What is a system prompt? Explained with a day-one job description

In one sentence: a system prompt is the set of standing instructions a company gives an AI model before any customer types a word, telling it what role to play, how to sound and which rules to follow in every conversation.

The everyday picture

Picture a new hire at a hotel front desk. On day one, the manager hands them a job description. It says who they are, who they serve and how to speak to guests. It also lists the rules: never promise a free upgrade, send billing questions to the accounts office.

Guests never see that page. They just walk up and ask questions. But every answer the new hire gives is shaped by it.

An AI model in a chat app works the same way. The company that built the app is the manager. The system prompt is the job description. You, typing your question, are the guest.

How it works, step by step

1. The app maker writes the system prompt (the manager writes the job description). It sets a role, a tone and rules. Anthropic's prompting guide says setting a role in the system prompt focuses the model's behaviour and tone for your use case.
2. The system prompt goes in first (the new hire reads the job description before the first guest arrives). It sits at the start of the conversation, ahead of anything a user types.
3. A user sends a message (a guest walks up to the desk). The app sends the system prompt and the user's message to the model together.
4. The model answers within the rules (the new hire helps the guest, the way the job description says). If a request clashes with the rules, the rules should win. OpenAI's Model Spec ranks the app maker's instructions above the user's, and the AI company's own rules above both.
5. The same page is used every time (every guest gets the same trained new hire). The model does not remember the job description between conversations. The app simply sends it again with each one.

Where the picture stops

• The new hire may not follow every rule. A real employee who breaks a rule can be warned. A model follows its system prompt most of the time, not every time. OpenAI notes that its production models do not yet fully match its own Model Spec.
• The job description is not locked in a drawer. A clever user can sometimes coax a model into repeating its instructions. OWASP, a non-profit security community, says a system prompt should not be treated as a secret or used as a security control.
• There is no real person reading it. The model does not understand the job the way a person does. It predicts a fitting reply, with the system prompt as part of what it reads.

Why it matters for your business

• One consistent voice. A good system prompt means every customer gets answers in the same tone, with the same boundaries, without staff retyping instructions.
• Fewer risky answers. Rules such as "don't give legal advice" or "hand refunds to a person" lower the chance of the assistant saying something you can't stand behind. They lower it; they don't remove it.
• Keep secrets out of it. OWASP warns against putting passwords, keys or connection details in a system prompt. Real permissions, such as who can see which account, belong in your software, not in the model's instructions.
• Cheaper changes. Changing how the assistant behaves is often a matter of editing the system prompt and testing it, not rebuilding the app.

An example

Imagine a 30-room guesthouse that adds an AI chat assistant to its website. The owner's developer writes a system prompt. It says the assistant is the guesthouse's booking helper, answers in short, friendly sentences and only discusses rooms, prices and local directions. It says to never confirm a booking itself, and to pass complaints to the owner's email.

A guest asks for a discount. The assistant explains the published prices and offers the owner's email, as the system prompt says. Room availability itself is checked by the booking system, not left to the assistant's judgement. The developer keeps the system prompt free of passwords, and tests changes on sample questions before they go live.

Frequently asked questions

What is the difference between a system prompt and a prompt?

A prompt is any message you give an AI model. A system prompt is the standing set of instructions the app maker puts in place first, which shapes every conversation, while your own messages change each time.

Can I see the system prompt of the AI tools I use?

Usually not. Users often don't know what instructions an app has given the model. Some AI companies publish the system prompts of their own chat apps, but most apps built on top of AI models keep theirs private.

Can a user override a system prompt?

They are not supposed to. Major AI models are designed to put the app maker's instructions ahead of the user's. But tricks such as prompt injection (hidden instructions slipped into a message or document) sometimes work, so important limits should also be enforced in the software around the model.

Do I need a system prompt if I just use ChatGPT or Claude?

The company behind the app already sets its own. Many chat apps also let you add custom or project instructions, which work in a similar way for your own conversations.

Related explainers

• What is a prompt? Explained with briefing a new contractor
• What is an LLM? Explained with a new hire who read the whole library
• What is a context window? Explained with a clerk's desk
• What is AI memory? Explained with an assistant's notebook

The DevDrift builds AI assistants with clear rules, approvals and audit trails: see how.

Sources

• Anthropic, Claude Platform Docs: Prompting best practices
• OpenAI: Model Spec (2025-04-11 version), chain of command
• OWASP Gen AI Security Project: LLM07:2025 System Prompt Leakage