The half nobody staffs

NeuroDialogs

The module that answers private messages — how a session is shaped, what bounds it, and every setting on the page.

In this guide

01What this page is 02Map of the page 03Control 04Prompts 05Rhythm 06Replying 07The four limits 08Safety 09Group promotion 10Conversations 11First run

01What this page is

NeuroDialogs answers the people who write to your accounts — someone who saw a comment, opened the profile and sent a message. It is the only module in the system where a real person is on the other end and can decide to press report.

Everything about the page follows from that. The accounts are not online waiting; they come online for a while, read what arrived, answer some of it at human speed, and go offline again.

The obvious design — keep every account connected and reply the moment a message lands — was rejected on purpose. An account that is online around the clock and answers within two seconds at four in the morning is not a person, and the pattern is visible from outside. Almost every default on this page exists to break that pattern, which is why a correctly running pool looks idle most of the time.

02Map of the page

Five regions, with a jump-nav across the top in this order.

01ControlStart and Stop, five counters, one row per account with what it is doing right now, and the live log underneath.
02Dialogs PoolTwo columns of accounts — available and in the pool — the same shape as the commenting pool, and the same one-module-at-a-time rule.
03PromptsPrompt presets, which one is active, the reply-length slider and the knowledge file attached to each.
04SessionsEverything about rhythm and restraint: Rhythm, Replying, Limits, Group Promotion, and a folded Safety group.
05ConversationsEvery thread, searchable, with the full exchange and a box to write in it yourself.

03Control

Start does not start a conversation with anyone. It puts the pool into service; each account then schedules its own first session and comes online when its turn arrives.

IllustrationNot the live panel — nothing here is connected: no state, no saving, no requests. Click a control to read what it does.

StartControl

What it does: Marks the module enabled and hands the pool to the runner. Accounts come online on their own schedule from that point, never all at once.
What refuses it: An empty dialogs pool, and an active prompt that cannot be assembled — one pointing at a knowledge file that has since been deleted, for instance. Both refuse with a message naming what to fix, rather than reporting a running module that does nothing.
How many at once: Three accounts online at a time, by default. A larger pool does not mean more simultaneous sessions; it means the turns spread further apart.
Never at night: Sessions only happen inside an account’s own waking hours. An account whose local time is the small hours does not come online, whatever the gaps say.
What a fresh start looks like: Nothing. Zero online, zero replies, for hours. The panel says so in its own words under the counters, because a correct run and a broken one look identical otherwise.

StopControl

What it does: Sessions in flight finish and their accounts go offline. Nothing new is scheduled.
Never gated: Stop always works, whatever the subscription says.

An account rowControl

The four states: Online — in a session right now. Waiting — awake, with the next session due at the time shown beside it. Asleep — outside its own waking hours. Paused — pulled out by the safety layer, with the reason next to it.
The three counters: Replies sent today, new people this account started talking to today, and how many of those people blocked or refused it today. A waiting badge appears when someone is unanswered.
Paused is the only one to act on: The other three are ordinary. A paused account has a Resume button on its own row and does not come back without it.

reading only — automatic replies are offControl, beside the state

What it means: The module is started and the accounts are working, but Automatic replies in Safety is off, so nothing is written.
Why that is useful: It keeps the accounts warm and the inbox current while a prompt is being reworked. It is a real operating mode, not a broken one.

Five counters sit above the account rows: accounts in the pool, online now, conversations, unread, replies today. Daily AI spend is deliberately not among them — it is a number nobody watches, and the limit that protects it works without anyone looking.

04Prompts

A prompt preset is a name, a template you write, a maximum reply length, and optionally a knowledge file. The active one is what every account in the pool speaks with.

IllustrationNot the live panel — nothing here is connected: no state, no saving, no requests. Click a control to read what it does.

