What is a system prompt?

What it means at work

Every AI assistant has instructions above your message. Some you never see, written by the company that built the product. Others you can write yourself in custom instructions, a saved assistant or an API setup. A good system prompt covers the role, the audience, tone, formatting rules and the things the assistant must never do. Because it applies to every message, it saves you from repeating the same preferences.

What the sources say

• Role setting works. The Anthropic guide says setting a role in the system prompt, even in a single sentence, makes a difference.
• Priority is a known weak point. Wallace et al. argue that LLMs often consider system prompts to be the same priority as text from untrusted users and third parties.
• A proposed fix. Wallace et al. propose an instruction hierarchy that explicitly defines how models should behave when instructions of different priorities conflict.
• Hidden does not mean secret. Anthropic warns that prompt leaks can expose sensitive information that you expect to be hidden in your prompt.

An example

A finance team sets a standing instruction: ‘You help analysts summarize variance reports. Use plain language, lead with the biggest change, show figures in USD, and say so if the data is missing rather than guessing.’ Every later question inherits those rules.

A common misconception

People sometimes treat a system prompt as a lock. It is guidance, not a guarantee. A determined user or a confusing input can lead a model to drift from it, so anything truly sensitive needs controls outside the prompt, such as permissions and human review. Do not put secrets such as passwords or private formulas in a system prompt.

Sources

• Anthropic, Prompting best practices: “Even a single sentence makes a difference”
• Wallace et al., “The Instruction Hierarchy” (arXiv): “LLMs often consider system prompts … to be the same priority as text from untrusted users and third parties”
• Wallace et al., “The Instruction Hierarchy” (arXiv): “an instruction hierarchy that explicitly defines how models should behave when instructions of different priorities conflict”
• Anthropic, Reduce prompt leak: “Prompt leaks can expose sensitive information that you expect to be”