How to train your TBit AI agent on WhatsApp, step by step

Training is not programming

Training your TBit agent is not writing code or filling out technical forms. It's teaching it the context of your business like you would teach a new employee: who you are, what you sell, how you talk and what rules you have. The engine already knows how to chat, format WhatsApp messages and hand off to a person when needed. Your job is to give it the information only you have.

Step 1: the prompt defines who it is

The prompt is the agent's identity. That's where its personality, tone, how it sells and your brand's own policies go. A short, specific prompt performs better than a long one full of technical instructions.

What does go in the prompt:

• Personality and tone: friendly, formal, direct.

• How it sells: whether it suggests add-ons, whether it offers the combo first, whether it insists or not.

• Your own policies: asking for a deposit, confirming before scheduling, your business conditions.

What does NOT go in the prompt: schedules, prices or product lists. That data changes and is set up elsewhere. Schedules go in Scheduling and products in the Catalog.

Step 2: memories are what it knows

If the prompt defines who it is, the memories define what it knows. They are the business knowledge: policies, FAQs, conditions, promotions. The agent checks them to answer with real information.

For a memory to work:

• One idea per memory: "returns" and "shipping" are two memories, not one.

• In the customer's words, not internal jargon.

• With exact figures: deadlines, amounts and conditions.

How to write memories the agent will actually find is in the memories guide.

Step 3: configure, don't write

A good part of what an agent needs isn't written: you configure it once and the agent uses it on its own.

• Hours and services in Scheduling.

• Products, prices and photos in the Catalog, Store menu.

• Data you want to capture from the customer in Forms.

• Sales stages in the Pipeline.

Configuring instead of writing keeps the agent from answering with old data. A schedule in the prompt goes out of date; a configured schedule is always up to date.

Step 4: test by chatting

The best way to train is to talk to it like you were a customer. Write to it like someone undecided, like someone upset, like someone comparing prices. Notice where it hesitates, where it says too much or where it's missing information. Every failure tells you which memory to adjust or which line of the prompt to fix.

Step 5: let the agent tell you what it's missing

You don't have to guess. The training backlog gathers the questions the agent couldn't answer in real conversations. Review it: every unanswered question becomes a new memory or a tweak to the prompt. That way training stops being an event and becomes a short, steady habit.

What you DON'T need to do

There are instructions you can skip because the engine already handles them:

• You don't have to tell it "don't make things up": it answers with what it knows and, if it doesn't know, it says so and escalates.

• You don't have to tell it "escalate to a human": it already does that when the customer asks or it can't solve something.

• You don't have to tell it "answer briefly": it already uses the WhatsApp format.

All of this comes out of the box. What the engine already includes is in the engine guide. Focus your energy on what only your business knows.

Train with your own AI

Do you use ChatGPT or Claude? Connect them to your TBit agent and let them help you train it with your real data: on the platform go to Agent Settings, Connect your AI section, generate your token and add the connector https://api.tbit.app/mcp. Your assistant can review the prompt, the memories and recent conversations, and suggest improvements.