Stop Discord spam and scam links without slowing your own crowd

Why this goes wrong

Switching rules on takes a minute, and that is exactly the trap: turn everything live at once and at the first bit of trouble you no longer know which rule caused it. What you cannot see weighs more, though. A correct hit and a wrong one look identical in the log, and nobody comes forward after a harmless message of theirs was cleared away. People do not complain, they stop writing. Someone who loses two messages in a row posts no more links, asks no more questions with a screenshot, and drops their find somewhere else instead. That loss shows up in no number, while the work the team saves is visible immediately. The question is therefore not how strict a filter can be, but how you notice it has become too strict. That one does not answer itself.

What to decide first

Sort the rules by what they judge. Anything that recognises content (known scam addresses, foreign invites, a word list you genuinely agreed on) rarely gets it wrong and may go live first. Anything that judges form (capitals, addresses in general, many mentions, tempo) sooner or later catches everyone who is simply annoyed or wants to show a bug. That group comes later and one at a time. Before all of it the exception list belongs on paper, not after: which channels live on links, in which is a foreign invite precisely the point, which roles must never be touched. And the third decision saves the most grief: separate cleaning up from punishing. Removing a message is housework and needs no file. A warning is a statement about a person. Hang both on the same rule and you lose the chance to talk to somebody before the counters decide for them.

Step by step

1. First write down which channels and roles stay untouched by every check, and enter that before the first rule goes live.
2. Switch on exactly one rule (the one that recognises content and rarely misfires) and let it run for a week without changing anything else.
3. After that week read the log from top to bottom and count how many hits you would have decided differently in hindsight.
4. Only once that number is small do you add the next rule, again just one, again for a week.
5. Right at the end you decide whether a hit should also go on the record; while you are unsure, leave it at cleaning up.

The mistakes almost everyone makes

The strictest rule first: that is the classic. It looks like it settles everything in one go and then hits the help channel, where people link something precisely because they are stuck. The second mistake is tightening up after a single bad evening. One night is not a pattern, and whoever turns three screws at once afterwards can no longer attribute an effect to any change. The third grows slowly: after every complaint another role goes on the exception list, and six months later the rules only apply to newcomers, that is, to the group most likely to get something wrong by accident and least likely to have anyone speak up for them. The fourth is the quietest: nobody told the members these rules exist at all. A message that vanishes without a word feels arbitrary, even when it vanished for good reason.

Frequently asked questions

The one that recognises content instead of form: known scam addresses. No ordinary member sends those on purpose, you have nothing to maintain, and the rule passes no judgement on tone. Everything that rates volume, tempo or length can wait until week two.

By repetition. Same person, same channel, three attempts within a minute: somebody was trying to get a message through and did not understand why it kept disappearing. If one particular channel keeps showing up, the rule does not fit that room rather than the room being bad.

Once you have read a week of hits and would have decided almost every one the way the bot did. Before that the counter is adding up your own uncertainty, and it surfaces later as a punishment nobody can explain. Until then quiet cleaning up is entirely enough.

Build it by reversibility. A deleted message is annoying, a short pause is instructive, a removal ends the relationship. The first two steps the bot may take on its own; the last one, the one you cannot undo, belongs in the hands of a person who has read the history.

Which rule do I switch on first?

The one that recognises content instead of form: known scam addresses. No ordinary member sends those on purpose, you have nothing to maintain, and the rule passes no judgement on tone. Everything that rates volume, tempo or length can wait until week two.

How do I spot a false alarm when nobody complains?

By repetition. Same person, same channel, three attempts within a minute: somebody was trying to get a message through and did not understand why it kept disappearing. If one particular channel keeps showing up, the rule does not fit that room rather than the room being bad.

When may an automatic warning run alongside?

Once you have read a week of hits and would have decided almost every one the way the bot did. Before that the counter is adding up your own uncertainty, and it surfaces later as a punishment nobody can explain. Until then quiet cleaning up is entirely enough.

What does an escalation look like that punishes nobody for a slip?

Build it by reversibility. A deleted message is annoying, a short pause is instructive, a removal ends the relationship. The first two steps the bot may take on its own; the last one, the one you cannot undo, belongs in the hands of a person who has read the history.