System prompt vs user prompt: the differences
The system prompt is the standing instruction that defines an AI assistant's role, tone and limits, while the user prompt is the one-off request made at each exchange. The first is written once and applies to every conversation; the second changes with every question. This article explains what separates them, how to write each one and how to combine them in a small business. It extends our complete guide to prompt engineering for SMEs and our article on the anatomy of a good prompt.
Definition: two levels of instruction for the same model
The system prompt is the set of instructions given to the model before the conversation; the user prompt is the message the person types during it. An LLM receives both in the same stream of text, but with different labels ("system" and "user" in most programming interfaces). Models are trained to give more weight to system-level instructions.
A simple picture: the system prompt is the job description handed to a new employee on day one (who you are, how you speak to our customers, what you never do). The user prompt is the task you give them on Tuesday at 2 pm. You do not rewrite the job description for every task, and you do not slip a task into the job description.
You already use a system prompt without seeing it. When you open ChatGPT, Claude or Copilot, the vendor has already loaded its own operating instructions, which you do not read. Your message comes after.
The concrete differences between the two
The difference comes down to three criteria: who writes it, how often it changes and what it contains. The table below compares them.
| Criterion | System prompt | User prompt |
|---|---|---|
| Who writes it | A manager, an AI lead, an administrator | Each team member |
| Frequency | Written once, revised from time to time | Rewritten for every request |
| Content | Role, tone, rules, limits, default format | Specific task, today's data, context |
| Scope | All conversations of one assistant | A single exchange |
| Visible to the user | Rarely | Always |
| Example | "Always use formal address, never quote a price" | "Summarize this quote in five lines" |
A practical consequence: if you notice that your team members all paste the same introduction into their messages ("You are a sales assistant, you write formally..."), that is a signal. That part should move into the system prompt.
Where to find the system prompt depending on the tool
Depending on the tool, the system prompt goes by a different name, but it always plays the same role. In consumer apps it hides behind settings; in a programming interface it is written explicitly.
- In ChatGPT: "custom instructions" and the instructions of a custom GPT.
- In Claude: a project's "instructions" and profile preferences.
- In Copilot and business tools: the agent or assistant configuration set by the administrator.
- Programmatically (API): a dedicated field for the system message, separate from user messages. The exact field name varies by provider; check the documentation of the tool you use [To verify before sending: current labels at OpenAI, Anthropic and Google].
For a small business, the most common case is the first one: creating a shared assistant (a GPT or a Claude project) whose system prompt is written once by the AI lead, then used by the whole team.
How to write a good system prompt
A good system prompt fits on one page, describes stable rules and also says what not to do. Keep the same blocks as for an ordinary prompt (role, context, constraints, format), but only what never changes.
Here are the five elements to include:
- The role and the audience: "You assist the administrative team of a 12-person nonprofit."
- The tone and language: formal address, English, short sentences.
- The core rules: authorized sources, company vocabulary, mandatory mentions.
- The prohibitions: do not invent figures, do not give legal advice, do not process health data.
- The behavior when in doubt: ask a question rather than guess, or write [TO COMPLETE].
An example for a customer follow-up assistant: "You write payment reminder emails for the accounting department of a services company. You always use a formal tone. You never write a reproach. You invent no amount and no date: if data is missing, you write [TO COMPLETE]. You answer with an email subject, then the body of the message."
With this framework in place, the team member only has to write: "Invoice 2024-118, Dupont LLC, 2,400 EUR, 18 days overdue, first reminder." The amounts in this example are illustrative.
How to write the user prompt when a system prompt exists
The user prompt works best when it stays short: the task, today's data and the expected result. Whatever is already in the system prompt does not need repeating.
Three habits:
- Do not restate the role, it is already set.
- Provide the data specific to the request (names, amounts, a document excerpt).
- Only specify what departs from the default rule: "This time, a firmer tone, it is the third reminder."
Watch out for conflicting instructions. If the system prompt says "answer in three lines maximum" and the user asks for a detailed report, the model arbitrates, and not always the way you hope. In general the system instruction wins, but the result varies from one model to another. Test edge cases before rolling the assistant out.
Security: why the separation matters in a business
The system prompt is not a safe: a confidential instruction placed in it can be revealed or bypassed. So-called prompt injection attacks consist of slipping, into a message or a document the AI reads, an instruction that tries to override the system's rules. A received email containing "ignore your previous instructions" is the typical example.
Three precautions:
- Never write a password, an access key or confidential data into a system prompt.
- Do not give an assistant more rights than necessary (read-only rather than send rights, for instance).
- Have a human review any irreversible action (sending, paying, deleting).
These questions belong to usage governance. Our article on data confidentiality with AI sets the basic rules, and the one on shadow AI shows what happens when everyone configures their tool on their own. A company-approved system prompt is also a natural part of your AI policy.
Managing system prompts over time
A system prompt is a working document: it gets versioned, tested and revised. Treat it like an internal procedure.
- Name it and date it ("Reminders assistant v2, October 2026").
- Assign an owner, usually the AI lead.
- Keep about ten typical requests to test every change.
- Record what changed and why.
Plan a ten-request test after each revision; this is a field rule of thumb, to adapt to your use. It prevents the classic regression: you fix one flaw and create two others.
As for tokens, note that a very long system prompt is re-read at every exchange. In a pay-per-use tool, it weighs on cost and on the room left for your documents. A dense prompt beats a wordy one.
From good settings to a team skill
Knowing the difference between system prompt and user prompt is one of the skills that separates individual use from team use. It determines the quality, consistency and security of the assistants you deploy.
It is also a compliance topic. Article 4 of the European AI regulation (Regulation EU 2024/1689) asks organizations to ensure a sufficient level of AI literacy among the people using these systems. Our article on the Article 4 training obligation details what it implies. To go further depending on your tool, see our ChatGPT training for business and Claude training. The whole approach is in the prompt engineering guide.
How GrowthPerf supports your teams
GrowthPerf trains your teams to build shared assistants whose system prompt is clear, tested and approved. We are a Qualiopi-certified training provider, which in principle opens the possibility of funding through your OPCO depending on your situation (eligibility to be confirmed case by case).
A typical session starts from a real use in your organization (payment reminders, replies to inbound requests, meeting minutes), drafts the system prompt with you, tests it on your cases and documents it. The program is detailed in our prompt engineering training. To find out where to start, book a free 30-minute AI training audit.
Frequently asked questions
What is the difference between a system prompt and a user prompt? The system prompt sets the assistant's permanent rules (role, tone, limits). The user prompt contains the request of the moment. The first is written once by a manager, the second by each team member at each exchange.
Can you see ChatGPT's or Claude's system prompt? Not the one the vendor configured internally. You can, however, write your own in custom instructions, a project or a custom assistant.
Does the system prompt always take priority? Models generally give it more weight, but it is not an absolute guarantee. A contradictory instruction or an injection can produce unexpected behavior: test your edge cases.
Do you need a system prompt for occasional use? No. For a one-off question, a good user prompt is enough. The system prompt becomes useful as soon as several people repeat the same task.
How long should a system prompt be? One page is a good starting point. Beyond that, check that each rule adds something and remove duplicates.
Can you put confidential information in it? Better avoided. Assume its content can be revealed and include no password, key or personal data.
Who should write the system prompt in a small business? Ideally an AI lead, together with the future users. They know the company's rules and stay responsible for updates.