The seven rules
| Rule | What it detects | Most important setting |
|---|---|---|
| Banned words | Messages that contain a word from your list | Word list, separated by spaces or commas |
| Anti-spam | Too many messages in too short a time | Number of messages per time window |
| Invites | Discord invite links to other servers | - |
| Links | Any web addresses | - |
| Caps | Messages that are mostly shouted | Maximum share in percent, minimum length |
| Mention spam | Too many mentions in one message | Maximum number of mentions |
| Scam patterns | Common scams such as gifted games or fake login pages | - |
Which action fits what
Every rule gets its own action. You do not have to choose the same one everywhere, and you should not.
- Delete only, the message disappears, nothing else happens. Right for caps and links: usually carelessness, not intent.
- Warn, on top of the deletion, a warning is counted. Combined with the escalation ladder, repetition then leads to a timeout on its own.
- Timeout, mutes for the configured duration. Right for anti-spam, because there is usually an automated account behind it.
- Kick or ban, only sensible for the scam pattern filter. Someone out to harvest login details is rarely there by accident.
The two settings that really matter
Experience shows that it is not the choice of rules that decides whether Auto-Moderation is felt to be helpful or annoying, but the exceptions.
One rule at a time
We advise against arming all seven rules at once. Otherwise the usual course of events is: everything on, three days of complaints, everything off again.
- Week one: anti-spam and scam patterns. Both almost never hit a real user.
- Week two: invites and mention spam. Now you can see which channels need exceptions.
- After that: caps, links and banned words, the three with the highest share of false hits.
Maintaining banned words properly
The word list is more powerful than it looks, and therefore the most common source of trouble. Three notes:
- The filter checks whether the word is contained in the message. A short entry therefore also hits harmless longer words, the classic example is a ban that suddenly deletes place names.
- Upper and lower case make no difference.
- Start with a few unambiguous terms and add whatever actually comes up. A list of three hundred words copied from the internet produces more false hits than benefit.
Limits you should know about
- Edited messages: The bot checks edits as well, rewriting a harmless message afterwards does not get you through.
- Images and attachments: The text is checked, not the image content. The filter does not detect a screenshot containing a banned word.
- Evasive spellings: The word list does not reliably catch deliberate distortions made with special characters. Only human moderation helps against that.