3.2.2. Prompts as Configuration

💡 First Principle: The prompt is configuration you write in plain language. A system prompt sets the model's persistent role, rules, and constraints for the whole conversation; the user prompt is each specific request. Separating them gives you durable control: the system prompt governs behavior, the user prompt supplies the task.

A second powerful lever is grounding — supplying the model with relevant, trusted information (retrieved documents, data, context) inside the prompt so its output is based on that material rather than only its training-time patterns. Grounding is the primary practical defense against hallucination: instead of hoping the model "knows," you give it the facts to work from. This directly connects responsible AI (reliability) to a concrete technique you'll use in Foundry.

⚠️ Exam Trap: The system prompt and user prompt are not interchangeable. Putting your durable rules in each user message (instead of the system prompt) makes behavior inconsistent and easy to override. Persistent role and constraints belong in the system prompt.

Reflection Question: Why is grounding a more reliable way to get accurate answers than simply trusting the model's built-in knowledge? Tie this back to how generative models produce text.