A normal prompt is what you ask right now. A system prompt sits a level above the task, telling the AI how to interpret everything that follows: what role it’s playing, what rules matter, and what behavior should frame the conversation.
Think about hiring someone with a job description. Before you hand them today’s task, you tell them the job: you’re the office manager, you protect client confidentiality, you write in this tone, you don’t send anything without approval. Those instructions sit above any single task.
How it shows up
You usually don’t see the system prompt directly. The product has one, your organization may add one, and a project may add project instructions. In Claude Code, Codex, and other agent tools, you feel these standing instructions through behavior: the agent asks before risky actions, prefers certain tools, or follows a repo’s local rules.
Why you care
AI doesn’t just answer your last sentence, it follows all the instructions it has, with some carrying more authority than others. If the system prompt says never reveal secrets and a user prompt says print the API key, the higher-level instruction should win. That’s why system prompts tie into guardrails. Still, a system prompt is language, not a wall, so good systems pair it with permissions, review, scoped tools, and human judgment. It sets the working posture before the actual work begins.