Learn AI /Large Language Models /System prompt

System prompt

Also called: system message, system instruction.

Detail

In five seconds

What it is

A high-priority instruction that sets the assistant’s role, rules and priorities before it responds to later messages.

In plain words

It is the behind-the-scenes instruction that tells a chat assistant how to behave before it answers you.

An everyday example

A moderator is about to introduce speakers at an international youth forum. Before the audience sends written questions, the organiser hands the moderator a briefing note: keep introductions under one minute, use respectful language, avoid naming private individuals, and redirect off-topic questions. Then audience questions arrive, one after another. The moderator answers in a way shaped by the briefing note, even though the audience cannot see all of it.

Where the story stops working

A human moderator can deliberately interpret, question or refuse the organiser’s note with clear judgement. A language model does not follow instructions through human understanding; it generates text based on patterns in the full input. Also, real systems may add other controls beyond the system prompt, such as filters or tool restrictions, so the briefing note is only part of the picture.

Remember thisA system prompt gives the model high-priority standing instructions, but it is still context, not fine-tuning, an access control or a guarantee that every instruction will be followed.

Understand it

Why it matters

If you use or buy AI tools, the system prompt affects what the assistant tends to do, what it refuses, how it speaks and what risks remain. That matters for quality, safety, privacy and accountability. A well-designed system prompt can reduce confusion, set boundaries and make outputs more useful. A weak one can lead to off-task answers, policy breaches or overconfident text. For non-specialists, the key decision point is simple: do not assume the visible user message is the whole instruction set, and do not assume a system prompt alone makes an application safe or reliable.

How a system prompt shapes a reply

Show that the system prompt is earlier input that guides the model’s reply, but does not guarantee obedience.

01
1. System instruction added
The application places a high-priority instruction before the conversation, such as role, style, task limits or safety rules.
Draw a top input box labelled 'System prompt' with example chips: 'be concise', 'protect privacy', 'ask clarifying questions'.
02
2. Earlier messages included
Previous turns may also be included so the model can continue the exchange consistently.
Draw a second input box labelled 'Conversation history' beneath the first, with stacked message lines.
03
3. User message arrives
The user adds a request or question. It is part of the same context, but not usually the highest-priority instruction.
Draw a third input box labelled 'User message' with one sample line such as 'Summarise this in 3 bullets'.
04
4. Model generates text
The model processes all provided text together and predicts the next tokens for the reply.
Merge arrows from the three input boxes into one central box labelled 'LLM generates next tokens from full context'.
05
5. Reply reflects the instruction mix
The response is influenced by the system prompt, the prior messages and the user request. If they conflict, the application often expects the system instruction to take priority, but failures can happen.
Draw an output box labelled 'Assistant reply' with badges such as 'brief', 'safe', 'asked for clarification'.

A left-to-right flow diagram with three input boxes feeding one model box, then one output box. Read from top to bottom on the input side: system prompt first, then conversation history, then user message. Arrows merge into the model, then continue to the reply. Small priority labels show that system instructions usually outrank user instructions in the application logic.

Hovering or tapping each input box highlights the phrases in the output it most plausibly influenced; a toggle shows a second case where a conflicting user request tries to override the system instruction.

Describe this without the diagram

The diagram explains a chat system as a flow of text into a language model. First comes a system prompt containing high-priority instructions like role, style and safety limits. Second comes any earlier conversation history. Third comes the user’s latest message. These three text sources flow into the model together. The model then generates a reply token by token from the whole context. The final reply often reflects the system prompt strongly, for example by staying brief or asking for clarification, but the diagram also notes that following the system prompt is not guaranteed.

With reduced motion: A static flow chart shows three stacked inputs on the left: system prompt, conversation history and user message. Arrows merge into one model box in the centre, then point to one assistant reply box on the right. A note states: 'Earlier system instructions usually have higher priority in the application, but the model may still fail or be influenced by conflicting text.'

In many chat systems, messages are not all treated the same way. A system prompt is a special instruction placed before the user’s message. It usually tells the model what role to take, what style to use, what safety limits apply, and which tasks matter most.

What goes in? Usually the application sends a hidden or partly hidden instruction such as “answer briefly”, “do not reveal private data”, or “ask for clarification if the request is ambiguous”. Then the user’s message is added after that. Both become part of the model’s context, meaning the text available for generating the next output.

What happens? The model processes the whole sequence of text and predicts likely next tokens, meaning small pieces of text such as words or parts of words. The system prompt does not physically control the model like a switch. It influences the output because it is part of the input text the model is conditioned on, often with higher priority in the application’s rules.

What comes out is a response shaped by those earlier instructions. If the system prompt says “reply in Spanish” and the user asks a simple question in English, the output may still be in Spanish. But this is not a guarantee. The model can still produce poor answers, ignore instructions, or be manipulated by later text if protections are weak.

A system prompt is often confused with a guardrail or fine-tuning. A guardrail is a broader control mechanism around the model, such as filtering or policy checks. Fine-tuning changes model behaviour by further training on data. A system prompt does not retrain the model; it gives instructions for one interaction or session.

