System Prompts
The instructions that shape how a model behaves before a user says anything
What it is
A system prompt is a special input provided to an LLM that sets context, instructions, and constraints for how it should behave throughout a conversation. It appears before the user's messages and typically has higher priority, models are trained to follow system prompt instructions even when user messages conflict.
System prompts can instruct the model to: respond only in a certain language, adopt a specific persona, limit discussion to certain topics, use a particular format, or be aware of available tools. They're the primary mechanism by which companies customize model behavior for their products.
System prompts are not foolproof, they can be undermined through jailbreaking, and they don't override safety training baked in during RLHF.