Question

A developer is building a customer-support assistant. They want every response—regardless of which support agent is chatting—to consistently follow the company's refund policy, maintain a friendly tone, and never promise delivery dates. The conversation itself will span many back-and-forth turns with different customer questions. Where should the developer place these persistent behavioral instructions in the Messages API request?

Accepted Answer

In the system prompt, so the instructions apply consistently across all turns of the conversation

Answer

Repeated inside every user message so the model re-reads them each turn

Answer

In an assistant-role message at the start, prefilling the desired behavior

Answer

In a stop_sequences parameter to enforce the policy boundaries