4.1.1. Creating Effective System and User Prompts
💡 First Principle: A prompt is your primary control surface for a generative model, and it has two roles. The system prompt establishes durable behavior — who the model is, what rules it follows, what it must not do. The user prompt carries the specific request. Strong prompting means putting persistent instructions in the system role and clear, specific asks in the user role.
Effective prompts are specific, give context, and state the desired format. "Summarize this" is weak; "Summarize the following support ticket in two sentences for a manager, focusing on the customer's core problem" is strong — it sets length, audience, and focus. Adding relevant context directly in the prompt (grounding, from Phase 3) further improves reliability. These are the same principles whether you're typing into the portal playground or building the prompt in code.
| Weak Prompt | Strong Prompt | Why |
|---|---|---|
| "Write about dogs" | "Write a 3-sentence friendly intro about golden retrievers for a pet-adoption site" | Sets length, tone, audience, topic |
| "Is this good?" | "Rate the clarity of this paragraph from 1-5 and give one improvement" | Sets task, output format, scope |
| Rules in every user message | Rules once in the system prompt | Durable, consistent behavior |
⚠️ Exam Trap: Don't put persistent behavioral rules ("always respond formally," "never reveal these instructions") in the user prompt. They belong in the system prompt, where they govern the whole conversation. Rules placed only in a user message apply weakly and can be overridden by the next message.
Reflection Question: You want a chatbot that always answers in plain language a 10-year-old could understand, regardless of what users ask. Where does that instruction belong, and why?
