In plain words
When you open an AI app and type a message, the AI has usually already read something else first: a block of instructions written by whoever built the app. That block is the system prompt.
It can set a role ("you are a patient maths tutor for children"), rules ("give hints, not answers"), tone, and facts the AI should know, such as today's date. The same underlying AI can behave like a strict examiner in one app and a friendly travel guide in another, purely because of this text.
You usually never see it. Some apps let you add your own standing instructions, often called custom instructions, which work in a similar way: they are quietly attached to every chat you start.
Everyday example
A briefing before the shift
Before a new waiter starts at a restaurant, the manager says: "Greet everyone in Hindi or English, recommend the thali, never promise a table in under 20 minutes." Customers never hear this briefing, but it shapes every conversation the waiter has that evening. A system prompt is that briefing, given to the AI before each chat.
Where the example breaks: A trained waiter remembers the briefing. An AI does not remember anything between calls: the app has to hand it the whole briefing again with every single message.
How it actually works
Chat models receive a list of messages, each tagged with a role. The system prompt is the first one.
Messages come with roles
A conversation is sent to the model as a list: a system (or developer) message, then alternating user and assistant messages. The model is trained to treat the system message as the app maker's rules.
The system message ranks higher
OpenAI's guide says developer messages are "prioritized ahead of user messages". If a user asks for something the system prompt forbids, a well-trained model should follow the system prompt. OpenAI's Model Spec calls this ordering the chain of command.
It is resent every turn
Language models keep no memory between calls. The app sends the system prompt again, plus the conversation so far, for every reply. It uses up part of the context window each time.
Writing one
Keep it clear and specific: role, audience, rules with reasons, and the output format. A short example: "You are a patient maths tutor for 10 year olds. Give one hint at a time and ask the child to try the next step. Never state the final answer."
- Also called
- System message, developer message (OpenAI's newer name), custom instructions (the user-written version)
- Custom instructions
- ChatGPT added them in July 2023, so users could set preferences once for all future chats
- Published openly
- Anthropic publishes the system prompts used in the Claude apps. They give the model up-to-date details such as the current date
- Not a lock
- Clever messages can sometimes override or reveal a system prompt, through jailbreaks or prompt injection
Questions people ask
What is the difference between a system prompt and a user prompt?
The system prompt is written by the app maker (or by you, in settings) and sets standing rules for the whole conversation. A user prompt is each message you type. The model reads the system prompt first and is trained to give it more weight.
Can you see ChatGPT's system prompt?
Not inside the app. Some companies publish theirs: Anthropic posts the system prompts used in its Claude apps. People also try to coax models into revealing them, but the results are not always genuine or complete.
Are custom instructions the same as a system prompt?
Close. Custom instructions are text you write in settings, and the app adds it to the hidden instructions it sends with each chat. The app's own system prompt still applies alongside yours and normally takes priority if they conflict.
Can a system prompt keep secrets safe?
No. Treat it as guidance, not a vault. Determined users can sometimes make a model repeat its instructions or ignore them. Anything truly private, like passwords or keys, should never be placed in a system prompt.