What Is a System Prompt?

How it works

When you chat with an AI app, the model receives more than what you type. Each request typically bundles the system prompt, the conversation so far, your newest message, and often other material such as uploaded files, retrieved documents, or tool descriptions. The system prompt is placed first and labeled separately from your messages, so the model can treat it as the frame for the whole exchange. Depending on the provider, developers supply it in a dedicated field or as a specially labeled message. That is why you may also see it called a system message, system instructions, or (in some tools) a developer message.

Because it travels with every turn, the system prompt uses part of the model’s context window, the limited amount of text it can consider at once. A very long system prompt leaves less room for documents and conversation, and can add cost, since many AI services charge by the amount of text processed.

Many models are trained to give system instructions more weight than requests typed into the chat. That is a tendency, not a guarantee, which is why the limits below matter.

System prompt vs. your own prompts

Both give the model instructions, but they play different roles:

• The system prompt is set in advance, typically by the company or team that built the tool. It usually stays the same for the whole conversation and is shared by everyone using the tool, though apps often add details for each user, such as the date or account information.
• A user prompt is what you type in the moment: one question or one task.

Think of the system prompt as the briefing a new hire gets before their first shift, and user prompts as the individual requests that come in during it.

You can also write your own standing instructions. Many AI assistants offer features such as custom instructions, or instructions attached to a project or custom assistant. You describe your role, preferences, and preferred format once, and the assistant applies them to each new chat. These usually sit alongside the provider’s own system prompt rather than replacing it. Writing any of these well draws on the same skills as prompt engineering.

What a good business system prompt includes

For a customer-facing or internal assistant, a clear system prompt usually covers:

• Role and audience: who the assistant is and who it serves.
• Scope: what it should help with, and what it should decline or hand to a person.
• Tone and format: how formal to be, how long answers should run, and when to use lists.
• Rules: specific things it must always or never do, rather than vague values.
• Background information: facts it needs, such as hours, policies, or product details, plus an instruction to say it doesn’t know rather than guess.
• Escalation: when and how to pass a conversation to a human.

For an AI agent that can take actions, the system prompt also explains when and how to use each tool. Which tools it can actually reach is set separately in the app’s configuration. Write it in plain, direct language and spell out what you mean rather than expecting the model to infer it. Test it with real questions, including awkward ones, and keep a dated copy of each version so changes are deliberate and easy to undo.

Limitations to know

A system prompt guides behavior; it does not lock it down. Keep these limits in mind:

• It can leak. Users can sometimes coax a model into repeating its instructions word for word. Never put passwords, access keys for other software, private customer data, or anything you wouldn’t want made public in a system prompt.
• It can be overridden. Messages typed into the chat that try to talk the model out of its rules (often called jailbreaking) sometimes work. So can hidden instructions planted in a web page, email, or file the model reads. Both are commonly grouped under the term prompt injection.
• It isn’t followed perfectly. Long or conflicting instructions get applied unevenly, and the model can still make things up.

Treat the system prompt as one layer of protection, not the only one. Pair it with AI guardrails such as content filters, limits on what data and tools the assistant can reach, and human review before high-stakes actions.

Further reading

• Claude 101 WorkshopA beginner-friendly, hands-on Claude workshop that covers custom instructions, Projects, and styles, everyday ways to set your own standing instructions.
• AI EmployeesFor readers planning a business assistant, an example of AI systems set up for specific roles inside the tools a team already uses, with guardrails and support after launch.