Usage & Prompting
System Prompt
The standing instruction that applies before every conversation: role, tone, rules, taboos. Users don't see it, but it frames every answer, which makes it the most important steering instrument in operating an assistant. It is also a security issue: whoever can manipulate the system prompt changes the system's behavior from the inside.
An example
The system prompt sits in front of every conversation and sets role and limits: “You are the customer support assistant. Answer only about products from the attached list. Never state prices without pointing to the price list. For questions about contracts, refer to a human.” The user doesn't see this text. It takes effect with every single request, which makes it the cheapest place to regulate behavior for an entire system.
A closer look
The system prompt comes before every user input and, in most models, carries more weight. This is where role, tone, limits, and format requirements belong: everything that should apply to every request.
Why it matters
Two things are regularly overestimated. First, reliability: a system prompt is an urgent request, not a technical barrier. With unusual inputs or in long conversations it loses effect, and anyone who sets out to do so can often override it. Anything that truly must not happen additionally belongs in code, as a check before sending, not as a plea in the text. Second, secrecy: system prompts can often be extracted. Don't put anything in it whose disclosure would do harm.
Commonly misunderstood
It is not a security boundary. What is in it can be coaxed out with a little effort, and what it forbids can be rephrased around. Real boundaries belong in the application.