The standing instructions a model gets before any user says anything — its job description, rules, and personality. Users of a product never see it, but it shapes every reply.
Why it matters: when an AI product behaves oddly, the fix is often in the system prompt, not the model. It's also where a product's safety rules live.