How to talk to the builder

The builder isn't a help page or a search through a manual. It does the work itself: it creates the agent, writes its character, reads your website, connects channels and actions. By the time you've finished reading its reply, the work is already done.

Use plain words

You don't need to guess the "right" wording or name internal concepts. Just say what you want.

This works:

Rename it to “Smak Assistant”

Make it answer briefly and politely

Read my website smak.example.com

Make it answer in the customer's language

Add that we're open from 9 to 20, and until 17 on Sundays

I want it to be able to book a table for a customer in our system

Every morning at 9:00, send me a summary of the past day

This doesn't work, and you don't need it:

Open the “Knowledge” section and click “Add material”

There is no “Knowledge” section in the dashboard — nor any other form where you set up the assistant by hand. Just say what you want to end up with.

It asks one or two questions, not a questionnaire

If something is missing, the builder will ask — but briefly, and only about what matters. Where a detail is minor, it doesn't ask again: it picks a sensible value and says it out loud:

Set it to 9:00; the summary covers the conversations of the past day. Tell me if you want it different.

This is on purpose. A request that stays a question disappears from the conversation, and you find out about it when the morning emails don't arrive. An assumption is visible right away, and fixing it takes one sentence.

And when a question has a few ready answers — “where should it answer: on your website or in Telegram?” — buttons with the options appear under it. Tapping a button is your answer: you don't need to retype the word from the question.

Changes are applied as the conversation goes

This is the most important thing to understand about the dashboard: changes don't pile up until the end, and there is no “Save” button.

One turn looks like this:

1. You send a message.
2. The builder starts replying — the text appears as you watch.
3. It gets to work, the text pauses, and you see “Doing what you asked…”.
4. Below, cards for finished work appear one after another: “Renamed.”, “Source “Opening hours and address” added to the knowledge base.”, “Pages read: 12.”
5. The panel on the right re-reads itself — the numbers change.
6. The builder finishes its reply: what came out and what's next.

A card appears when the change has already been applied, not when the builder promised it. The text on the card is word for word what the tool itself reported, not a retelling.

The order on screen matches what actually happened

While a turn is running, the builder's reply sits above the cards — exactly as it happened: first “let me take a look”, then the work. When the turn ends, the final reply moves below the cards: it describes what has already been done.

What it does right away, and what it asks about

This isn't a whim of the model — it's a platform rule, and it's always the same.

Does right away, and tells you right away what it did:

• the agent's name;
• character: role, tone, language, greeting, suggested questions, refusal text;
• the look of the chat window on your website: title, color, position, theme;
• new knowledge;
• an action that reads someone else's system (checks, looks something up);
• a key saved under a new name;
• turning a trigger on or off;
• how long an action waits for a response.

Asks and waits for a clear “yes”:

What	Why it asks
a change to the agent flow	customers will see the change right away
switching on a raw instruction instead of the assembled character	customers will see the change right away
rewriting the text of an existing topic in knowledge	customers will see the change right away
changing a trigger's schedule	customers will see the change right away
connecting or disconnecting a channel	customers will see the change right away
an action that writes to someone else's system	the result will land in someone else's system
testing an action for real, running a trigger now	the result will land in someone else's system
connecting an external server (MCP)	the result will land in someone else's system
inviting a person to the dashboard	the result will land in someone else's system
removing an action, channel or trigger	this can't be undone
saving a key over an existing one	this can't be undone
adding a skill from the internet, or a draft the assistant wrote after reading someone else's text	this is someone else's text, and the agent will act on it from now on
canceling a subscription, revoking a builder key, removing a person from the team or changing their role, deleting an agent, reissuing its key, erasing a visitor's data	this can't be undone
replying to a customer on your behalf, sending or correcting a training-mode draft, approving or rejecting an agent's action	this goes to another person — and in your name
approving an app's connection by code, signing out on all devices	the result happens outside the conversation
anything on this list, and smaller things next to it (a key, a payment, an archive), if in that same turn it read someone else's content — a website, a search, a customer's words	you didn't write that text, so it won't act on it without you

When it asks, the builder says in its own words what exactly will happen and why it's asking — not just “please confirm”.

