What a chatbot's system prompt is

When you talk to a company's chatbot, there is part of the text the model reads that you cannot see: the instructions written by whoever set it up. That is where it says, for example, that the bot only answers questions about products and shipping, that it uses the informal "you", that it does not promise discounts and that it passes complaints to a person.

The system prompt is the simplest tool for giving an assistant a character and boundaries. But it is not a safe: well-written instructions reduce errors, they do not eliminate them, and a determined user can try to get around them.

An example in an SMB

An estate agency puts a chatbot on its website. In the system prompt it writes down the areas it works in, that no price valuations are given in the chat, that every viewing request must be collected with name, phone number and property, and that the tone is courteous but direct. Without these lines the same model would answer on any subject and would hazard estimates the agency does not want to give.

The most common mistakes

• Writing it once and never reading it again, even when price lists, opening hours and services change.
• Relying on the system prompt to protect confidential data: what the model must not say, it should not have in the first place.
• Filling it with contradictory rules: the model chooses which one to follow, and it does not always choose well.

All the words of AI, explained simply, are in the glossary.