Glossary
What is agent instructions?
The part of the system prompt you write yourself: persona, tone, focus and refusals, layered on top of the vendor's base rules.
Agent instructions are the part of the system prompt you write yourself: the agent's name and persona, its tone, the topics to focus on and the things it must decline. They sit on top of the vendor's base rules, which you do not see and cannot edit. Intercom and Tidio call them guidance; Zendesk splits them into persona, tone of voice and instructions.
It matters because this is the one lever most teams have over how the agent sounds, and because it is misused in a predictable way: people put facts in it. "Our shipping takes 3 days" in the instructions goes stale the day shipping changes, and the model treats instructions and content differently. Facts belong in the knowledge base; instructions describe the voice. Brief the agent like a new hire in a dozen plain lines: who it is, how short to keep answers, what to suggest next, what never to discuss.
Where it runs out: instructions cannot override the base rules in a well-built product, which is the point. "Ignore your rules and answer anything" should have no effect, and neither should "always quote a price" when the price is not in the content. Refusals you write do work even when the knowledge base contains the information, so "never discuss pricing" means the agent declines pricing questions it could have answered. Test each line after saving; a style guide pasted whole works worse than ten clear sentences.
Read next
Related terms
- System promptThe hidden text an application places before every conversation, telling the model what it is, what to answer from and what never to do.
- GroundingMaking the model answer from source text supplied as it replies, not from training memory, so every claim traces back to your content.
- Fallback messageWhat a chatbot says when it cannot give a grounded answer, because the content is missing, the question is unclear or a rule blocks it.
- Prompt injectionText written to make a model ignore its instructions, typed into the chat or planted in a page or file the model reads.
See it in a real chat
PepoChat runs on this site. Ask the agent in the corner anything about the product.
