On top of the system alerts, you can create as many custom alerts as you want: Alerts → New alert (Your alerts section).
The two detection modes (you can combine them)
Trigger prompt: describe in natural language, in one sentence, what the AI should watch for. Examples: "The fan mentions a date or offers to meet in real life", "The fan says they are going through a hard time", "The fan mentions a big budget to spend". This prompt is evaluated by the AI on every fan message, with the context of the recent messages. Detection is deliberately cautious: the alert only fires when the criterion is clearly met.
Keywords: a comma-separated list of words (for example: meetup, address, phone). As soon as a keyword appears in the fan's message, the alert fires immediately, without going through the AI.
The "Stop the conversation" option
On top of notifying you, each custom alert can:
Stop the conversation: the AI pauses on the conversation, which then waits for human intervention. Save this for cases where you always want to take over (for example: the fan brings up meeting in real life).
Without this option, the alert is purely informative: the AI keeps handling the conversation.
Assignment per model
Like system alerts, your alerts are assigned model by model (Models → your model → Alerts tab). You can therefore have alerts specific to certain creators.
After creation
Remember to set up the alert's Delivery to receive it on Telegram: with no destination checked, it only appears on the Notifications page (see "Choose where your alerts go").
Each trigger shows the reason (the keyword found, or the AI's justification) and a direct link to the conversation.
Your alert fires too often? Make the prompt more precise and remove keywords that are too common: a keyword always fires the alert, without going through the AI. It does not fire enough? Describe the situation in other words, or add the words your fans actually use.