A key in your message is already consent

One exception: if you sent the key or token itself, the builder doesn't ask again — it connects right away. A secret lives exactly as long as the turn it was written in, so a second question would cost you sending the key once more.

“Yes” works the first time

The platform remembers exactly what the builder asked, and after your “yes” it does exactly that — without rebuilding the request or asking a second time. Consent to replace a topic's text is one consent, however it carries it out; consent to remove several sources also covers removing some of them. It will only ask again about something it hasn't asked about yet.

If you talked about something else between the question and your “yes”, the “yes” no longer applies to it — it will ask again.

“Uh-huh” doesn't count as consent

Consent has to be clear and given to this particular question. A nod given to an earlier question doesn't count as consent.

If it did something without asking, that means it can be undone with one word.

When it shows a flow change, it shows the price too

The agent flow is the sequence of steps that make up one of its replies. If you ask for a change that makes every conversation more expensive (for example, one model writes the reply and another checks it), the builder tells you before the change, together with the question:

This graph makes up to 2 calls to the model per turn and takes up to about 40 s.

So you see the cost while you can still say “no”. If the flow adds nothing to the price of a conversation, money won't come up: the question comes when there's something to ask about.

It will also warn you separately if the planned setup won't manage to answer a live visitor in time — nobody waits for minutes in a chat, and work like that belongs in a schedule, not in a conversation.

When something goes wrong mid-turn

If a turn breaks off, you'll see a red banner with the reason and — most importantly — a hint on what to do:

• “What managed to apply is still there — see above. Don't ask for the same thing again: the finished part would be repeated too.” Part of the work is done. Repeating the request would create a second agent or send a second email. Look at the cards above and ask only for what's missing.
• “Nothing was changed — you can say the same thing again.” Safe to repeat.
• “Looks like the connection dropped. Nothing was changed.” Same as above.
• “Too many messages in a row — wait a moment.” Just a pause.

If the top says “The builder is unavailable right now”, there's nowhere to write; the conversation will be back when the builder is.

What it won't do

• It won't get around your permissions. If your role is operator, it can't create an agent — and it will say so plainly instead of staying silent.
• It won't make up facts about your business. It knows hours, prices and terms only from your words or your website. If you haven't said, it will ask.
• It won't go into someone else's dashboard. It works only where you're talking to it from.
• It won't show a saved key. It can save a key but not read it — not even for itself. Lost it? Issue a new one in the service where you got it.
• It won't take a password in the conversation. It prepares changing your email or password, or deleting your account, and gives you a link to a screen where you type the password yourself. If you did write a password in the conversation, change it.
• It won't reply to a customer or approve an action on its own. Only when you directly ask about a specific conversation or card — and it first shows what exactly will go out.
• It won't do anything with money, keys, the account, the team or “Conversations” for someone else's program with a key. This is only for a person signed in to the dashboard.
• It won't say “done” if it didn't happen. A report about work that wasn't done counts as worse than a refusal here.

The dashboard in words

Anything in the dashboard can be asked of the builder — not just things about the agent. A key for your program, your plan and credits, email and data archive, the team, deleting an agent, replies to customers in “Conversations”: it does it itself and shows you the result, instead of sending you off to look for a button.

Three things stay the same.

1. This is only for you. These requests are carried out only for a person signed in to the dashboard, and with the same permissions as the buttons: the owner handles payment and the team; keys and action approvals are for admins. Someone else's program with a key — the agent's secret key or a builder key — can't buy a plan, issue itself a key, change your email or reply to a customer on your behalf.
2. Anything irreversible, and anything involving money, waits for your “yes”. The builder says what exactly will happen and asks with “Yes” / “No” buttons. Consent covers exactly what it asked about: a “yes” for one person, one card or one text doesn't allow anything else.
3. Passwords don't go into the conversation. For anything confirmed with a password, it prepares the step and gives you a link to a screen where you type the password.