Prompt templatePrompts → editor

What it does: The instruction the model answers under. Written as plain text, with tokens substituted before each call.
The tokens: Eight, offered as buttons under the box: {message}, {sender_name}, {message_language}, {context}, {account_id}, {account_username}, {account_phone}, {account_first_name}.
Substitution: Plain text replacement, like the commenting persona. Braces that are not a known token are left alone.
What is appended for you: The length instruction is added to the end of every prompt automatically, so there is no need to repeat it in the text.

Max reply lengthPrompts → editor

What it does: Caps the reply, in characters, as an instruction appended to the prompt.
Default: 300 characters.
Range: 40 to 1000. The lower bound is real, not a formality — a 40-character cap forces the kind of terse message a person actually sends, and anything below it produces fragments.
Why shorter is usually right: A long, well-structured paragraph arriving in a DM from a stranger reads as generated. Short reads as typed.

Knowledge filePrompts → editor

What it does: Attaches a document whose contents go into the prompt, so the model states your prices, dates and links instead of inventing them.
Accepted: Plain text and Markdown — .txt, .md, .markdown — up to 512 KB. Anything else is refused rather than accepted and fed to the model as noise.
How much is used: The first 12,000 characters. A longer file is cut at a line break near that point and marked truncated in the picker, so it is visible which files are only partly in play.
Not a search index: The whole file goes into every generation. That is deliberate at this size, and it is also why the cut exists — the text is re-sent on every single reply.

Deleting a knowledge file that a preset still points at does not fail quietly. The preset stops assembling, and the module refuses to start until it is fixed — which is better than accounts confidently answering questions with nothing behind the answer.

05Rhythm

When an account comes online is demand-driven, not a timer. An inbox with people waiting pulls the next session in; an empty one lets it drift out. That is why there are two gap ranges rather than one interval.

IllustrationNot the live panel — nothing here is connected: no state, no saving, no requests. Click a control to read what it does.

Gap between sessions — inbox emptySessions → Rhythm

What it does: How long an account stays offline when nobody is waiting for an answer.
Default: 120 to 300 minutes — two to five hours.
Picked how: A fresh random value inside the range after every session, not a fixed cadence. A constant interval is a metronome in the traffic pattern.

Gap between sessions — people waitingSessions → Rhythm

What it does: The same thing, for an inbox with unanswered people in it. Shorter, because a lead that waits five hours is usually gone.
Default: 20 to 60 minutes.
The trade: Shorter converts better and looks less human. This pair is the main dial between the two.

Session lengthSessions → Rhythm

What it does: How long an account stays online per visit.
Default: 20 to 40 minutes.
Still runs when capped: An account that has hit its daily reply limit comes online anyway: it reads, marks things read and writes nothing. That is a real state, not a wasted session.

Max extensionSessions → Rhythm

What it does: Extra minutes a session may run past its planned end while the other person is still actively replying.
Default: 20 minutes.
Why it exists: A person pulled into a live conversation does not stop mid-sentence because a timer expired.
Zero: Turns the extension off — sessions then end exactly on their planned length.

06Replying

Two different delay distributions, because a cold first reply and a follow-up inside a running conversation are not the same act. One range for both produced either implausibly fast strangers or uselessly slow conversations.

IllustrationNot the live panel — nothing here is connected: no state, no saving, no requests. Click a control to read what it does.

First reply to a strangerSessions → Replying

What it does: How long before the first answer to someone new.
Default: 300 to 2700 seconds — five to forty-five minutes.
Why so wide: An instant answer to a cold DM is the clearest bot tell there is. The width matters as much as the length: a consistent delay is its own signature.

Reply inside a live conversationSessions → Replying

What it does: The delay between messages once a conversation is already running.
Default: 40 to 180 seconds.
Fast on purpose: That is what a person engaged in a chat actually does.

Context messagesSessions → Replying

What it does: How many previous messages of the conversation are sent to the model with each reply.
Default: 10. The field accepts 0 to 50.
At zero: Every reply is written with no memory of the conversation.

