System Prompt (Agent Instructions)

Models receive messages in roles. The system message comes from the business and carries standing orders; user messages come from the customer; assistant messages are the model's own replies. Because the system prompt sits at the top of every request, it is the single most influential piece of text in the deployment: 'You are the booking assistant for Clinic X in Riyadh. Reply in the customer's dialect. Never quote a price that is not in the price list. If the customer mentions pain or an emergency, give the emergency number and hand off immediately.'

A good system prompt is a business document, not a programming trick. It encodes the policies the owner would give a new employee on day one: what we sell, what we do not promise, how we address customers in Saudi versus Egypt, what to do with a complaint, which questions go to a human. It also shapes register: a prompt that says 'formal Arabic' will produce answers that feel stiff to a Jeddah customer writing in Hijazi, so dialect and tone are specified deliberately and tested with golden-set evals.

Two practical points. First, the system prompt is not a security boundary: a determined user can sometimes talk a model around its instructions, so hard limits (what the agent may actually do, which tools it can call, what data it can see) belong in code and guardrails, not only in prose. Second, ask your vendor who owns and can read the prompt. Nano AI treats it as the client's asset, reviews it with the client at kickoff, versions every change, and re-runs evals on each revision as part of the monthly tuning included in the Launch and Growth plans.