If your meeting tool produces a transcript, AI can turn it into notes people actually read.
The prompt
You are writing notes for people who missed this meeting.
From the transcript, produce:
## Summary
3-5 sentences.
## Decisions
- Each decision, one line.
## Action items
| Owner | Task | Due |
Only include tasks someone explicitly agreed to. Use "unassigned" if no owner was named.
## Open questions
- Questions raised but not answered.
Rules: don't invent owners or dates. Keep names as written.Code
def meeting_notes(transcript):
r = client.chat.completions.create(model="claude-sonnet-4-6", temperature=0.2, messages=[
{"role": "system", "content": PROMPT},
{"role": "user", "content": transcript[:200000]},
])
return r.choices[0].message.contentFor very long meetings, summarise in parts first (see summarising long PDFs, the same map-reduce idea).
Make it automatic
- Watch a folder or your meeting tool's webhook for new transcripts.
- Post the notes to your team chat or email them to attendees.
- Save action items to your task tracker.
Tip: The rule “don't invent owners or dates” matters. Without it, models helpfully assign tasks to whoever spoke last.
Claude, GPT, Gemini, DeepSeek, image and video models through one OpenAI- and Anthropic-compatible endpoint. Prepaid, pay only for what you use.