System Prompts & Personas
Theory
Every chat request arrives as a structured list of messages, each tagged with a role. Before tokenization, role markers wrap each message so the model distinguishes developer intent from user input.
system"You are a concise support assistant. Only answer billing questions."
user"Why was I charged twice this month?"
assistantmodel generates here
| Role | Written by | Purpose |
|---|---|---|
| system | Developer | Instructions, persona, constraints, format |
| user | End user | The human's actual request |
| assistant | Prior model turns | Conversation history the model conditions on |
System prompts shape outputs in three ways: persona, operational constraints, and format.