System prompt
What is System prompt?
The standing instructions given to a model before any user message: its role, its rules, its tone and what it must never do. The user asks the question. The system prompt sets the behaviour around every answer. Get it wrong and the assistant is inconsistent or off-brand. It is usually the first thing to fix when an AI assistant misbehaves.
Why it matters
A system prompt is the cheapest, fastest lever on how an assistant behaves, and that ease hides two traps. The first is treating it as a security control. It steers behaviour, but a determined user can often talk the model past its instructions or coax them into the open, so anything that truly must not happen needs enforcing in code, not just requesting in the prompt. The second is fragility. Small changes in wording can shift behaviour more than you would expect, so a prompt running in production deserves the same version control and testing as code. It is powerful precisely because it is so easy to change, and that same ease means an untested edit can break the assistant for everyone at once.
In practice
A company ships a customer assistant whose system prompt sets its role, its limits and a firm rule never to discuss competitors. A support engineer later tweaks one line to fix a tone issue and, without testing, loosens that rule by accident. Because the prompt was under version control, the change is caught in review and rolled back before release. The instructions are treated as production code, not a text box someone edits on a whim.
