---
title: "Give each source of your chatbot its own skill"
description: "You tell your chatbot which sources it can use, and yet it wanders off. With a short instruction and one skill per source, you draw that line much tighter."
language: en
source: https://coachsteff.live/en/een-skill-per-bron
---

# Give each source of your chatbot its own skill

A group of participants built a chatbot for career questions in a workshop. They gave it a mix of internal documents and sources on the internet. The prompt stated very clearly which sources it could trust, and those were only Belgian ones. When they checked the source references, a foreign site about social security was included anyway. Perhaps you have experienced this yourself. You write a clear rule in your instruction, and your AI does something else with it anyway.

Then the reflex is to make the prompt longer and stricter. I understand that, but I would choose a different path. You can actually define your sources much more firmly if you take them out of your main instruction.

Copy the prompt at the bottom of this article. With that, you let your AI write a separate skill for each source.

## An instruction works differently from software

We are used to software doing what you say. If this happens, then it does that. An instruction for an AI is always non-deterministic. It is not an if-then-else. You actually never quite know what is going to happen. You can give it rules, and then you have to see how strictly it sticks to them.

The group had also tested that in a funny way. They asked the chatbot to end every answer with three emojis. It did that neatly. Only it sprinkled a few more throughout the text, just for fun. So a rule that you find crystal clear still leaves some room for an AI.

And the more rules you stack in one prompt, the greater the chance that they become vague or contradict each other. An AI struggles with that.

## Frequently asked questions

**Why does my chatbot use sources that I excluded in the prompt?**

- An instruction for an AI is non-deterministic. You can give rules, but you never quite know how strictly it sticks to them. The more rules in one prompt, the greater the chance that they become vague or clash.

**What do you mean by one skill per source?**

- You give your agent a short main instruction and a separate skill for each source it is allowed to use. Each skill states what it can and cannot extract from that source. That way the boundary around your sources becomes much clearer.

**Is my chatbot completely reliable after that?**

- You are never completely sure, because an AI works differently from regular software. Test with questions to which you know the answer and look in the source references. If it wanders off, you sharpen the skill of that source.

**Where can I learn this hands-on?**

- In AI agents and in Working smarter with Copilot in M365 you build agents and skills yourself, with guidance. You can find the dates at [Workshops & Events](https://coachsteff.live/agenda).