Skip chanceSessions → Replying

What it does: The chance of deliberately leaving an answerable conversation for the next session instead of answering it now.
Default: 15%. The slider runs from 0 to 90%.
Why: Nobody clears their whole inbox every time they open it.

Reply languageSessions → Replying

What it does: Either answers in whatever language the message arrived in, or pins one language for every reply.
Default: Match the sender.
Fixed: Reveals a second picker for the language itself.

Two more things happen without a setting: consecutive incoming messages are folded into one reply rather than answered one by one, and the typing indicator runs for roughly as long as a person would take, if typing simulation is on.

07The four limits

Four independent numbers, because no single one expresses the risk. Zero means no limit on any of them.

IllustrationNot the live panel — nothing here is connected: no state, no saving, no requests. Click a control to read what it does.

Replies per personSessions → Limits

What it does: The total number of replies one conversation may ever receive.
Default: 5. It is a lifetime count per conversation, not per session or per day.
What it prevents: One person being pestered. A conversation that reaches it is closed with the reason shown in the inbox.

Replies per sessionSessions → Limits

What it does: The most one account may write in one visit.
Default: 8.
What it prevents: One sitting turning into a blast.

Replies per daySessions → Limits

What it does: One account’s total daily exposure.
Default: 25.
On reaching it: The account still comes online and reads for the rest of the day. It simply writes nothing.

New people per daySessions → Limits

What it does: How many strangers this account starts talking to in a day.
Default: 5.
The important one: Unique non-contacts messaged is the closest available proxy for what actually trips Telegram’s own flood protection. Of the four, this is the one to keep low on fresh accounts.

Chats read per sessionSessions → Limits

What it does: How many conversations an account opens and reads in one visit.
Default: 15. The minimum is 1 — unlike the four above, this one has no unlimited setting.

08Safety

The group is folded by default, with its current values written along the header so nothing is hidden by being shut. Everything in it is enforced in code — a prompt instruction is not an enforcement mechanism, and models break instructions regularly.

IllustrationNot the live panel — nothing here is connected: no state, no saving, no requests. Click a control to read what it does.

No links before N exchangesSessions → Safety

What it does: Withholds any link the model writes until the conversation has been through this many back-and-forth rounds.
Default: 3.
What counts as a link: More than http addresses: t.me and tg:// forms, bare domains, and an @mention too — an @channel funnels exactly like a link and carries the same risk sent unprompted.
The exception: If the person explicitly asked for it, one round is enough.
Why in code: A link in the first message to a stranger is the fastest route to a spam report. The prompts all say not to; the gate is what makes it true.

Auto-pause at block rateSessions → Safety

What it does: Pauses an account once this share of its recent sends comes back as a block or a privacy refusal.
Default: 25%.
Why it is the number to watch: Recipients blocking an account is the earliest externally visible sign it is heading for a ban — days before the ban itself.
After it fires: The account appears paused in Control with the reason beside it and stays out until you resume it by hand.

Daily AI spend limitSessions → Safety

What it does: Caps what the whole pool may spend on generation in a day, in dollars.
Default: 5.
On reaching it: Replying stops for the rest of the day. This is why the number is not a counter on the Control panel — it works without being watched.

Automatic repliesSessions → Safety

What it does: Off keeps accounts coming online and reading without writing anything.
Default: On.
Where it shows: The Control panel wears a reading only badge for as long as it is off, so the state is never silent.

Stop after sending a linkSessions → Safety

What it does: Ends automatic replying in a conversation once a link has gone out.
Default: On.
Why: The link was the point of the conversation. Continuing past it is where an answer turns into pestering.

Answer the backlogSessions → Safety

What it does: Answers unread messages that arrived before the module was started, rather than only marking them read.
Default: Off.
Why off: A pool started on dozens of stale conversations produces exactly the burst of outbound messages that gets accounts banned — the one thing the rest of this page exists to prevent.

