From install to confident moderation

The mod card

When the AI flags a message, a card lands in your mod channel with the content, the author's history, and which rule the AI thinks was broken. Every card is decided by a human:

• ✅ Approve — Repost Message — the AI got it wrong. The message is restored, and the AI learns from the correction.
• ❌ Reject as Flagged / ❌ Reject — Select Rules / ❌ Reject — Other Reason… — confirm the violation (as flagged, against rules you pick, or with your own reason). A strike is recorded and the decision becomes precedent the AI applies to repeats. Some servers see a single ❌ Reject button instead — it opens the same rules picker with the AI's citation already selected, so agreeing is just Reject → Confirm (rolling out gradually).
• ⚠️ Warn — close the flag with a warning to the author instead: no strike, and it doesn't teach the AI.
• 🗑️ Allow Withdrawal — when the author asked to withdraw, let the message go quietly: no strike, no precedent.

Cool-off holds — what members see

When enabled, a borderline message can be held briefly instead of being flagged outright. The author gets a private notice with Edit Message (the edit is re-checked immediately), Post Anyway (sends it to moderator review — never straight to the channel), or Delete. Most incidents end here without a moderator lifting a finger.

Strikes & escalation

Each rejected flag records a strike. At the thresholds you configure in Settings, strikes escalate — a temporary posting pause, then a ban. The dashboard's Activity tab shows every member's count, and admins can reset strikes there. Warnings, withdrawals, and expired holds never add strikes.

Appeals (opt-in)

Turn appeals on in Settings and rejection or ban notices carry an appeal button (⚖️ Appeal this decision / ⚖️ Appeal this ban). The member writes their reasoning, and a fresh appeal card reaches your mod channel with ✊ Uphold and ⚖️ Overturn. Overturning reverses the strike (and lifts the ban, for ban appeals); for decision appeals it also withdraws what the AI learned from the original rejection, so reposting the cleared message won't trigger an automatic repeat-rejection. One appeal per decision; if no moderator reviews an appeal within 14 days — sooner on servers with a short log-retention setting, where the review window closes before the logs expire — it's closed as upheld and the member is told no one got to it in time.

Community flagging

Members can flag a concerning post by reacting with the emoji you configure. Enough independent flags send the post to your mod channel for review — with built-in rate limits so coordinated brigading can't weaponize it.

Edits are covered too

Editing a message re-runs the same screening as posting one — sneaking a violation in via an edit doesn't work.