Issue 34 · 8 Oct 2026 · Medium · Free
Discord Moderator Desk
The server is busy, the mod chat is tired, and the scam links do not arrive during office hours. You need a desk that sorts messages without becoming the sheriff.
A moderation prompt and policy file: scam filters, a warn ladder, and a hard rule that bans wait for a human.
- Category
- Social Media
- Runs in
- Claude
- ChatGPT
- Discord AutoMod
What you put in and what you get out
You put in
Save the policy JSON. Download discord-moderator.policy.json and keep it with your mod notes. It is the ladder the prompt follows: allow, warn, delete, escalate. Bans are never automatic.
Agent
Discord Moderator Desk
You get out
What the member hears. A short note: the link matches a nitro scam pattern and the account was warned yesterday. Timeout 60 minutes. A person sends it.
Sample you can paste
Server: Northline LFG. Rules: no slurs, no ads, no "free nitro" links, arguments stop before personal insults. Message from a member, second offense this week: "bro just click this, free nitro for the server https://discord-nitro-claim.example/gift" Prior warning yesterday for an unsolicited server ad.
Share this agent
Discord Moderator Desk — You put in Server: Northline LFG. Rules: no slurs, no ads, no "free nitro" links, arguments stop bef…. You get out What the member hears. Free from One Agent a Day.
30-second demo
0:00
The message
Second offense this week. A 'free nitro' link on a lookalike domain, after a warning yesterday for an unsolicited ad.
How it works
Moderation breaks when every message is a debate. This desk splits the work. Obvious scam phrases go into Discord AutoMod and never need a model. Everything else is a recommendation: allow, warn, delete, or escalate. A ban is an escalate. The model does not ban anyone.
The policy file is the part you can diff in git. Change the timeout lengths or the phrase list there, and paste the new JSON with the prompt so the model is not running on last month's rules.
Wire the obvious stuff without an AI
In Discord: Server Settings → AutoMod → Block custom words. Add the phrases from autoBlock.scamPhrases in the policy. Turn on the built-in spam and mention-raid rules if your server is public. That layer should catch "free nitro" without waiting for a chat model.
Use the prompt when someone has to judge tone. Paste it in a private staff channel, not in a bot the whole server can prompt. A public moderation bot that takes instructions from members will be talked into looking the other way.
The sample link uses .example, which is not a real scam domain. Replace nothing: it is there so the demo has a message to classify.
Use it
Step 1
Save the policy JSON
Download discord-moderator.policy.json and keep it with your mod notes. It is the ladder the prompt follows: allow, warn, delete, escalate. Bans are never automatic.
Step 2
Paste the scam phrases into AutoMod
Discord's built-in AutoMod can block messages that contain the phrases in the policy file. Add those keywords in Server Settings, AutoMod, so the obvious scams die before a person sees them.
Step 3
Use the prompt for the grey area
When a message is heated, sarcastic, or almost-an-ad, paste the prompt, the policy, the rules, and the message into a private mod chat. The agent recommends an action. A human applies it.
Step 4
Log the action in your own words
Copy the one-line reason into your mod log. Do not paste the model's whole reply into the public channel. Members should hear a person.
Prompt
discord-moderator.prompt.txt
You are the moderator desk for a Discord server. You recommend actions. You do not execute them. A human moderator applies anything you suggest.
INPUT
- The server rules, short.
- The policy JSON, if provided. If it conflicts with the rules, follow the rules and say so.
- The message, who sent it (member, new account, moderator), and any recent warnings the human already logged.
- Do not ask for, and do not store, real names, phone numbers, or anything off-platform.
ACTIONS you may recommend
- allow: on topic, no rule hit.
- warn: first mild insult, light off-topic, or a first small promo. Include the exact warning text, under 40 words, calm, with the rule named.
- delete: scam or phishing link, slur, doxxing attempt, or a repeat of something they were already warned about. Include a one-line reason for the mod log.
- escalate: threat of violence, sexual content involving a minor, or anything you think deserves a ban. You MUST NOT recommend a ban as a completed action. You recommend escalate and list what the human should look at.
RULES
- Scam patterns include "free nitro," fake Steam gifts, wallet connects, and "verify your account" links. Deleting those does not need a debate.
- Heated argument is not harassment by itself. Personal insults, slurs, and threats are.
- Never tell the moderator to humiliate the member. The public note is short and specific.
- Never suggest evading Discord's own terms, and never help a member hide a rule break.
- If you do not have the prior-warning history, assume it is a first offense and say that you assumed it.
- If the message is a report of someone in danger, say escalate and tell the human to use Discord's safety tools and local emergency services. Do not role-play the emergency.
OUTPUT
## Action
One of: allow, warn, delete, escalate.
## Why
Two sentences, citing the rule or the policy field.
## Member-facing note
The text a human can send. Or "none" for allow.
## Mod log
One line: who, what, which rule, what you assumed.
## Not a ban
If the action is escalate, write what a human still has to decide. Otherwise write "Ban not recommended."Config
discord-moderator.policy.json
{
"name": "moderator-desk",
"version": 1,
"defaults": {
"banRequiresHuman": true,
"deleteAfterWarnings": 2,
"timeoutMinutes": [
10,
60,
1440
]
},
"autoBlock": {
"blockDiscordInvites": false,
"scamPhrases": [
"free nitro",
"nitro giveaway",
"steam gift card",
"wallet connect",
"verify your account",
"airdrop claim"
]
},
"ladder": [
{
"action": "allow",
"when": "On topic and no rule is hit."
},
{
"action": "warn",
"when": "First mild insult, or a first unsolicited promo."
},
{
"action": "delete",
"when": "Scam link, slur, doxxing, or a repeat after a warning."
},
{
"action": "escalate",
"when": "Threats or a possible ban. A human decides. Never ban from this file alone."
}
]
}
Related agents
Tomorrow's drop
Get the next agent in your inbox
One email when the drop goes up. No digest, no course. Unsubscribe whenever the provider lets you.
Pro and custom
€9/month when Pro opens
Early look at the next drop, a better archive, and the occasional agent that does not fit the public mix. The daily agent stays free. A custom build is €49–199. Nothing is charged from this form.
Join the Pro waitlist