System Prompts: Setting the Rules for an AI

Imagine hiring a brilliant new assistant who can do almost anything you ask, but who arrives on day one with no idea what your business does, who your customers are, or how you like things done. You could explain it all afresh in every single conversation, or you could write one clear briefing note that sits on their desk and quietly shapes every reply they give. In the world of artificial intelligence, that briefing note is called a system prompt, and learning to write a good one is one of the highest-leverage skills you can pick up.

This guide explains what a system prompt actually is, how it differs from the messages you type day to day, why it has such an outsized effect on the quality and safety of an AI's output, and how to write one that consistently gets you the behaviour you want. No coding is required, and you do not need to understand the maths behind the models to use this well.

What a system prompt actually is

Most modern AI assistants are built on large language models, and when you interact with one there are usually two kinds of instruction at play. The first is your ordinary message, the question or request you type in the moment. The second, sitting quietly behind the scenes, is the system prompt: a standing set of instructions that tells the model who it is meant to be, what it should and should not do, and how it should respond. To understand why this matters, it helps to know a little about how large language models work, because the system prompt is simply the first and most influential piece of context the model reads before it generates anything.

Think of it as the difference between a one-off instruction and a job description. If you tell a colleague "summarise this email," that is a single task. If you tell them "you are our customer support lead; always reply warmly, never make promises about refunds, and escalate anything legal," that is a role they carry into every interaction. The system prompt is the job description, and everything the AI does is coloured by it.

Where the system prompt lives

In a consumer chat tool, the system prompt is often written by the company that built the product, and you never see it. When you build your own AI feature or assistant, however, you write the system prompt yourself, and it becomes one of the most important design decisions you make. It is set once and applies to the whole conversation, which is what gives it such reach.

Why system prompts matter so much

Because the system prompt is read first and frames everything that follows, it has a disproportionate influence on the model's behaviour. A vague system prompt produces a vague, generic assistant. A precise one produces an assistant that sounds like your brand, stays inside your rules, and refuses politely when asked to do something it should not. The same underlying model can feel completely different depending on the instructions it is given, much as the same talented person performs very differently depending on how well they are briefed.

System prompts also do quiet but vital safety work. They are where you tell the model what it must never do: never share internal pricing, never give medical or legal advice, never invent information when it does not know an answer. This connects closely to the broader challenge of why AI models sometimes make things up, because a well-written system prompt can instruct the model to admit uncertainty rather than guess, which dramatically reduces confident-sounding errors.

The building blocks of a good system prompt

Most effective system prompts cover a handful of recurring elements. They establish a role or persona so the model knows whose voice to adopt. They state the goal so the model knows what success looks like. They set boundaries, listing what is off-limits. They define tone and format, so replies arrive in the style and structure you expect. And they often include a few examples, because showing the model what good looks like is frequently more effective than describing it. These elements work together, and leaving one out usually shows up as a weakness in the output.

How to write an effective system prompt

The best system prompts are specific, positive, and ordered sensibly. Specific beats vague every time: "reply in three short paragraphs, avoiding jargon" works far better than "be helpful." Positive instructions, telling the model what to do, generally land better than a long list of prohibitions, though a few firm boundaries are essential. And order matters, because the most important instructions are best placed early where they carry the most weight. Many of the same habits from everyday prompt engineering apply here, just at a higher, more permanent level.

It also pays to be concrete about edge cases. Tell the model what to do when it does not know an answer, when a request falls outside its remit, and when a user becomes frustrated. Spelling these out turns an unpredictable assistant into a dependable one. For more advanced control, techniques covered in our guide to advanced prompting can be applied inside a system prompt to handle complex multi-step behaviour.

Iterate, do not perfect

No one writes a flawless system prompt on the first try. The reliable approach is to draft a version, test it against real questions, watch where it drifts or breaks, and refine the wording. Each round tightens the behaviour. Treat the system prompt as a living document that improves as you learn how your users actually behave, rather than a one-off you set and forget.

System prompts in AI agents and assistants

As businesses move from simple chat tools to more capable AI agents that can take actions, the system prompt becomes even more important, because it now governs not just what the AI says but what it is allowed to do. When an assistant can look up an order, send an email, or update a record, the system prompt is where you define which of those tools it may use and under what conditions. This is closely tied to how AI agents use tools, and a clear system prompt is the difference between an agent that acts responsibly and one that oversteps.

If you are building a practical assistant for your business, the system prompt sits at the heart of the wider job of integrating AI agents with your business tools. It is the layer where your policies, your tone, and your guardrails all come together, and getting it right early saves a great deal of correction later.

Common mistakes to avoid

The most frequent error is vagueness: a system prompt so general that the model has nothing concrete to hold onto. The second is overload, cramming in so many rules and contradictions that the model cannot honour them all and quietly ignores some. A third is forgetting to test, then being surprised when the assistant behaves oddly in situations the prompt never anticipated. Keeping the prompt focused, internally consistent, and well-tested avoids nearly all of these. When you do need to combine several capabilities, it is often cleaner to break the work into stages using prompt chaining rather than overstuffing a single instruction.

Used well, the system prompt is the quiet lever that turns a generic model into an assistant that genuinely reflects your business. Spend time on it, test it against reality, and revisit it as you learn, and you will get far more reliable and on-brand results than any amount of clever in-the-moment prompting can deliver. If you would like help designing an AI assistant around your own rules and tone, you can always get in touch with our team.

Frequently asked questions

References

1. OpenAI. "Prompt engineering guide." platform.openai.com.
2. Anthropic. "Prompt engineering documentation." docs.anthropic.com.