The instruction you never see
A system prompt is the standing instruction supplied to a model before the user’s conversation starts. Separate from the prompts you type, it contains rules such as who the assistant is, how it should write, and what it must decline.
Where an assistant’s character comes from
The same underlying LLM behaves very differently depending on the system prompt. The name an assistant uses, its tone, its answering policy, and its list of prohibitions come from the system prompt its developers wrote, together with training. A customer support assistant that identifies with a company and talks only about that company’s products is doing so because of one.
You can write your own
This is not only a developer mechanism. Custom instructions, GPTs, and project features in various services all let users set their own system prompt.
- Standing preferences — answer in English, lead with the conclusion — without repeating them
- Purpose-built assistants, such as an English learning coach
- Fixed working context: company information, document formats
Limits
System prompts are powerful but not absolute. Attempts to make a model break its standing instructions through crafted input are called prompt injection, and they remain a security problem. Contradictory instructions also make behavior unstable, so keep your own short, clear, and consistent. Noticing that you write the same preamble every time, and promoting it to a system prompt, is the step that separates casual from practiced use.
If you have typed the same preface three times in chat, that is the signal. Opening the settings screen of the assistant you use most is worth doing once.