Builder key. “Give my Claude full access” — the builder issues a key, shows it once, and gives you a ready-made setup for your program (Claude Code, Cursor, Claude or any other app that supports MCP). With this key the program can do with your agents the same things you do in the conversation with the builder. Money, the account, the team, new keys, replies to customers and action approvals are off limits to it. After you refresh the page, the key is no longer in the conversation — only a note that it was issued remains. To revoke: “revoke the builder key” — all such keys stop working at once.

This isn't the same as an access key to an agent

“Give my Claude access to this agent” is a different key: with it, the program can only talk to one agent, like a visitor. “Full access”, “configure”, “manage” — that's a builder key. If it's unclear which one you need, the builder will ask.

Workspaces. A conversation belongs to the workspace where it was started. When you ask to switch to another workspace, create your own or leave a team, the builder does it, the dashboard switches on its own, and a line above the input field shows where you are now. The previous conversation stays in its workspace — go back there and you'll find it.

A file in the conversation. Attach a file with the paper clip and say what to do with it. The file itself stays in your browser: the builder tells the dashboard where to put it, and the dashboard puts it there — that's why the builder says “putting it” rather than “added”, and how it ended is shown in a line above the input field. The buttons under the file (“Learn it”, “Send to customers”, “Make actions from it”) do the same without words.

It works with the agent the panel shows

If you have several agents, the builder talks about the one whose name is shown in the panel. To switch, tap another agent's name among the chips at the top of the panel.

When it creates another agent, the builder does not jump the panel over to it on its own: you were just talking about a different one.

Several chats about one agent

You can have more than one conversation about an agent. Above the conversation there are two buttons:

• New chat starts an empty conversation about the same agent. The previous one doesn't go anywhere.
• Chats shows all your conversations about this agent, with the most recent at the top. If there are several chats, the button shows their number, and the open chat's name appears next to the agent's name.

In the list you can:

• open a chat — and continue where you left off; if the builder left a question with buttons, the row is marked “awaiting your reply”;
• rename it (pencil) — by default a chat is named after your first request in it;
• delete it (trash can) — the chat moves to “Deleted” at the bottom of the same list, stays there for 30 days, and can be brought back with “Restore” all that time. After that it's gone for good.

Good to know:

• The agent is the same in all chats. A chat is just a conversation. What's done in one chat applies everywhere, and in a new chat the builder knows the agent just as well: it looks at the agent itself, not at the previous conversation.
• Deleting a chat doesn't undo what was done in it. Changes the builder made there stay; you can roll them back with the word “undo” from any of your other chats, or in the change log.
• Chats are personal. In a team, everyone sees only their own conversations with the builder — the owner too.
• The conversation you had before appears in the list as the first chat. It has no time next to it: the platform didn't record when you last talked in it.
• Up to 50 chats per agent (deleted ones don't count). Hit the limit? Delete one you don't need.
• You can say “let's start over”, but a new conversation is started with the button: the builder replies in the chat where you asked, and it will tell you so.

The dashboard remembers which chat you were in: after refreshing the page, or after switching to another agent and back, the same chat opens. A new tab opens the chat you used last.

“Back” in the browser

Anything that opens on top of the conversation — a sheet from the panel, the chat list, a confirmation window, and on a phone the panel itself and the menu — closes with the browser's “back” button or the “back” gesture on your phone. It doesn't take you out of the dashboard. “Forward” opens the sheet again, and refreshing the page keeps an open sheet open. “Back” doesn't erase unfinished text in an edit window: the window asks whether you really want to close it.

