System Prompt
Also: System Message, System Instruction, Instruction système, Message système, Prompt système, Systemanweisung, Systemnachricht
The hidden set of instructions that defines how an AI assistant behaves before any user types a question: its role, tone, limits and rules.
What It Is
A system prompt is the standing set of instructions given to an AI assistant before any conversation starts. It sets the assistant's role, tone, boundaries and priorities, and the user never sees it. Where a user prompt is a single question ("Summarize this contract"), the system prompt is the frame around every answer: "You are a compliance assistant for a European bank. Cite the source clause. Never give legal advice. If unsure, say so." The user asks; the system prompt governs how the assistant is allowed to answer.
Why it matters
For a senior leader, the system prompt is where brand voice, risk controls and policy actually live inside an AI product. A weak system prompt is how a customer-facing bot goes off-brand, invents a refund policy or answers a question it should have refused. When your team deploys an AGI chatbot for customer service, the difference between "helpful and on-message" and "legal exposure" is often a few paragraphs of system prompt. It is cheaper to change than retraining a model and faster to test, which makes it a genuine governance lever rather than a technical detail. Treat it as a controlled document: who can edit it, who approves changes, and how you check it still behaves after an update.
How it works
The system prompt is inserted at the top of the model's context every time, ahead of the user's message, so the model reads its instructions first. In practice it competes for attention with the user's input, which is why a determined user can sometimes talk around it ("prompt injection"). A CMO reviewing a marketing assistant should ask to read the actual system prompt, not a summary of it: check that it names the audience, forbids unsupported claims, and states what the assistant must decline. Good practice is to version it, test it against adversarial questions before launch, and log outputs so you can trace a bad answer back to a specific instruction and fix it.
Frequently asked questions
What is a system prompt?
A system prompt is the standing set of instructions given to an AI assistant before any conversation begins: its role, tone, boundaries and priorities. The user never sees it, but it frames every answer the assistant produces. Where a user prompt is a single request like "Summarize this contract", the system prompt decides how the assistant is allowed to respond to it.
What is the difference between a system prompt and a user prompt?
The user prompt is the question typed in the conversation; the system prompt is the permanent frame placed above it. A user prompt changes at every turn, a system prompt stays identical across all conversations and all users. One asks, the other governs what an acceptable answer looks like.
Why should a non-technical executive care about the system prompt?
Because brand voice, risk controls and policy actually live there inside an AI product. A weak system prompt is how a customer-facing bot goes off-brand, invents a refund policy, or answers a question it should have refused. It is cheaper to change than retraining a model and faster to test, which makes it a governance lever rather than a technical detail.
Can a user bypass the system prompt?
Sometimes, yes. The system prompt is inserted at the top of the model's context before the user's message, but it competes for attention with that input, so a determined user can occasionally talk around it. This is known as prompt injection, and it is the reason adversarial testing before launch matters.
What should I check when reviewing a system prompt before launch?
Read the actual text, not a summary of it. Check that it names the audience, forbids unsupported claims, and states explicitly what the assistant must decline. Then treat it as a controlled document: version it, define who can edit and who approves changes, test it against adversarial questions, and log outputs so a bad answer can be traced back to a specific instruction.