Plain-English explanation
When you chat with a character on a companion app, you only see your own messages and the character's replies. Behind the scenes, the platform is also sending the model a block of instructions before the conversation even starts, telling it who it's supposed to be, how it should talk, what it should and shouldn't do, and sometimes formatting rules for its responses. That instruction block is the system prompt, and it's the main technical lever platforms use to turn a general LLM into a specific persona.
How it works
Most LLM APIs support distinct message roles (commonly "system," "user," and "assistant") and treat the system role as higher-priority context that frames how the model should interpret everything else. A typical companion-app system prompt might describe the character's name and personality, instruct the model to always respond in character, set boundaries on content, and specify formatting like using asterisks for actions. This is sent alongside the visible conversation on every single message, since the model has no memory between separate API calls; it re-reads the system prompt and conversation history from scratch each time, within the limits of its context window.
Adult-AI use
The system prompt is where a huge amount of a companion app's actual product behavior lives: its persona enforcement, its content policy (what the character will and won't do), and often disclaimers or safety instructions required by the platform. It's also a common target of "jailbreaking" (users trying to get a model to ignore its system prompt's restrictions), which is one reason some adult platforms instead run a genuinely uncensored model rather than relying on a system prompt alone to hold a content boundary in place.
How it differs from persona
Persona is the concept: a character's identity and personality. System prompt is the technical mechanism most platforms use to implement that persona: the actual written instructions the model receives. A persona could theoretically be built into a model through fine-tuning instead of a system prompt, but character chatbots generally use a system prompt, fine-tuning, or both together.
Limitations and misconceptions
- System prompts aren't unbreakable. Because they're just text competing with the rest of the conversation for the model's attention, sufficiently long or adversarial user input can cause a model to drift from or override its instructions.
- Platform operators may store or review system prompts. Some models can be prompted into revealing or paraphrasing their system prompt, so platforms shouldn't treat it as a secure place to store sensitive logic.