Fill empty sessions with warmup actionsSessions → Safety

What it does: With nothing to answer, the account scrolls, reads and uses its saved messages instead of sitting online doing nothing.
Default: On.
Which actions: Active Warmup’s own action library. This does not enrol the account in that module — it borrows the behaviour for the length of the session.

Never reply to these peopleSessions → Safety

What it does: A list of people no account will ever answer.
Format: An @username or a numeric id, one per line or comma-separated.

Two more guards run without a setting of their own. Outgoing text is compared against what the pool has recently sent, and a reply too close to one already used is withheld and rewritten next session — one preset across a hundred accounts otherwise converges on identical phrasing, which is the textbook signature of a spam network. And a short, deliberately narrow set of incoming conversations is never answered automatically at all: payment demands, accusations and threats, apparent minors, anything crisis-shaped. Those are handed to you in the inbox instead.

09Group promotion

Separate from the link gate, and often confused with it. The gate decides whether any link may go out yet; this decides how often, once past the gate, a reply also mentions one of your groups.

IllustrationNot the live panel — nothing here is connected: no state, no saving, no requests. Click a control to read what it does.

Promoted groupsSessions → Group Promotion

What it does: The groups a reply may promote. One per line, as @group or a t.me link.
Empty: The feature is simply off — no group instruction is added to the prompt at all.
Where else this is written: Group Parser’s promote action writes into this same list, so a group promoted from there appears here.

Promote every N repliesSessions → Group Promotion

What it does: At most one group mention per this many outgoing messages, counted per conversation.
Default: 5.
Zero: Never promote, whatever is in the list above.
Enforced where: In code, when the prompt is composed — the model is not asked to limit itself.

10Conversations

Every thread the pool has, searchable, with the exchange on the right. A conversation that has stopped says why.

Reason it stopped	What it means
link sent	A link went out and Stop after sending a link is on.
reply limit reached	The conversation hit Replies per person.
blacklisted	This person is on the never-reply list.
they blocked us	The recipient blocked the account.
needs a human	The incoming screen caught something that must not get an automated answer.
stopped manually	You wrote in it yourself.
Telegram service account — never answered	Telegram’s own service messages. Never answered, by design.

IllustrationNot the live panel — nothing here is connected: no state, no saving, no requests. Click a control to read what it does.

The reply boxConversations

What it does: Sends a message as that account, from you.
What it costs: The conversation. Writing in it takes it over — the engine stops answering that thread automatically until you hand it back.
Why: Two authors writing into one chat minutes apart, possibly contradicting each other, is worse than silence.
Handing it back: A resume action on the thread returns it to automatic answering.

The warning strip at the top of Control counts two things worth acting on: accounts auto-paused for safety, and conversations waiting on a human. Both are the kind of thing that will not resolve itself.

11First run

The defaults on this page are already the cautious configuration. On a fresh pool, most of it is worth leaving alone.

Put accounts in the dialogs pool. The one-module rule applies here as everywhere: an account that is commenting has to leave that pool first.
Pick or write a prompt. Duplicate the closest built-in rather than starting from an empty box, and attach a knowledge file if there are prices, dates or links the answers must get right.
Leave Rhythm, Replying and Limits at their defaults. They are the conservative setting already, and New people per day is the one to lower rather than raise on fresh accounts.
Open Safety once and read it. Everything in it is on by default except Answer the backlog, which should stay off on a pool that has any history at all.
Press Start, then leave it. The first sessions are hours away, and nothing being online is the expected state.

After that, the two things worth checking are the paused count and the blocked counter on each account row. A rising block count on one account is that account being disliked; a rising count across the pool is the prompt.

Saving settings never needs a restart — they apply from the next session. The prompt is the exception in the other direction: the module refuses to start at all on a prompt that cannot be assembled, so a broken preset is caught at the button rather than discovered in someone’s DMs.