You are building a customer-facing LLM product used by thousands of users with different backgrounds, languages, and intents. Design a system prompt architecture that is robust across this diversity. Cover: how to handle adversarial users attempting prompt injection, how to make the prompt work across different user sophistication levels, how to prevent the model from going off-topic, how to structure instructions for reliability (ordering, specificity, negative examples), and how to test system prompt robustness before shipping.