A few phrases worth knowing

Say	And it will
“undo” / “put it back the way it was”	undo everything it did in the last turn
“put it back the way it was yesterday”	show restore points and ask which one to go back to
“the bot isn't working”	first check what's broken, and only then change anything
“it gave a silly answer”	look at the log of recent turns and tell you the reason
“what came out?” / “what's missing?”	re-read the conversation and name first what you asked for and didn't get
“read my website”	read pages from the sitemap and via links, and tell you in numbers: how many it found and where, how many it read, what didn't open and why — see What the agent knows
“add another page …” / “read only this page”	add just that page; only that page will be re-read too
“re-read the website now” / “when did you last read the website?”	re-read without waiting a week — or tell you when it read it, when it will next, and which pages disappeared or changed
“don't re-read this page” / “don't re-read the website”	keep the page (or the whole site) as it is now
“take this page from the website” / “keep it as I wrote it”	decide what happens to your edit when the page on the website has changed; your edit never disappears without your word
“remove the missing pages”	remove from knowledge the pages that are no longer on the website (until then they're only hidden from answers)
“let it ask me if it's not sure”	teach the agent to ask you questions: the question appears in the panel and in a notification, and after you answer, the agent continues on its own
“make the window available in English too”	translate the greeting, suggestions and your other window texts, and tell you what was left untranslated — window languages
“don't show the window on /app pages”	remove the chat button and window from your site's editor or admin area — where the window doesn't appear
“make the chat a panel on the right so it doesn't cover the site”	the chat becomes a full-height side panel and pushes the page aside instead of covering it — chat as a side panel
“put “Assistant” on the button and use my icon”	the chat button becomes a pill with an icon and a label — button with a label
“keep a list of what you find”	give the agent a notebook — cards with deduplication that you see in the panel (“Collected”), see What the agent collected
“go through all the venues / every listing”	add a step that goes over each item; when there are hundreds — in batches, from the notebook
“no more than 15 a day from each page”	set the limit for each account separately, not for the action as a whole
“don't write to people who opted out”	turn on the stop list; a person who changes their mind can ask to be contacted again
“show me first, then send”	put the action on approval: nothing happens until you click
“connect my email: the address is this, the password is that”	connect the mailbox; for well-known services it figures out the server itself
“connect my email, but don't reply to anyone yourself — just read”	connect a mailbox that only reads: emails don't wake anyone and stay unread; the agent searches and reads them when you ask
“every morning at 8 — a summary of overnight emails”	set up a schedule: the agent looks through new emails itself and writes to you who they're from and what they're about
“send emails yourself, don't ask me”	emails on your behalf go out without an approval card; it asks for consent. After reading someone else's email in the same conversation, approval is still required
“let me know when an email comes from …”	set a trigger on a new email from that sender
“don't transcribe voice messages”	turn off voice recognition: a person who sends a voice message gets “please type your message”; “transcribe voice messages again” turns it back on (Voice messages)
“make an image for a post” / “draw a sale banner in our website's colors”	give the agent the ability to draw and tell you the price; it asks you to name your website's colors or takes them from the chat window — see Images and voice
“read this text aloud” / “answer me with voice”	give the agent text-to-speech; “me” means only your personal chat, “everyone” means everyone in Telegram and WhatsApp; “answer with text” turns it back (more)
“turn on training mode” / “show me the replies first”	a person in “Conversations” sees the assistant's replies to customers first; “send it yourself if I haven't replied in 5 minutes”, “turn off training mode”, “summarize the corrections” (more)
“show contacts” / “how are replies rated” / “where can I see that a customer is unhappy”	list the contacts customers left and the ratings; in “Conversations” the same shows in the conversation itself and as marks in the list — see What the customer left in the window
“connect Viber” / “connect WhatsApp”	tell you step by step where to get the token, and connect it
“my email password changed” / “here's the new bot token”	replace the access in the same channel — without disconnecting and without a new address
“turn off Telegram for a week” / “while the bot is off, tell people: we're on vacation until the 10th” / “keep the bot silent while it's off”	turn the channel off without disconnecting it; anyone who writes gets one sentence (the standard one, yours, or none), once per half hour and without spending credits (more)
“try connecting again”	re-subscribe the messenger to the agent's address if it didn't work the first time
