What is system prompt?

A system prompt is the hidden text an application places in front of every conversation, telling the model what it is, what it may answer from, how to behave and what it must never do. The visitor never sees it. In a support product the vendor writes most of it, the base rules that keep the agent grounded and polite, and leaves a slot for the customer's own instructions.

It matters because most of a product's behavior lives there. Two products on the same model with different system prompts answer the same question differently: one declines off-topic questions, one chats; one says "I don't know", one guesses. When a vendor demonstrates tone control or refusals, it is usually a line in the system prompt. The 12 instruction examples show what a well-written slot of it looks like.

Where it runs out: a system prompt is advice to the model, not a lock. Visitors can sometimes extract it by asking, so it should hold no secrets, no API keys and nothing you would mind a customer reading. It cannot make the model know facts; facts belong in retrieval. And an instruction in the system prompt can be contradicted by text the model reads later, a page or a tool result, which is why prompt injection is a separate concern and why well-built products wrap retrieved content as data rather than as instructions.