System Prompt
A hidden instruction layer sent to an LLM before the user's input that configures the model's behavior, persona, constraints, and context. The system prompt is where AI agents receive their role definition, business rules, and operational guidelines.
What is a System Prompt?
When you interact with an AI assistant, there are typically two layers of input: the system prompt (which you do not see) and the user message (which you write). The system prompt is set by the developer or operator and instructs the model on how to behave: what role it plays, what it is and is not allowed to do, what context it should always have in mind, and how it should format responses.
A simple system prompt might be "You are a helpful customer service agent for Acme Corp. Always respond in Dutch. Do not discuss competitor products." A complex operational system prompt might include detailed business rules, the structure of the ERP data model, escalation logic, and examples of how to handle specific exception types.
Why System Prompts Matter
The system prompt is the primary mechanism for turning a general-purpose LLM into a specialized operational agent. The same underlying model — say, GPT-4o — behaves entirely differently depending on the system prompt. With one system prompt it is an invoice processing agent that knows your chart of accounts and approval thresholds. With another it is a customer service bot that knows your return policy and escalation contacts.
Role definition: "You are an accounts payable processing agent. Your job is to extract invoice data and match it against purchase orders in the system."
Business rules: "If the invoice total exceeds EUR 50,000, always route to the CFO for approval regardless of PO match status."
Output format: "Always return results as structured JSON with the fields: extracted_data, confidence_scores, match_status, recommended_action."
Constraints: "Never post directly to the ERP without a confidence score above 0.90 on all mandatory fields."
System Prompts in Operations
For operational AI deployments, the system prompt is where institutional knowledge lives. It encodes what a trained accounts payable clerk knows implicitly: which suppliers always invoice in a different currency than the PO, which cost centers own which purchase categories, what "delivery complete" means in the context of partial shipments. Getting the system prompt right is a significant part of deploying a reliable operational agent — it is not a configuration file, it is the operational rulebook that determines whether the agent behaves like a competent team member or a confused intern.