Prompting is a real skill

The same model gives wildly different results depending on how you ask. These patterns reliably improve quality:

• System prompts — set role, tone, and rules up front.
• Be specific — spell out format, audience, and constraints.
• Few-shot — show 2–3 examples of the output you want.
• Chain-of-thought — ask the model to reason step by step for hard tasks.
• Structured output — request JSON so your code can parse it.

System: You are a precise assistant. Reply ONLY as JSON.
User: Extract {name, email, intent} from: "Hi, I'm Sara (sara@x.com), I want a demo."

Takeaway: treat prompts like product specs — clear, testable, and versioned.