AI Assistant — the brain behind the replies

What it is

The AI agent is the engine that generates every automatic reply in your Instagram and WhatsApp DMs. You write a prompt (the agent’s instructions) and the AI follows it to the letter throughout every conversation.

Every account starts with Lume, the default AI trained by ZeroChats. You can test it, modify it, clone it or create new agents from scratch.

How to configure it

Create or select an agent

From the agent tab you can:

Create new AI agent — opens the dialog where you define the agent’s name and language.
Clone — copies the entire configuration of the current AI into a new agent.
Delete AI — removes the agent. You can’t delete the last remaining agent or one that’s in use by a business.

Agent fields

Field	What it’s for
AI name	Identifies the agent. Maximum 40 characters.
Language	Spanish, English or Portuguese. Sets the language the AI will use with the lead.
Prompt	The agent’s instructions: personality, tone, goals, topics to cover and to avoid.

Edit the prompt

You have two ways to modify the prompt:

Edit with AI — write what you want in natural language (“Make my AI agent more effective”, “Add personality to my AI agent”) and the AI rewrites the prompt for you. There are quick suggestions like Use the right tone for my business or Use emojis in replies.
Edit manually — change the prompt text directly in the editor.

Every time you save, the version is recorded in the Prompt history, where you can view, simulate, restore or delete previous versions.

Quality bar

Next to the prompt you’ll see a quality bar that scores your instructions from 0 to 10 and tells you what to improve: goal clarity, tone definition, objection handling, examples, critical rules, etc. It’s a quick guide to know whether the prompt is ready before simulating or activating it.

Simulator

The Simulator button opens a test chat where you talk to the AI just like a lead would. Use it to validate the prompt before activating it in production. You can return to a previous point with Return to this point.

AI activation

The AI activation switch turns automatic replies for incoming messages on or off. It’s the agent’s master switch.

Delegations

Delegations let one agent pass the conversation to another specialized agent when a specific situation occurs.

From Manage Delegations → Add Delegation:

Choose the target AI in Delegate to AI.
Describe the Situation in which the handoff should occur (e.g. When the customer asks about technical support).
Save. The rule appears in Current Delegations.

The lead is transferred silently: they never see a switch message. From that moment on they converse with the target agent, with its own prompt, content and sales links.

How a conversation works

The lead writes first. ZeroChats rebuilds the context, then the AI replies with short messages and a small delay.

The lead sends a DM on Instagram or WhatsApp (or replies to a Story from Instagram).
ZeroChats rebuilds the conversation context with that lead.
The AI generates short, natural messages that match the agent’s prompt and language.
It sends them with a small delay to mimic a person typing.

Best practices

Be specific about the agent’s personality and tone.
Explicitly list the topics the AI should and should not address.
Define a clear goal (book a call, send a link, qualify a lead).
Anticipate the most common objections and explain how to respond to them.
Keep the prompt focused: the AI follows your instructions literally.
For critical rules, write them in UPPERCASE and repeat them more than once in the prompt; it helps the AI prioritize them.
If a real conversation didn’t go the way you wanted, copy the chat from the leads panel and paste it into Edit with AI indicating what you would have wanted it to reply. The AI rewrites the prompt to fix that case.
Before activating big changes, test them in the Simulator.

How to use examples in the prompt

Before we get into it, one idea you have to be crystal clear about: everything you write in the prompt, your agent reads as instructions it has to follow. It doesn’t automatically tell apart “this is a rule” from “this is just an example so you understand the idea.” If you put a specific phrase in quotes, the most likely outcome is that it will drop it verbatim to your real customers. That’s why you have to be careful with examples.

Think of it this way: your agent is like a brand-new, extremely literal employee. If you teach it to greet customers by telling it “say hello with: ‘Hi, how can I help you?’”, it will drop that exact phrase to everyone, for months, with no variation. If instead you tell it “greet the customer in a friendly, brief way, adapting to how they write”, it will improvise different greetings every time.

When to add an example and when not to

Don’t add examples when what you want to regulate is:

Greetings, farewells or polite formulas. If you put a sample greeting, that greeting will repeat identically conversation after conversation. Your customers will notice it as artificial.
The tone or personality of your agent. How your agent “sounds” isn’t taught with sample phrases: it’s described with words.
How to present your product or offer. If you put a specific sales phrase as an example, that phrase will appear over and over, even when the real customer is asking about something completely different.
Responses to typical objections. Every customer raises an objection in their own way. If you put a closed sample reply, your agent will drop it even if it doesn’t quite fit what the customer said.
Follow-up templates or any fixed script. If you write it as an example, you’ll see it copied word for word.

