AI, explained
What is a system prompt?
It is the standing brief you give an assistant before anyone talks to it: who it is, what it must and must not do, and how it should sound.
Every capable language model is a generalist until you tell it what job it is doing. The system prompt is that briefing. It is separate from the user's messages and carries more authority: it is where you set the persona ("you are a support assistant for a Lisbon property firm"), the rules ("never quote a price, hand pricing questions to a human"), the format ("reply in the customer's language, under 120 words"), and the guardrails ("if you are unsure, say so and offer to escalate").
Why it is the highest-leverage part of a build
Most of the difference between an AI feature that embarrasses a company and one that quietly works is in the system prompt, not the model choice. A vague prompt produces a vague, over-eager assistant that invents policies and answers questions it should refuse. A precise prompt produces something that stays in its lane. Getting this right is cheaper and faster than fine-tuning and should always be tried first.
What a good system prompt contains
- Role and scope. One clear sentence on what the assistant is for, and an explicit list of what it should not attempt.
- Behaviour rules. How to handle uncertainty, when to escalate to a human, what it must never do (make commitments, quote prices, give legal or medical advice).
- Voice and format. Tone, length, language, and structure, so output is consistent and on-brand.
- Grounding instructions. In a RAG system, an order to answer only from the retrieved documents and to say when the answer is not in them.
The security angle people forget
A system prompt is not a vault. Users can attempt prompt injection, feeding instructions that try to override your rules, so the prompt should assume it will be probed. Keep genuinely sensitive logic and permissions in code and tools, not in prose the model might be talked out of following. Treat the system prompt as guidance the model usually follows, backed by hard limits the model cannot exceed.
In practice
In our production deployments the system prompt is versioned and tested like code, because a one-line change to it can shift the behaviour of thousands of conversations. If an existing assistant is going off-brand, refusing valid requests, or answering things it should not, the fix is almost always in the prompt and its guardrails. An AI audit can pinpoint where the instructions are letting you down.
Frequently asked questions
What is a system prompt?
A system prompt is the standing instruction given to an AI model before any user message. It sets the assistant's role, rules, tone, and output format, and it applies to the whole conversation. Users typically do not see it, but it governs how the model responds, which makes it the most important control in most AI builds.
What is the difference between a system prompt and a user prompt?
A user prompt is the individual message a person types. A system prompt is the persistent brief that sits above all user messages, defining the assistant's identity and boundaries for the entire session. The system prompt carries more authority and shapes how every user message is interpreted and answered.
Can a system prompt keep an AI secure?
Only partially. A good system prompt reduces bad behaviour, but users can attempt prompt injection to override it, so it should not be the only line of defence. Keep sensitive permissions and destructive actions enforced in code and tools rather than relying on the model to obey instructions it might be manipulated into ignoring.
Related
Ready to put AI to work?
Book a discovery audit and we will map the highest-ROI AI agents and automations for your business.
Book a discovery audit →