Prompting Best Practices

Introduction

After thousands of prompts in production, the patterns that consistently win are surprisingly simple. Master these ten and you'll outperform 90% of self-taught prompters.

Beginner analogy: Like cooking — the basics (sharp knives, hot pan, salt your food) matter more than fancy techniques.

Understanding the topic

Core concepts to understand:

• Give the model a role and a goal.
• Be specific — replace vague adjectives with measurable criteria.
• Provide examples when the task is unusual.
• Specify format — JSON, markdown, bullets, length.
• Use delimiters (### or XML) to separate sections.
• Tell the model what to do and what NOT to do.
• Ask for reasoning first, answer last, when correctness matters.
• Set temperature (0 for facts, 0.7+ for creativity).
• Iterate with a small test set, not one example.
• Version your prompts like code.

Syntax reference

Visual workflow / architecture:

   ┌──────────────────────────────┐
   │  Role      → "You are…"       │
   │  Task      → "Your job is…"   │
   │  Context   → "Here's the data"│
   │  Examples  → 1–3 input/output │
   │  Format    → JSON | markdown  │
   │  Constraints → length, tone   │
   │  Reasoning  → "think step by  │
   │              step then answer"│
   └──────────────────────────────┘

Real-world use

Anthropic publishes the same 10 rules in its official Claude prompting guide; OpenAI's GPT best-practices doc echoes them. They aren't tricks — they're the industry standard.

Best practices

• Treat each rule as a checklist — apply all that apply.
• Re-read the prompt as if you've never seen the task.
• If the model's output is bad, edit the prompt — not the data.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Give 5 prompting best practices and why each works.
• Q2. When does asking the model to reason first hurt performance?
• Q3. What's the difference between zero-shot and few-shot prompting?
• Q4. How do you decide on temperature?