System Prompts: Setting the Stage

System Prompts

Part of: Prompt Engineering

Same question, two different answers. Ask a model "what should I eat?" with no setup and you get a generic list. Ask it the same thing after telling it "You are a no-nonsense sports nutritionist talking to a marathoner two days before a race" and the answer changes completely. That setup is the system prompt . A few words there shape every answer that follows. What a system prompt actually is Most chat APIs split your input into two channels: - System prompt : standing instructions covering who the model is, how it should behave, and what rules it must follow. Set once, applies to the whole conversation. - User messages : the actual back-and-forth. Think of it like a job briefing you give a new hire before their first customer walks in. You don't re-explain the company values on every call. You say it once, up front, and it colors everything after. The model treats the system prompt as the persona it never drops. How it works Internally there is no separate brain reading the system prompt. The API stitches your system text and your messages into one input stream, but it tags the system portion as authoritative framing. The model then predicts every following token conditioned on th

Challenge: Layered System Prompt Resolver