The half nobody staffs
NeuroDialogs
Answer private messages with AI: how a run works, accounts taking turns, the limits that keep accounts safe, prompts and the inbox.
01What it does
NeuroDialogs answers the people who write to your accounts, for example after reading one of their comments. Replies are written by AI with your prompt.
Accounts are not online all the time. During a run they take turns: an account comes online, answers a few chats, and leaves; then the next one comes.
02How a run works
- Start opens a run. It lasts Run length (default 8 hours); the progress bar shows how far it is and when it ends. Then the module stops by itself. Press Start again for the next run.
- Accounts come online one after another, 20 to 60 seconds apart, never more than Accounts online at once (default 2). When one leaves, the next in line starts.
- Accounts with people waiting for an answer go first. After that, the account whose last turn was longest ago.
- In its turn an account reads its inbox, answers up to Chats per turn (default 3) conversations, browses a little (AI Protection) and goes offline. A turn lasts at most Turn length (default 8 minutes) and ends sooner when the chats are done.
- Messages from the last 3 days are answered, newest first. Older conversations nobody answered are backlog and are only answered if you switch on Answer the backlog.
- Accounts that are paused, still in their Supervise rest, or have not passed the account check are skipped.
03Set it up
Add accounts to the Dialogs Pool. An account works in one module at a time; one that is commenting must leave that pool first.
Open Persona and make a prompt active: copy a built-in or create your own. Attach a knowledge file if replies must state prices, dates or links correctly.
Check Sessions. The defaults are careful. On fresh accounts, keep New people per day low. Press Save settings.
Choose an AI Protection level (Medium is the default).
Press Start and watch Statistics: who is online, who is in line, how many people are waiting.
Start refuses an empty pool and a prompt that cannot be built (for example, one whose knowledge file was deleted). The error says what to fix.
04Sessions settings
Settings apply after Save settings, from the next turn. Nothing needs restarting. Presets (above Sessions) save these settings under a name; Load fills the form, then you save.
| Setting | Default | What it does |
|---|---|---|
| Run length | 8 h (1–24) | How long a run lasts after Start. |
| Accounts online at once | 2 | How many accounts can be in a turn at the same time. |
| Chats per turn | 3 (1–20) | How many conversations an account answers before it leaves. |
| Turn length | 8 min (2–60) | The longest an account stays online in one turn. |
| Pause before a reply | 40–180 s | Wait before each reply, with the typing indicator on top. Inside a turn it is capped at 60 s. |
| Context messages | 10 (0–50) | How many earlier messages of the chat the AI sees. |
| Skip chance | 15% | Chance of leaving a chat for a later turn, like a person who does not answer everything at once. |
| Reply language | Match the sender | Or Fixed, with a language you pick. |
Limits (0 means no limit):
| Limit | Default | What it caps |
|---|---|---|
| Replies per person | 5 | Total replies one conversation can ever get. Then the chat stops with “reply limit reached”. |
| Replies per session | 8 | Replies one account sends in one turn. |
| Replies per day | 25 | Replies per account per day. After that the account still comes online and reads, but writes nothing. |
| New people per day | 5 | How many strangers one account starts talking to per day. The most important limit; keep it low on fresh accounts. |
| Chats read per session | 15 | How many chats an account opens and reads in one turn (minimum 1). |
Group Promotion: list your groups (one per line) and set Promote every N replies (default 5; 0 = never). A reply mentions one of them at most once every N messages in a conversation. Groups allowed in Group Parser land in this list.
Safety (folded):
| Setting | Default | What it does |
|---|---|---|
| No links before N exchanges | 3 | Any link or @mention the AI writes is held back until the chat has had this many back-and-forths (one, if the person asked for a link). Enforced in code. |
| Auto-pause at block rate | 25% | Pauses an account when this share of its recent messages ends in a block or refusal. It stays paused until you press Resume. |
| Daily AI spend limit | 5 USD | For the whole pool. When reached, replying stops for the day. |
| Automatic replies | On | Off: accounts still come online and read, but write nothing. Control then shows “reading only”. |
| Typing simulation | On | Shows “typing…” for about as long as a person would take. |
| Stop after sending a link | On | No more automatic replies in a chat once a link has gone out. |
| Answer the backlog | Off | Also answer people who wrote more than 3 days ago and were never answered. |
| Never reply to these people | Empty | @username or numeric id, one per line or comma-separated. |
Two guards run without a setting: a reply too similar to what your pool recently sent is held back, and conversations about payment demands, threats, apparent minors or a crisis are never answered automatically; they wait for you in the inbox as “needs a human”.
05Prompts and model
A prompt is a name, the text the AI answers with, a maximum reply length and an optional knowledge file. The active prompt is used by every account in the pool.
- Tokens you can insert: {message}, {sender_name}, {message_language}, {context}, {account_id}, {account_username}, {account_phone}, {account_first_name}.
- Max reply length: 300 characters by default (40–1000). Shorter reads more like a real person texting.
- Knowledge file: .txt or .md, up to 512 KB. The first 12,000 characters go into every reply, so the AI states your real prices and links instead of inventing them.
- Model: the same five models as Neurocommenting, set separately here. Default Grok 4.3. The daily spend limit is counted in the chosen model’s real price.
06AI Protection
The AI Protection block under Control. While an account is online in its turn, it also reads channels and groups, scrolls, looks at posts and stories, likes and looks through its settings between replies. Low: about every 3–7 minutes. Medium (default): every 1–3 minutes. High: every minute or two. Off: the account only answers.
It never reads, marks read or archives private chats, so it cannot hide a message NeuroDialogs has to answer.
More: AI Protection across the modules.
07Inbox and statistics
| Account state | Meaning |
|---|---|
| online | In its turn right now. |
| in line | Waiting for its next turn. |
| retrying | The last turn failed (for example, it could not connect); it tries again at the time shown. |
| paused | Stopped by a safety check, with the reason. Press Resume on the row; it does not come back by itself. |
Each row also shows replies sent today, new people today, people who blocked it today, and how many people are waiting. A warning strip counts auto-paused accounts and conversations waiting on a human.
Conversations lists every chat. A stopped chat shows why: link sent, reply limit reached, blacklisted, they blocked us, needs a human, stopped manually, or Telegram service account.
If you write in a chat yourself, the AI stops answering it (“You’re handling this conversation”). Press Hand back to AI to let it continue; it will not re-answer what you already handled.
08Common mistakes
- Expecting every account to be online. Only Accounts online at once are, taking turns; that is by design.
- Forgetting that a run ends. After Run length the module stops; press Start for the next run.
- Raising New people per day on fresh accounts. It is the limit that protects them most.
- Switching on Answer the backlog on accounts with many old chats. They all get answered in a burst.
- Writing in a chat and waiting for the AI to continue. Press Hand back to AI.
- Leaving paused accounts. Read the reason, fix it, press Resume.