It can make sense to include an example (carefully) when:

Your business uses a very specific word or term that your agent wouldn’t use on its own and you want to lock it in (for example, a particular way of naming your services or internal stages).
You want to set a specific response structure that’s hard to explain in words. Even then, describing it almost always works better than exemplifying it.
There’s a very specific case that rules don’t quite cover and an example disambiguates. These cases are fewer than they seem.

Signs an example will be copied literally

It’s in quotes.
It’s a complete, polished sentence, ready to drop.
You introduce it with something like “Example:” or “The AI should say:”.
You put it at the end of the prompt, right before closing it.
It contains the real name of your product, a specific price or very specific data.
It’s the only example you provide, with no variants showing the idea can be expressed in many ways.

How to write an example properly, when you do need one

Use several, not one. A single example automatically becomes the agent’s reply. Three or four variants, different from one another in wording and length, teach it that the idea can be expressed in many valid ways. If you can only think of one version of your example, better skip it: turn it into a rule.

Don’t put them in quotes. Quotes suggest to your agent “this is what I have to say verbatim.” Better embed the examples within the explanation, in a bulleted list, or as descriptions without quotes. The less they look like an “official phrase,” the lower the risk of literal copying.

Templates with blanks beat full sentences. Instead of writing a complete reply word for word, write the shape of the reply. Something like “brief greeting + reference to what the customer said + one specific question” works much better than a closed sentence ready to drop. You’re teaching the idea without imposing the words.

Don’t place examples at the end of the prompt. The last thing your agent reads carries the most weight when responding. If you close the prompt with a sample phrase, that phrase becomes the leading candidate to appear copied. Put examples in the middle and always close with rules, never with samples.

Flag them explicitly as examples. After the example block, add a clear line like:

The samples above are guidance. Never copy them word for word. Always adapt the vocabulary, length and form to each real conversation.

This isn’t decoration: it helps a lot. Your agent lives alongside other instructions that do tell it to copy things exactly (like the URLs of your links). It pays off to make clear that your examples don’t work that way.

Don’t put real business data inside an example. If you write the name of your product, a specific price or a concrete deadline inside the example, that phrase with that data will get locked in and reappear even when it doesn’t fit. Data belongs in the business info field (a different field from the prompt). In prompt examples, use generic terms or blanks like [product name].

Common mistakes worth avoiding

Adding a single “perfect” example thinking it’s enough. It ends up as a fixed mold.
Writing formats like “Customer: … → AI: …”. Your agent learns to drop exactly what you wrote after “AI:”.
Quoting verbatim what you don’t want it to say (“never say: ‘dear customer’”). Sometimes just seeing it written activates it anyway. Better describe the problem without writing the forbidden phrase (“avoid outdated and overly formal formulas”).
Mixing examples and rules in the same paragraph with no visual separation.
Filling the prompt with examples thinking more is better. More examples = more things that can come out copied verbatim.

What to write instead of an example

Almost everything people try to teach with examples is better taught with descriptions and rules. It’s counterintuitive, but it works much better.

To regulate tone and how it speaks, use adjectives and specific restrictions, not sample phrases:

Warm but professional. First-person basis. No archaic formalities. No emojis except in greetings.
Very short messages, max two sentences for simple questions.
Avoid empty politeness formulas and chained confirmation questions.

This covers every possible conversation without any of them coming out cloned.

To regulate behavior, use conditional rules, not scripts. Instead of putting a specific reply phrase, describe the behavior step by step:

When the customer asks about prices:

If it’s unclear which plan they mean, ask them to clarify before giving the price.

Give the price of the relevant plan and at most two differentiating benefits.

Don’t list every benefit or push for the close.

End with an open question that invites further conversation.

This works for infinite ways a customer can ask about price, and your agent replies with its own words every time.

To teach how to handle situations, talk about the customer’s state and what your agent should achieve, not what it should say:

When the customer shows doubts: acknowledge the specific doubt they mentioned, offer specific information about that particular doubt (not generic answers), and close by inviting further conversation without pushing for the close.

You’re describing what to do, not what to say. Your agent picks the words, you control the behavior.

To stop your agent from talking about things it shouldn’t, use clear prohibitions, without including the bad phrase as an example:

Never offer discounts that aren’t described in the business info.
Don’t mention plan prices the customer hasn’t asked about.
Don’t promise delivery times or specific results.

Clear rules, in the negative, without writing the phrase you don’t want to appear.