“keep only the last 90 days in the set”	remove from the set everything the agent hasn't touched for longer; asks for consent
“throw away this set”	remove the whole set — this is how you get under the limit on sets; asks for consent
“show what you'd do with “I want two pizzas”” / “do a rehearsal”	run a rehearsal: the agent answers for real, but only records each outside action — no order is placed, no email goes out. You see both the reply and the “would have done” list
“show what the morning summary will do”	rehearse the schedule's background turn — without waiting for nine o'clock and without sending anything
“put a consultant in my online store”	give you a line of code; the window sees the product the person is viewing, its card and the “add to cart” button on its own — see Chat window
“here's a link to our products: …” / “let it know all our products”	look for the product file on your website itself (or take the link you gave), read it, and keep re-reading it on its own; the assistant searches it during conversations — see Store products
“here's a link to my calendar” / “connect my iCloud calendar”	explain where to get the secret address or app password, connect the calendar, and tell you what the assistant can do with it: only see busy times, or also book (with your permission) — see Calendar
“I work from 9 to 18 on weekdays”	the assistant looks for free slots only within these hours
“Here's a skill from the internet — add it: …”	read the SKILL.md, tell you what the skill will make the assistant do and, separately, the risky parts; it adds it only after your “yes”. Scripts from the skill are not taken or run — see Skills
“What skills does my assistant have?” / “Forget the skill about reports”	list skills and drafts; remove a skill (“undo” brings it back)
“I want customers to pay right in the chat” / “connect LiqPay”	give you steps for where to get LiqPay keys and check them; from then on the assistant gives payment links for a product, order or service itself — see Payment in the conversation
“a consultation costs 500 UAH” / “no more than 5,000 at a time”	save the price list and the cap per link: a larger amount will wait for your approval
“so I don't end up with a pile of invoices” / “no more than three links per person”	set a limit on unpaid links per conversation per day (5 by default); paid ones don't count — see Payment in the conversation
“who paid?” / “latest payments”	show the latest payments with their status; it doesn't make refunds — only the payment service's dashboard does
“let it answer where an order is” / “I use WooCommerce (Horoshop, Prom)”	recognize the platform from your website and tell you step by step where to get the key in the admin panel; the assistant shows the status only to someone who gave the order number and phone number from the order — see Where's my order
“is the consultant selling?” / “how many clicked add to cart?”	give you the window's funnel: how many opened it, asked, clicked “add to cart”, went to a messenger — see Is the consultant selling
“what did people search for and not find?”	list customer searches with no matches in the catalog — see Searched, not found
“I want to know when a customer asks for a person” / “where will requests reach me?”	tell you where requests go now — to your chat with the assistant, to the on-duty person in Telegram, to email — and say plainly if they go nowhere but the dashboard; from a chat you reply to the customer by number: “123456: text” — see Conversations
“don't disturb me from 22 to 8” / “remind me after 5 minutes if nobody took it” / “don't email me about requests”	set quiet hours (during them the customer is told when they'll get a reply), change the time of the single reminder, turn off emails to the owner
“send me the gaps every Monday by email” / “a summary every day” / “send the summary now” / “don't send the summary”	change when and where the summary of unanswered questions and everything waiting on you arrives, or send it now; it costs no credits — see Summary
“when you don't know — give them my Telegram”	put buttons with your messengers under the replies; the person's question will already be filled in the message
“I want my own assistant like OpenClaw” / “an assistant for myself in Telegram that tells me every morning what I have today”	set up a personal assistant step by step: ask for the bot token and give you a code, set up a morning summary, and email and calendar if you want; at the end it tells you what's set up and what's missing — see Personal assistant
“write to me every morning, but if there's nothing new — don't”	add a condition to the schedule: on a morning with no news nothing arrives (more)
“this is my personal bot — let it answer only me”	give you a one-time code for your chat with the bot; after that the bot knows the owner is writing, and answers strangers with one sentence at no cost (more)
“what does my bot remember about me?” / “forget about my old job”	tell you the assistant's notes about you (and name separately the ones saved after reading someone else's text) / forget one; “forget everything about me” — asks for consent (more)
“keep my ideas journal” / “save this file for later”	check whether the assistant can run code and add that ability if not; from then on the assistant puts files in its persistent folder itself — only when you write, not a customer (more)
