A prompt is instruction text that drives an agent. In a RAG system, it is a distinct artifact that can be improved without retraining the model. section 2.1.1 section 2.6
An offline improvement loop compares prompt variants using a signal, such as an LLM judge, and keeps the better version. The model stays fixed; the prompt changes. section 1.2.1 section 2.6
Lead with the verdict
A support agent answers whether a return after 30 days is allowed. Adding “Begin your reply with a direct yes or no, then explain” makes the answer more direct. section 1.2.1
Before: You are a support assistant. Answer the customer's
question using the policy documents.
After: You are a support assistant. Answer the customer's
question using the policy documents. Begin your reply
with a direct yes or no, then explain.Compare the original support instruction with its revised version. section 1.2.1































































