matter more than models. A transcript where everyone is “Speaker 2” is useless three months later.

I used to fix names and titles in MacWhisper. Now every transcript lands in Command OS, the app I built to run the company, and I do it there. I name a speaker once and the name sticks for every later meeting. Then I can ask for every meeting where two specific people talked.

The processing runs in pi, an agent harness. The agent reads the transcript in chunks, checks names against my people list, and writes back to Command OS. Tasks and long-term memories are only proposals. Nothing gets in without my yes.

And for the weeks when I don’t get to it, a background job picks up anything untouched after 72 hours. It writes a short TL;DR and links the people. Busy weeks don’t lose meetings.

The prompt

Here’s the core of it, trimmed so it works without my setup:

You process a meeting transcript. Don't optimize for a faithful summary.
Optimize for future usefulness. For every part, ask:
"What from this will be useful in 3-6 months?"

1. Verify names. Speech recognition breaks names. Check every person and
   company against the list I give you. If unsure, ask me.
2. Split the content into three layers:
   - Evergreen: decisions, recurring patterns, role clarity, lessons.
   - Operational: follow-ups, blockers, prep for the next meeting.
   - Noise: small talk, debate details, who-said-what chronology. Drop it.
3. Pick the smallest output that fits:
   full note / short note / one-line stub / nothing.
   Don't invent depth for an operational meeting.
4. Note format: TL;DR (max 6 bullets, written to me), Decisions,
   Open questions, Action items.
5. Propose tasks and long-term memories. Never create them. Wait for my "yes".
6. Put each insight in one place only. No copies across note, tasks and memory.

The important line is the question. Not “what happened?” but “what will I need later?”

Ask that, and most of the meeting disappears. Good.