What is my agent's prompt?

It's the set of instructions that program how it thinks and responds: its personality and limits, the step-by-step sales guide, the rules it always follows, and what it must never do under any circumstance. You edit it in Settings, Training, Prompt.

Your assistant's knowledge base: the exclusive place it pulls information from to respond. Your frequently asked questions, services, prices, hours, and shipping policies live there.

How do I get the agent to pass a case to a human?

In the prompt's Advanced Settings you define the Resolution Policy: what limits it has and when it hands off. For example: if the customer asks for a refund, have it validate the case without promising the refund and hand the chat off to support.

How do I control the tone of the responses?

With the Response Style: tone, length, and formality. For example, a professional but friendly tone, paragraphs of no more than three lines, and prioritizing empathy if the customer shows annoyance.

Help center

Training and memories for your agent

How to train your agent in TBit: instructions, memories, and the difference between what it always knows and what it learns from each customer.

Published on July 31, 2026Updated on October 10, 20265 min read

Training is the key instruction that guides your assistant through your sales or service process. Here you define how it thinks, how it responds, and where it pulls information from.

Watch the video: TBit Learning, How to train your TBit

The Prompt

Go to Settings → Training → Prompt → Edit. You'll see a description of your business that the AI loaded automatically when your TBit was created. The prompt isn't a chat message: it's the set of instructions that program the AI's "brain." To build it, define:

Personality and limits: the assistant's tone of voice and the scope of its functions.
Step-by-step sales guide: the exact flow it should follow in the interaction.
Golden rules: the guidelines it must always respect.
Restrictions: what it must not do or say under any circumstance.

The Prompt editor: your assistant's operating brain.

Advanced settings

Inside the prompt, in Advanced Settings → Edit, you define two pillars of its behavior:

Resolution Policy: how it reacts, what limits it has, and when it hands off to a human. E.g.: "If the customer asks for a refund, tell them you'll review the case, don't promise the refund, and transfer the chat to Support."
Response Style: tone, length, and formality. E.g.: "Professional but friendly tone, no jargon, paragraphs of no more than three lines. If the customer writes in all caps or seems upset, prioritize empathy."

The Prompt editor with draft/publish and Advanced Settings.

Memories

Memories are your assistant's brain and knowledge base: the one place it pulls information from to answer. It includes FAQs, services, prices, hours, or shipping policies.

Golden rule

For the assistant to check this base, explicitly include the instruction "Use your memories" inside the main prompt. Gather your FAQs, product PDFs, hours, and shipping or return policies.

Training memories: the agent's knowledge base.

Analyze

The Analyze section reviews all the information you've uploaded and gives you improvement tips so your TBit answers better. Take them as a guide: the analysis suggests, but you know your business. If something you know matters (hours, a policy) doesn't show up in the prompt, add it even if the analysis doesn't ask for it.

Analyze: check how well trained your agent is and get recommendations.

Test without fear: draft and publish

You don't have to "get it right the first time." Every change you make to the prompt is saved automatically as a draft, without affecting your real customers. So you can iterate calmly:

You edit the prompt (or the memories) and the change stays in draft.
You message from your test line (your personal WhatsApp): the agent already replies with the new instructions, just for you.
When you're happy with it, click Publish to production and only then does the change apply to everyone.

Improving is a practice

Fine-tuning the prompt comes from practice: you write, test, spot a contradiction or an answer that "slips," adjust, and test again. The draft is exactly for that: experimenting without risk to your real operation.

Have it introduce itself as a virtual assistant

Set up the prompt so the agent identifies itself as a virtual assistant from the start, to set the right expectation with the customer. Save the word "advisor" for when you actually transfer the conversation to a person on your team.