System prompt
A system prompt is the set of instructions given to an AI model before a conversation starts, usually written by the developer or business deploying it, that defines its role, tone, rules and the information it should use.
Users normally never see the system prompt. When you chat with a support bot on a shop's website, the shop has already told the model something like "You are the support assistant for this store; answer only from the attached policies; be brief and friendly". Your questions arrive after that.
A system prompt is different from your own prompt, which is the message you type, and from fine-tuning, which changes the model itself. It shapes behaviour for one deployment without retraining anything, which is why it is the first tool businesses use to customise an AI model.
How it works
AI services keep the system instructions separate from the user's messages and the model's replies. The system prompt comes first and is sent along with every request, so the model reads it each time before it reads the latest question. Models are trained to give it more weight than user messages, though not absolute authority.
A good system prompt covers who the assistant is, who it is talking to, what it may and may not do, which sources to rely on, the format of replies and what to do when it does not know. Many deployments also insert retrieved documents or the customer's account details alongside it.
System prompts are not a security boundary. Users can sometimes trick a model into revealing or ignoring its instructions through prompt injection. Rules that must hold, such as never issuing refunds, belong in the surrounding software and guardrails, not only in the prompt. Instructions also compete for attention: the more rules a prompt contains, the more likely one is to be ignored in an unusual conversation, so the most important rules should be stated plainly and early.
Example
A physiotherapy clinic sets up a booking chatbot. Its system prompt says the bot represents the clinic, answers only from the attached list of treatments, prices and hours, never gives medical advice, suggests calling the clinic for urgent pain, replies in the language the patient writes in, and keeps answers under eighty words.
In testing, a patient asks whether their knee injury needs surgery. The first version of the bot offers an opinion. The clinic adds an explicit instruction to decline diagnostic questions and point to a consultation, retests with ten similar questions and keeps a copy of each version of the prompt so changes can be traced. The clinic also notices that the bot occasionally answers in a stiff, formal tone. One sentence added to the system prompt, with a sample reply in the clinic's friendly style, fixes it. Showing an example tends to work better than describing the tone in adjectives.
Why it matters for your business
If you deploy any AI tool for customers, the system prompt is where most of its behaviour is decided, and it is cheap to change compared with retraining. A vague prompt produces a bot that wanders off topic or makes promises; a precise one keeps it useful and safe. Ask any supplier to show you the system prompt they wrote for you and how they tested it. Keep the prompt under version control with dates, so that if behaviour changes you can see whether the prompt or the underlying model changed.
Common mistakes
- Writing a one-line system prompt and expecting the model to know your policies and limits.
- Putting secrets such as passwords or private API keys in the system prompt.
- Relying on the prompt alone to stop risky actions instead of enforcing limits in the software.
- Changing the prompt without retesting it on the same set of real questions.
Questions people ask
What is the difference between a system prompt and a user prompt?
The system prompt is written by whoever deploys the AI and sets standing instructions for every conversation. The user prompt is the message a person types in that conversation. The model reads both, and it is trained to give the system prompt more weight.
Can users see the system prompt?
Not normally, but they may be able to coax a model into revealing it. Treat a system prompt as private but not secret: never put passwords, keys or confidential data in it, and assume a determined user could read it.
How long should a system prompt be?
As long as it needs to be to cover role, audience, rules, sources, format and what to do when unsure. For a simple support bot that might be a few hundred words. Longer is fine if every line earns its place, since each instruction adds tokens to every request and so a little cost.
