Write an explicit scope definition in your system prompt — a clear statement of what the agent handles and a specific redirect instruction for anything outside that scope — and the agent will stay on topic without needing you to police every conversation.
Why Agents Go Off Script
An AI agent goes off script when its system prompt has no clear boundaries around what it is for. Without scope definition, the agent defaults to being generally helpful — which means it will answer any question a student asks, regardless of whether it is within your intended use case. A student asks your course FAQ agent for relationship advice, and a scopeless agent tries to help. A student asks about a competitor’s platform, and an unbounded agent gives a thoughtful comparison. These are not malicious failures — they are the natural result of an agent that was never told where its job ends.
Writing Scope That Works
A scope definition has two parts: what the agent handles and what it does not. Both need to be explicit. “You are a student support agent for [Campus Name]. You answer questions about course content, campus navigation, session schedules, and community guidelines” — that is the positive scope. Then: “You do not provide personal coaching, medical or legal advice, opinions on topics outside the course curriculum, or information about competitors. If a student asks about something outside your scope, acknowledge their question warmly and redirect them: ‘That is a bit outside my area here — for that, I would reach out to [name] directly or check [resource].'”
The redirect instruction is as important as the scope definition itself. An agent that says “I cannot help with that” feels cold and unhelpful. An agent that says “That is outside my area, but here is exactly where to go” feels warm and competent — it stays on brand even when it is staying in bounds.
Testing Your Scope Instructions
After writing your scope section, test it deliberately. Ask the agent three questions it should handle well and two it should redirect. Check whether the redirect language feels on-brand or abrupt. Adjust until both the in-scope answers and the out-of-scope redirects sound like something you would say to a student yourself.
What This Means for Educators
As a coach or trainer, a well-scoped agent is a reliable one. Students learn quickly what the agent is for and stop asking it things outside its role — not because it refuses harshly, but because the redirect is always helpful and specific. That clarity builds trust, and trust is what keeps students engaged with your campus long-term.
The Simple Rule
Define what the agent handles, define what it does not, and write the exact redirect language it should use when it hits the boundary. Scope definition plus redirect instruction equals an agent that stays on topic without feeling like a locked-down system.