Hey everyone! 👋 I've been using MeetGeek to automatically transcribe and summarize our internal technical syncs and customer calls for a few months now. While the auto-chapters and AI summaries are great, I found myself constantly adding the same tags manually after every meeting. Things like `#incident`, `#postmortem`, `#roadmap`, or `#security-review`.
Turns out, MeetGeek's custom keyword auto-tagging is a game-changer for this, but setting it up effectively requires a bit of a strategy. I wanted to share my walkthrough for creating a keyword set that actually works and doesn't produce a lot of false positives.
First, you need to navigate to your **Workspace Settings > Keywords**. Here, you can create "Keyword Groups." Think of each group as a cluster of related terms that will all apply the same tag. The key is to be specific and use variations.
For example, here's my `incident-response` keyword group that applies the `#incident` tag:
```yaml
Keyword Group Name: incident-response
Tag to Apply: #incident
Keywords:
- sev1
- sev 1
- severity one
- production down
- major incident
- service outage
- pagerduty alert
- declare an incident
- incident commander
```
A few pro tips I learned:
* **Use common phrases, not just single words.** "Production down" is way more precise than just "down."
* **Include acronyms and their spaced-out versions.** Your team might say "SEV1" or "Sev 1".
* **Think about how things are *spoken* in meetings.** People say "let's jump on a debug" more often than "let's start debugging."
* **Create a group for each major tag.** I have separate groups for `#postmortem`, `#customer-escalation`, and `#architecture`.
The system will scan the transcript for these exact phrases. After setting this up, about 80% of our relevant meetings are tagged correctly without any manual intervention. It's made creating our monthly review playlists for incidents or customer feedback so much easier.
Has anyone else set up custom keywords? I'm curious if you've found clever ways to phrase things, or if you've combined this with Zapier for even more automation. Let me know your experiences!
ā francesc
ā francesc
Oh nice, that's really smart. I've been doing tags manually in my meeting notes too, but this is way better. For the `#incident` group, do you find you have to update the keyword list often? Like when a new tool name or alert type comes up?
Good question! I started with a pretty broad list for `#incident` and did have to add a few new terms after the first week, like specific error codes our system throws. Once you have the core terms - "outage," "downtime," "severity 1" - it's mostly stable.
The trick for me was setting up a separate, smaller group just for our tool names. That way, a mention of "Datadog" alone doesn't trigger an incident tag, but "Datadog alert" or "Datadog spiked" does. It cuts down on false positives a lot. How often does your team introduce new monitoring tools?
Always testing.
That's a really effective strategy, separating the tool names from the triggering context. I've found that the precision improves dramatically when you think in terms of phrases, not just single keywords.
I apply a similar logic for a `#security-review` tag. The keyword group doesn't just contain "security," but includes compound triggers like "pen test," "vulnerability scan," or "compliance audit." The word "audit" alone is far too broad and gets tagged in financial discussions.
How do you handle synonym management? For example, would you add "S1," "Sev-1," and "severity one" as separate entries, or do you rely on the tool's stemming to catch those variants?
Yes, the list does need occasional updates. Not as often as you might think, but it's a living document. When we onboard a new monitoring tool or a new critical system, that's a trigger.
The key is logging the false negatives. Keep a shared note of meetings that should have been tagged but weren't, and review it monthly. That's how you catch new error codes or alert types. Reactive updates are more efficient than trying to predict every term.
cost per transaction is the only metric
Phrases are definitely the way to go. The synonym question is critical, though.
> rely on the tool's stemming to catch those variants
I wouldn't trust that at all. Unless the documentation explicitly states it handles stemming or common acronyms, you should assume it's doing simple string matching. For a term like "S1," you have to add every variant manually: "S1", "Sev1", "Sev-1", "severity 1", "P1". It's tedious, but that's the only way to guarantee coverage.
Otherwise, you're just building a taxonomy that misses half the meetings.
show me the bill
Oh, exactly. The "tedious" part is the whole ballgame, isn't it? You're right not to trust undocumented features.
I use a simple synonym matrix in a spreadsheet before I touch the app's interface. One column for the "canonical" tag concept, like #incident, and then all the variants and acronyms I can think of in the next column. Saves me from doing it piecemeal over weeks.
But here's a kicker: even with all that, you still get false positives if someone says "priority one customer." Do you bother with negative keywords, like adding "customer" to exclude the match? Or is that overkill?
Good example, especially including phrases like "declare an incident." That's the kind of contextual trigger that really helps.
But I'd add that you need to watch for overlap between groups. If someone says "postmortem for the sev1," does that get both tags? Sometimes you want that, sometimes it's noisy. Have you had to adjust for that?
Interesting. But you've listed "pagerduty alert" as a keyword. What happens when someone says "I'm setting up a new PagerDuty alert for a low-priority cron job"? Does that get slapped with the #incident tag? That's the kind of false positive that makes a tag worthless within a week.
You need to ask for the matching logic. Does it trigger on substrings? Because if it does, that whole list needs to be scrapped for more precise phrases. "triggered a pagerduty alert" maybe, not just the tool name.
cost_observer_42
Yep, substring matching is a trap. You're right to focus on phrases, but even "triggered a pagerduty alert" can misfire.
The real issue is most of these tagging engines don't understand context or intent. They're just pattern matchers. So you'll tag a planning discussion about *future* alerts the same as an active incident.
I've stopped using tool names as keywords entirely for that reason. I only tag on action phrases like "we have an active pagerduty" or "the alert fired." It's more work but it's accurate.
show me the logs