What is a system prompt?
Tools, assistants and prompting
A system prompt is a higher-priority instruction that tells an AI assistant how it should behave, such as its role, tone, boundaries, source-use rules or output format. Different platforms use slightly different names, but the practical idea is the same: set the working rules before the user asks for the current task.
Reviewed by Jackie, Head of Learning & Development, Levellers · Last reviewed 8 June 2026
What this means
A system prompt is the instruction layer that sits above the user's current request. It can define the assistant's job, tone, limits, preferred sources and the format it should follow. In OpenAI's current Responses API this higher-priority layer may be expressed as developer instructions, while Google documents the same idea as system instructions.
For a business user, the simplest way to think about it is this: the system prompt sets the working rules, while the user prompt asks for the task of the moment.
Why it matters
System prompts matter when a team wants more consistency than a blank chat box can provide. They are useful for assistants that repeatedly summarise meetings, draft replies, route knowledge or structure outputs in a fixed way.
They also help keep role, tone and format aligned across different users. Anthropic recommends system prompts for role-based consistency, and Google advises placing critical behavioural constraints and output requirements in the system instruction or at the start of the prompt.
How it works
OpenAI says its instructions parameter gives the model high-level guidance on how it should behave, including tone, goals and examples of correct responses. In its API reference, instructions delivered with developer or system roles take precedence over user instructions.
Google's Gemini documentation similarly says system instructions guide model behaviour, and Anthropic shows system prompts being used to set role and personality. In practice, the system prompt is combined with the current task prompt and any supporting context the assistant receives.
Examples
Meeting-summary assistant: always separate decisions, actions, owners, deadlines and risks.
Client-support drafting assistant: write in plain English, stay within approved service language and flag any missing facts before drafting.
Internal knowledge assistant: answer from supplied material, state when the answer is not in the source set and avoid pretending certainty.
Operations assistant: return outputs in a fixed table or JSON shape so the team can reuse them in a wider workflow.
Common misunderstandings
It is not the whole prompt. It sets the broader behaviour, while the user prompt still supplies the job to do.
It is not a guarantee of correctness. A system prompt can guide behaviour, but provider and government guidance still emphasise testing, validation and human control.
It is not the same as governance. This is an operational instruction layer, not a complete control framework. That is an inference from provider documentation, supported by separate government guidance on oversight and lifecycle controls.
It should not become a dumping ground. Too many vague or conflicting rules can make an assistant harder to test and maintain.
Risks and boundaries
The main risk is assuming that higher-priority instructions are enough on their own. They are not. The UK Government AI Playbook says organisations should validate high-risk decisions influenced by AI, test systems before deployment and keep meaningful human control at the right stages.
If the assistant handles personal data or sensitive firm information, teams also need a clear data boundary. ICO guidance says AI use with personal data should follow a risk-based security approach and data minimisation principles. A system prompt can tell a model what to do, but it does not replace the need for secure process design.
What to do next
Define the workflow before writing the system prompt. Decide what the assistant is for, what sources it may use, what a good output looks like and where a person needs to check the result.
Then test the prompt on real examples, keep a version record and adjust only what improves the workflow. OpenAI supports reusable prompt versions, and Anthropic's tooling explicitly supports testing and version tracking for prompt templates.
Have a question or a suggestion, or want to understand how we research and review these guides? Read about our editorial standards and how to reach us.
FAQs
Is a system prompt always called a system prompt?
No. Terminology varies by platform. OpenAI often uses developer instructions in the Responses API, while Google refers to system instructions, but both describe a higher-priority instruction layer.
Can users override a system prompt?
Provider documentation describes system or developer instructions as higher priority than user instructions, but that does not mean the output will always be perfect. Testing and safeguards still matter.
Does a system prompt make an assistant safe?
No. Safe use still depends on data boundaries, review standards, validation and ongoing monitoring.
Sources
UK Government AI Playbook is used for human control and validation boundaries.