How “The briefing note before a public event” maps onto the real thing

In the story	In the AI system
briefing note from the organiser	system prompt sent before the user message
audience questions	user prompts or later messages
moderator’s replies shaped by the note	model output influenced by earlier instructions
rule to avoid naming private individuals	safety or policy instruction in the system prompt

A worked example

An approved internal assistant is configured to draft a concise working-group update from a non-sensitive section of an implementing partner’s progress report.

Starting from: System prompt: "You assist with working-group updates. Return exactly three bullets: delivery, constraints and next step. Attribute reported claims, preserve uncertainty and add no facts." User message: "The implementing partner reports that two consultations were completed. The attendance total is awaiting validation. The next activity was moved pending venue confirmation."

The application supplies the system prompt in the model’s context before generation, together with the later user message.
The model conditions its response on both texts, giving the higher-priority instructions precedence over ordinary user wording.
The requested structure determines the three bullet categories, while the attribution rule keeps the implementing partner as the source of the claim.
The uncertainty instruction leads the model to retain the qualification about the attendance total rather than presenting it as confirmed.

Result: • Delivery: The implementing partner reports that two consultations were completed. • Constraints: The attendance total is awaiting validation, and the venue is not yet confirmed. • Next step: The next activity will proceed after venue confirmation.

In a United Nations workplace

A political affairs officer uses an approved internal assistant to turn non-sensitive excerpts from an implementing partner’s progress report into three bullets for a UN Country Team working group. The assistant’s system prompt requires neutral wording, clear attribution and an explicit flag wherever a figure is still awaiting validation.

The system prompt sets standing instructions before the officer submits the excerpt. It can shape role, format and treatment of uncertainty across requests, without the officer repeating those requirements each time.

Before you try thisA system prompt is not an access control or a confidentiality safeguard. The officer must not enter personal data, protection information or security-sensitive details unless the tool is approved for that material. The output still requires human review against the source report.

Where you may hear it

During an ICT governance board review of an internal assistant
At a vendor demo in a procurement meeting
In clearance of terms of reference for a chatbot pilot
During a technical review before a tool is connected to internal files

Avoid confusion

Often confused with

Common misconceptions

Common misconception: The system prompt is the same thing as the user prompt, just written first.
Correction: Both are text instructions, but they play different roles. The system prompt is usually inserted by the application and is meant to set the assistant’s role, limits and priorities across the interaction. The user prompt is the person’s request. Many systems treat conflicts in favour of the system prompt, though not perfectly.
Common misconception: If the system prompt says something clearly, the model will always obey it.
Correction: A system prompt strongly influences output, but it is not a hard guarantee. Models can still fail, drift, follow conflicting text, or produce unsafe or incorrect content. Real reliability usually needs extra controls such as filtering, tool permissions, testing and human oversight.
Common misconception: A good system prompt changes the model itself.
Correction: It does not retrain or alter the model’s stored weights. It changes the current interaction by adding instructions into the context. Once that session ends, the underlying model is the same unless separate training or fine-tuning has been done.

Quick check

Which statement best describes a system prompt?

Go deeper

The simple version

It is the behind-the-scenes instruction that tells a chat assistant how to behave before it answers you.

The technical version

In a chat-formatted LLM application, a system prompt is a high-priority text segment prepended or otherwise supplied as part of the model’s context before generation. It typically encodes role, behavioural constraints, output format, tool-use policy and safety instructions. The model then generates a completion by conditioning on the entire token sequence, which may include system, developer, tool and user messages depending on the platform. The system prompt is not the same as fine-tuning, because it does not modify parameters or model weights. It is also not identical to a guardrail, because a guardrail may involve external policy classifiers, retrieval constraints, moderation layers or execution limits outside the prompt itself. In practice, instruction priority is partly a product design choice of the application and partly an empirical behaviour of the model, which means compliance can be strong but not absolute. This is why system prompts are useful but insufficient as a sole safety mechanism, and why they are relevant to prompt injection risks: hostile text later in the context may try to override or manipulate earlier instructions.

Reviewed external resources

EncyclopaediaWikipedia search: System prompt (opens in a new tab)Wikipedia
This search results page helps readers find a range of Wikipedia entries related to system prompts, and suits people who want a broad starting point rather than one fixed article.
search results, not one pagelink checked 2026-07-31
CourseLLM Course (opens in a new tab)Hugging Face
This course gives structured teaching on how large language model systems work, and suits learners who want a more technical but guided explanation.
Deeperlink checked 2026-07-31
VideoLarge language models, explained visually (opens in a new tab)3Blue1Brown
This video offers a visual explanation of large language models, and suits readers who learn best from intuition and diagrams before formal details.
Intermediatelink checked 2026-07-31
VideoVideo explanations: System prompt (opens in a new tab)YouTube search
This search results page helps readers compare different video explanations of system prompts, and suits people looking for multiple teaching styles rather than one source.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: System prompt (opens in a new tab)Reddit search
This search results page surfaces community discussions and practical questions about system prompts, and suits readers who want to see common confusions and practitioner experiences.
search results, not one pagelink checked 2026-07-31

Where this came from

Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

The one thing to remember

A system prompt gives the model high-priority standing instructions, but it is still context, not fine-tuning, an access control or a guarantee that every instruction will be followed.