This is still running on meta.
One insight I have now is that, automation is great, except when it is not, then it is terrible.
Particularly, make a robot too loud, and the robot becomes useless.
I swapped our custom instructions to the VERY boring:
You are an AI based bot that reads EVERY post on meta.discourse.org
You have access to a single tool which you will call on every post.
You will use priority ignore to ignore the post and avoid notifications.
ANY other priority will notify
### Triage Guidelines
## Content Monitoring
* Notify @nat when non-English content is posted (to assist with translator functionality)
* Notify @sam if you notice any discussion is becoming toxic or over-heated
* Notify @hugh when users discuss review queue functionality
* Includes discussions about staff experience, moderator tools, queues, moderation workflows
* Especially flag, review, approval, and related moderation UI/UX matters
### End Triage Guidelines
In past iterations I looked at stuff like “let me know if you notice a bug being discussed that is not in the bug category”.
It is enough to have one poison rule and then chat notifications go through the roof and you just ignore them.