The meeting that produced 18 duplicate tasks
First version of the Meeting to Task extractor. One meeting recorded in two platforms. Claude parsed the transcript twice. The ops lead opened ClickUp Tuesday morning and saw 36 tasks where 18 belonged.
The robot did exactly what we asked. The fix was to teach it which source wins.
What happened
Acme ran a two-hour ops review. Fireflies captured it. The same call was also recorded in Zoom's native transcript. Both ended up in our pipeline. Claude saw two transcripts, extracted 18 tasks from each. ClickUp got 36. The ops lead noticed within an hour and sent a polite "um, what".
The fix
We added a deduplication check on the source meeting id, not the transcript text. Fireflies gets priority because it has cleaner attendee data. The Zoom transcript becomes an archive, not a task source. Two weeks later, clean runs: 14 tasks from a two-hour meeting, all attributed, all deduped.
The lesson
Deduplication runs at the ingest layer, not the write layer. If two sources can produce the same task, one has to be declared canonical before Claude ever sees it. Every automation we ship now has a "canonical source" rule baked in before it writes anywhere.
Official Notion Builder · Partner Program
3 months free Notion Business for your team
New and existing Notion users. Tracked via PartnerStack. Activate through our Builder referral. PartnerStack tracked. No extra cost to you.
The audit is thirty minutes. The cost of waiting is not.
Book a scoping call. We map what is manual, what can run automatically, and what the build would actually cost in time and money.