meeting-minutes

Read the full transcript in chunks, count utterances per speaker, then produce a minutes document with a one-line summary, schedule, agenda discussion, numbered decisions, and a three-column action item table with owner, task, and deadline. Cross-check the previous meeting's minutes so recurring references are assigned correctly.

Convert the mp3 to 16kHz mono opus with ffmpeg, then send it to the Groq Whisper API (whisper-large-v3) with a prompt listing domain-specific proper nouns to improve accuracy. Run this on a remote machine since the pipeline is resource-heavy, and note that personal recordings lack speaker diarization.

Create a page in the minutes database with attendee relations pointing to person database pages, incrementing the meeting number from the latest entry. Due to a known Notion MCP bug, number properties like meeting number must be entered manually in the browser after page creation.

Check the bot log for decoder errors indicating a participant's audio was never recorded, then ask attendees for personal recordings before publishing anything. Never reconstruct missing content from surrounding chat text, and mark any reconstruction with a clear notice separating in-meeting content from post-meeting discussion.

A follow-up bot reads recent minutes bodies as evidence when judging whether watchlist items are done. Statements describing verified demonstrations close items as green, while vague or speculative phrasing can wrongly close items or leave finished work flagged as overdue.