System vs User vs Assistant Roles

When you call a chat completion API, you don't just send a string. You send an array of messages, each with a role: system, user, or assistant. The roles matter — they're not labels, they're behavioural switches.

The three roles

[
  { "role": "system", "content": "You are a terse assistant. Reply in under 30 words." },
  { "role": "user", "content": "Explain RAG." },
  { "role": "assistant", "content": "RAG retrieves relevant documents, then asks the LLM to answer using them as context." },
  { "role": "user", "content": "Give an example." }
]

• system — instructions the model treats as persistent rules. Style, persona, constraints, safety policies.
• user — the human's input.
• assistant — past model responses, included to maintain conversation history.

Why the system role is special

System messages have higher precedence than user messages in the model's instruction hierarchy. If a user message says "ignore previous instructions," a well-aligned model will defer to the system message.

This matters for production: put your core rules in system, never in user. A user can override anything in their own message; they can't easily override a system message.

Multi-turn conversations

To maintain a conversation, you must resend the entire message history every turn. The model is stateless — it has no memory of previous calls. Every request includes everything that's been said so far.

This has cost implications. A 50-turn conversation re-sends 49 turns of context every time. Use prompt caching (Level 4) to amortise this.

Common patterns

• One-shot tasks — [system, user] only. No history needed.
• Chatbots — [system, user1, assistant1, user2, assistant2, ...] — full rolling history.
• Agents — [system, user, assistant(tool_call), tool(result), assistant, ...] — interleaved tool calls (Level 3).

The system prompt as a contract

A good system prompt is a tight contract:

1. Identity — "You are a customer support assistant for Acme Corp."
2. Capabilities — "You can answer questions about our products, pricing, and shipping."
3. Constraints — "Do not discuss competitors. Do not promise discounts."
4. Tone — "Friendly, professional, concise. Maximum 3 sentences per reply."
5. Escalation rules — "If the user asks about a refund, respond with: 'Let me connect you to a human agent.'"

These five blocks cover 90% of production system prompts. Most production bugs come from missing one of them.