Dialogs: AI runs the conversation

Dialogs answer incoming messages on behalf of your accounts. You describe once who you are and what you want from the conversation — then AI runs the chats itself: every account keeps its own manner, all dialogs live in a shared inbox, and when a human is needed, AI raises its hand — and you take the conversation over with one click.

Launch flow

• Select accounts and write the conversation rules — or start from a ready template.
• Create a personality for every account and try replies in the sandbox.
• Launch and watch the inbox — step in when AI calls you.

Conversation rules

The rules are one instruction for all accounts: who you are, what you offer, the tone and the boundaries. So you never start from a blank page, there are templates for the usual jobs: 'Traffic to a channel', 'Selling access', 'Collecting leads' and 'Reviving a base'. A click inserts the full text — then you shape it to taste.

What the rules must contain

The model re-reads this text before every reply. The more specific it is, the less the model invents. Six blocks cover 90% of cases:

• Who you are. Not 'a manager' but a real person with details: what they do, how long they have been around, why they are writing.
• One goal. What has to happen by the end of the conversation — a signup, a deposit, a subscription, a lead.
• What to say about the product: 2–3 facts that can be repeated. Everything else the model will make up if you do not write it down.
• When to send the link. Without this the model drops it in the very first message and the conversation is over.
• Bans in plain words: what must never be promised, what must not be discussed, what not to do.
• When to call the operator: which questions the model must not decide on its own.

Example: casino traffic on a revshare deal

A working instruction in full — copy it and rewrite it for your own project:

• You are a regular player who has been into casinos for a couple of years. Not a manager, not support, not a salesperson.
• Goal: bring the person to a signup through the link and a first deposit.
• About you: you mostly play slots, lately on one project — fast payouts and decent cashback.
• You send the link only when the person asks where you play. Before that it is just a conversation.
• Never promise a win, never name a payout percentage, never say 'a big win is guaranteed'. If asked directly, say it is a lottery and money can be lost.
• Write no more than two sentences in a row. Emoji no more than once every three messages.
• If the person says they are under 18 — say goodbye and close the dialog.
• If they ask about withdrawals, demand guarantees or a bonus for a specific amount — call the operator.

Mistakes that stop rules from working

• A goal as a slogan: 'sell the subscription'. The model cannot tell what to do with an actual message. Write the action: 'bring them to a signup through the link'.
• No bans. The model fills the gap itself — inventing a discount, a deadline or a guarantee you do not have.
• A two-page instruction. By the end the model is repeating itself. Keep it within one screen.
• Nothing about when to stay silent. Then AI answers everything, including what needs a human.
• Testing the rules on real people. That is what the sandbox is for — nothing goes to Telegram there.

A personality per account

Twenty accounts giving identical answers expose automation instantly. So on top of the shared rules every account gets its own personality: who they are, what they live by, how they write. The panel generates personalities itself — from your rules, for each account separately — and any of them can be opened and edited by hand. The launch stays closed until every participant has a ready personality: that is the guard against a faceless choir.

The reply sandbox

Before letting AI talk to real people, write a line as the other person and see how a specific account would answer. The sandbox respects both the shared rules and the chosen account's personality, and nothing is sent to Telegram — so you can calmly tune the tone.

The inbox: every chat in one window

All conversations of the run sit in one list with filters: new, needs attention, with the operator, with AI, blocked. Open a dialog and you see the whole history; 'Take over' moves the conversation to you and AI goes silent, 'Return to AI' hands it back. When AI is unsure — the instruction said to stay silent, or no safe answer was found — the dialog is marked 'Needs attention' with the reason. Your reply goes out from the same account that ran the chat: the other person notices nothing. Tags, notes and reactions are at hand, and the 'History' tab lets you return to past conversations — the account will only write to the people you picked.

How much and how long to work

• A per-account limit: by replies or by working time — once spent, the account stops on its own.
• How many conversations at once: the rest wait in line; a slot frees up when the other person stays silent for half an hour.
• Conversation memory: 10 to 80 messages — more memory, sharper answers, pricier requests.
• Stop words and a 'do not reply' list — topics and people AI will not touch.

Together with Mailing

The strongest combo in the panel: Mailing makes the introduction and sends the first message, and the 'Hand over to Dialogs' mode brings every replier straight into the inbox. From there AI runs the conversation by your instruction, and you join at the decisive step — that is how one person comfortably holds hundreds of chats in parallel.