“what's in your folder?” / “delete that file” / “clear the whole folder”	list the files and how much space is used / delete one / delete all — asks for consent; to download a file, use the “Agent files” row in the panel
“let my wife write to it too” / “disconnect my Telegram”	give you a code for one more person / disconnect a person (not the bot itself)
“let customers continue the conversation in Telegram”	connect a Telegram bot — “Continue in Telegram” appears in the window, and the bot knows what was discussed on the website (more)
“make the window square-cornered and add our logo” / “make it completely my own”	change the window's look — from a preset or with your own style; you don't need to re-insert the line of code on your website
“give my Claude access to this agent” / “make a key for Cursor”	issue an access key (shown once) and a ready-made setup for your program; the program will talk to the assistant like a visitor, and its calls will spend your credits — see Agent as a tool
“there are duplicates in knowledge” / “a source got added twice” / “the text in a topic is doubled”	show which topics are the same and where text repeats, and after your word remove the copies (one stays from each group) — see Duplicate sources
“who has access to the agent?” / “take away Cursor's access”	list issued keys / revoke a key — the program loses access immediately; asks for consent
“make the agent visible to other agents” / “describe the agent for others: …”	open the assistant's public card (there isn't one by default) / save the description other programs see
“let it post from the client's account”	connect someone else's account by the service's address; you'll just need to send the client the consent link
“give my Claude full access” / “so Cursor can configure my agents”	issue a builder key (shown once) and a ready-made setup for your program; it tells you what the key can and can't do — see below
“revoke the builder key” / “Claude showed a code — approve it”	revoke all builder keys / tell you which program is asking and approve it — both after your “yes”
“I want Pro” / “how much does it cost” / “buy more credits”	show plans and packs with prices, open payment and give you a link to the payment page — see Credits and plan
“how many credits are left” / “I paid but my plan didn't change” / “cancel my subscription”	tell you the balance and which agents the credits went to / check the payment with the payment service / cancel auto-renewal — asks for consent
“export all my data” / “is the archive ready”	order an archive and give you a download link as soon as it's ready — see Your account
“change my email” / “the confirmation email doesn't arrive” / “delete my account” / “sign out of all devices”	resend the email and sign out of all devices itself; it prepares an email change or deletion and gives you a link to a screen where you enter your password — the password doesn't go into the conversation (Your account)
“who's on my team” / “make Olena an admin” / “remove Petro” / “revoke the invitation”	show the team and invitations / change a role / remove a person / revoke an invitation — see Team
“I want my own dashboard too” / “switch to Ivan's team” / “leave this team”	create your own workspace / switch the dashboard to another workspace / remove you from the team; the conversation stays in the workspace where it started
“delete this agent” / “bring back the deleted one” / “make a new secret key”	delete the agent (and tell you until when it can be brought back) / bring it back / reissue the key and show the new one once — all after your “yes”
“a customer asks to delete their data”	erase the conversation, the memory about this person and the contact — permanently; asks for consent
“what's going on in conversations?” / “reply that we'll deliver tomorrow”	list who's waiting for a person, drafts and actions awaiting approval / write a reply in your words, show the exact text and send it after “yes” — in your name (Conversations)
“send as is” / “correct it like this: …” / “approve” / “reject, because …”	decide on a training-mode draft / approve or reject an agent's action — only the one you named, and after “yes”
“here's the price list, learn it” / “send this file to customers” / “make actions from this file” (file attached with the paper clip)	tell the dashboard where to put the file: in knowledge, in files for customers, or in actions; how it ended — a line above the input field

You don't need to talk technical

The builder doesn't ask for what it can find out itself: server addresses, IDs, formats. If it does ask for something like that, that's its mistake — say “figure it out yourself”.

When something technical can't be avoided — a bot token, a key to someone else's service — it gives numbered steps for where exactly to get it and tells you what it looks like. It applies the key you send right away and won't repeat it in the conversation.

The agent sees photos, receipts and screenshots with no setup at all — in the chat window on your website, in Telegram, Viber, WhatsApp, Slack and in email. To have it draw an image or answer with voice, tell the builder — see Images and voice.

Next

• The panel beside the chat — what each row means.
• If it didn't come out right — undo and restore points.