What Is a System Prompt? How AI Tools Get Their Instructions

A system prompt is a set of instructions given to an AI model before you start your conversation. It tells the model how to behave, what role to play, what rules to follow, and what tone to use. You do not see the system prompt when you use tools like ChatGPT, but it shapes every response the model gives you.

How system prompts work behind the scenes

When you open ChatGPT and type a message, your text is not the only input the model receives. Before your message, there is a system prompt that OpenAI has written. It might say something like: "You are ChatGPT, a helpful assistant. Be concise and accurate. Decline requests for harmful content."

This system prompt is processed as part of the conversation's context, consuming tokens just like your messages do. The model treats it as foundational instructions that apply to the entire conversation. Your messages are then interpreted within the framework those instructions establish.

Think of it this way: if the AI model is an employee, the system prompt is the job description and employee handbook they read on day one. Your individual messages are the tasks you assign throughout the day. The employee follows both, but the handbook sets the baseline expectations.

Every commercial AI tool uses system prompts, even if they do not show them to you. When you use a customer service chatbot on a Kenyan e-commerce site, the system prompt likely specifies the company name, product catalogue, return policy, and response style. That is why the bot "knows" about the company without you providing context.

Why system prompts matter for your work

Without a system prompt, an AI model defaults to generic behaviour. It will be helpful, but in a broad, unfocused way. A good system prompt narrows the model's behaviour to exactly what you need.

Consistency across interactions. If you use AI to draft customer responses for your business, a system prompt ensures every response follows the same tone, format, and policy guidelines. Without it, the model might be formal in one response and casual in the next.

Role specialisation. You can instruct the model to act as a Kenyan tax advisor, a children's English tutor, or a technical support agent for your software product. The system prompt provides the domain context so you do not have to repeat it in every message.

Guardrails and boundaries. A system prompt can tell the model to avoid certain topics, always cite sources, never make up statistics, or stay within a specific scope. This is especially important for business applications where inaccurate or off-topic responses have real consequences.

Writing your own system prompt

If you use the API versions of ChatGPT, Claude, or other models, you can write your own system prompts. Some platforms (like ChatGPT's custom GPTs or Poe's bot creation tool) also let you set system prompts through a user interface.

A practical system prompt has four parts:

Role. State who the AI should be. "You are a customer service assistant for [Company Name], a Nairobi-based delivery service."

Context. Provide background information the model needs. "Our delivery area covers Nairobi, Kiambu, and Machakos counties. Standard delivery takes 2-4 hours. Express delivery takes 1 hour and costs an additional a reasonable cost."

Rules. Set constraints. "Always respond in English. If a customer asks about areas outside our delivery zone, politely explain we do not serve that area yet. Never promise delivery times shorter than our standard."

Format. Specify output structure. "Keep responses under 100 words. Use a friendly but professional tone. End each response by asking if the customer needs anything else."

The more specific your system prompt, the more predictable the model's behaviour. Vague system prompts produce vague results.

System prompts vs regular prompts

Your regular prompts (the messages you type in a conversation) operate within the boundaries set by the system prompt. The system prompt is persistent, applying to every exchange. Your regular prompts are individual requests.

When there is a conflict between the system prompt and a user message, models generally prioritise the system prompt. If the system prompt says "never discuss pricing" and a user asks about pricing, the model will typically decline or redirect. This is why system prompts function as reliable control mechanisms.

However, system prompts are not foolproof. Determined users can sometimes override system prompt instructions through creative prompting techniques. This is known as "jailbreaking" and is an active area of AI safety research.

We explain system prompts alongside other foundational terms in our AI Automation glossary. Related terms include hallucination (what happens when guardrails fail).

FAQ

Can I see the system prompt of a tool I am using?

Usually not. Most commercial tools hide their system prompts. Some users have extracted system prompts through indirect techniques, but providers regularly update their prompts and add protections against extraction.

Do system prompts use up my token limit?

Yes. The system prompt counts toward your conversation's total token usage. A very long system prompt leaves fewer tokens available for your messages and the model's responses. Keep system prompts focused and concise, ideally under 500 words.

Do I need to write a system prompt to use ChatGPT or Claude?

No. The standard chat interfaces include a pre-written system prompt. You only need to write your own if you are building a custom application via the API or creating a custom GPT.