Core Principles

• Be Clear & Detailed - Treat Claude like a new intern; explicit instructions always win
• Show Your Work - Use scratchpads/thinking tags (hidden from users but improves accuracy)
• Structure Everything - XML tags are your friend (Claude was trained on them)
• Test & Iterate - Scientific approach: test prompts repeatedly, refine constantly

Key Techniques

Few-Shot Prompting

• Provide realistic examples of input → ideal output
• Show 2-5 examples before your actual task
• Use concrete, specific examples (not generic ones)

Chain-of-Thought

• Add "Think step by step" after instructions
• Use <thinking> tags to show reasoning
• Breaks complex tasks into manageable steps

XML Formatting

<instructions>Your task here</instructions>
<examples>Your examples</examples>
<input>Actual content</input>

• Clear separation of prompt sections
• Trained specifically for this format

Role Assignment

• "You are an expert [X]..." at the start
• Sets context and tone
• Helps shape response style

System Prompt Structure

1. Role/Persona - Who is Claude?

2. Task Description - What should it do?

3. Context - Background information

4. Examples - Show don't tell

5. Instructions - Step-by-step guidance

6. Output Format - Specify format needed

Production Best Practices

✅ DO:

• Use scratchpads for complex reasoning
• Provide grounding data (RAG)
• Give clear success criteria
• Repeat critical instructions
• Version and A/B test prompts

❌ DON'T:

• Use vague instructions
• Skip examples for complex tasks
• Forget to specify output format
• Leave edge cases undefined

Collaboration Model

Prompt Engineers + Subject Matter Experts = Best Results

• Engineers handle structure/optimization
• SMEs provide domain knowledge/examples
• Iterate together until success criteria met

Quick Wins

1. Add "Think step by step in <thinking> tags"
2. Provide 3 concrete examples
3. Use XML to structure your prompt
4. Specify exact output format wanted
5. Give Claude a relevant role/persona