On this page
Definition
A system prompt is a hidden or fixed instruction set passed to a language model that establishes its persona, rules and constraints before any user input is processed.
Simple explanation
When you open ChatGPT and type a question, there is often a message you never see — one written by the company or developer — telling the model how to behave: be concise, never give legal advice, always reply in British English, and so on. That is the system prompt.
It works like a job brief given to an employee before their shift starts. The customer only sees the employee's replies, but the brief shapes every one of them.
Why it matters
System prompts are how businesses customise a general-purpose model into a specific product — a customer support bot, a coding assistant, a tutor — without retraining the underlying model.
They also carry safety and brand guardrails: what the assistant should refuse, how it should handle sensitive topics, and what tone it should use consistently across thousands of conversations.
How it works
- 1Set the roleDefine who the assistant is (e.g. 'You are a helpful billing support agent').
- 2Set the rulesState what it must and must not do.
- 3Set the formatSpecify tone, length and structure preferences.
- 4Load onceThe system prompt is sent alongside every message in the conversation.
- 5User messages layer on topThe user's actual questions are processed within the boundaries the system prompt set.
Real examples
Products named for illustration only. Inclusion is not an endorsement.
- ChatGPT custom instructionsLets users set a personal system prompt for how ChatGPT should respond to them.
- Customer support chatbotsCompanies write system prompts restricting a bot to their own products and policies.
- Claude's constitution-style promptsAnthropic layers behavioural guidance into the system prompt used across Claude products.
Advantages
- Lets one base model serve many different products and tones.
- Provides a consistent baseline behaviour across all conversations.
- Can enforce safety rules without changing the model itself.
- Cheaper and faster to update than retraining or fine-tuning.
Limitations
- Users can sometimes get a model to ignore or reveal its system prompt.
- Long or overly restrictive system prompts can make responses feel stilted.
- It competes for space in the context window alongside the conversation.
- It cannot guarantee behaviour — the model may still deviate on edge cases.
Common misunderstandings
- ClaimA system prompt makes a model's behaviour unbreakable.RealityIt strongly influences behaviour but can be bypassed with adversarial prompting, known as a 'jailbreak'.
- ClaimSystem prompts and user prompts carry equal weight.RealityMost models are trained to prioritise system-level instructions over user requests when they conflict.
Frequently asked questions
Can I see a product's system prompt?
Not usually — most companies keep them private, though some can be extracted or are voluntarily published.
Is a system prompt the same as fine-tuning?
No. A system prompt is instructions given at run time; fine-tuning changes the model's underlying weights.
Do all AI chat tools use a system prompt?
Most production AI products do, even if it is invisible to the end user.
Can a system prompt stop a model from making mistakes?
It can reduce certain errors but cannot eliminate hallucinations or guarantee factual accuracy.
How long can a system prompt be?
As long as the model's context window allows, though longer prompts leave less room for the actual conversation.
The Tool Money Lab perspective
The system prompt is one of the highest-leverage, lowest-cost tools available to anyone building on top of an AI model. Getting it right often matters more than which model you choose.
We treat system prompts as living documents, not one-off settings — they need testing, versioning and revisiting as a product's needs change.
Conclusion
A system prompt sets the stage before a conversation with an AI model even begins, shaping its role, tone and rules for everything that follows.
It is a simple mechanism with an outsized effect on how useful and safe an AI product feels in practice, which is why most serious AI products invest real effort in writing and testing theirs.