Keep your system prompt focused on five essentials — identity, audience, job, constraints, and tone — then move all detailed background information into a knowledge base that the agent retrieves on demand. This architecture gives your agent everything it needs without bloating the prompt.
The Temptation to Paste Everything In
When educators first build AI agents, the instinct is to paste everything into the system prompt. Your full course outline, all your FAQs, your student policies, your program history, your teaching philosophy — the thinking is “the more context the better.” But this approach creates a bloated prompt that is hard to maintain, harder to update, and often makes the agent less focused rather than more capable.
The principle to internalize is this: the system prompt is for permanent instructions, not reference information. Instructions tell the agent how to behave. Reference information is what the agent looks up when it needs a specific answer. These belong in different places.
What Belongs in the System Prompt
Your system prompt should answer five questions in plain language: Who is this agent? Who does it serve? What is its primary job? What must it never do? What tone should it use? If you can answer all five in 300-500 words, you have a well-scoped system prompt. Anything beyond that is probably reference material that belongs in a knowledge base, not in the prompt itself.
Test your prompt length by asking: does this sentence tell the agent how to behave, or does it give the agent information to look up? Behavioral instructions go in the prompt. Lookup information goes in the knowledge base. A sentence like “Always respond in a warm, encouraging tone” is behavioral — it belongs in the prompt. A sentence like “The enrollment deadline for the Spring cohort is March 15th” is lookup information — it belongs in the knowledge base, where the agent can retrieve it when a student asks and where it can be updated without touching the system prompt.
What This Means for Educators
For coaches and consultants building campus agents, separating instructions from reference material makes your agent dramatically easier to maintain. When your program details change — new cohort dates, updated pricing, new module names — you update the knowledge base, not the system prompt. The agent’s core behavior stays stable while its factual knowledge stays current. This is the architecture that scales as your program grows.
The Simple Rule
System prompt = who the agent is and how it behaves. Knowledge base = what the agent knows and can look up. Write a tight system prompt. Build a well-organized knowledge base. Connect the two. That combination gives your agent the right context without turning your prompt into a document only you can parse.