3.4 KiB
Core Prompt Engineering Techniques
Anthropic's official prompt engineering techniques for Claude AI models.
1. Be Clear and Direct
What it is: Provide explicit, unambiguous instructions
Why it works: Claude 4.x models are trained for precise instruction following
When to use:
- Any prompt where clarity matters
- Production systems requiring consistent output
- Complex tasks with specific requirements
How to apply:
- State exactly what you want
- Specify format, length, and tone
- Avoid vague or ambiguous language
- Include specific constraints and requirements
Examples:
❌ Bad:
Tell me about Python
✅ Good:
Write a 300-word explanation of Python's list comprehensions for intermediate developers.
Include:
- 3 practical examples
- Performance considerations
- Common pitfalls to avoid
2. Use XML Tags for Structure
What it is: Organize prompts with XML tags like <instructions>, <example>, <context>
Why it works: Claude was trained with XML tags, naturally recognizing them as structural elements
When to use:
- Separating data from instructions
- Complex prompts with multiple sections
- When working with variable content
- Production systems requiring clear boundaries
Common tags:
<instructions>- Main task description<example>- Few-shot examples<context>- Background information<document>- Input data to process<output_format>- Expected result format<constraints>- Limitations and rules
Example:
<instructions>
Analyze the customer review below and extract structured sentiment data.
</instructions>
<review>
The product arrived damaged and customer service was unhelpful.
</review>
<output_format>
Return JSON with:
- "sentiment": "positive" | "negative" | "neutral"
- "confidence": 0.0 to 1.0
- "key_issues": array of strings
</output_format>
3. Chain of Thought (CoT)
What it is: Ask Claude to think step-by-step before providing final answer
Why it works: Breaking down problems leads to more accurate, nuanced, and reliable responses
When to use:
- Math and logical reasoning problems
- Complex analysis tasks
- Multi-step processes
- When accuracy is critical
- Debugging and troubleshooting
How to apply:
- Add "Think step by step" to your prompt
- Request reasoning in
<thinking>tags - Ask Claude to show its work
- Guide the thinking process with specific steps
Examples:
Basic CoT:
Solve this problem step by step:
<problem>
A store sells apples for $2 each. If you buy 5 or more, you get 20% off.
How much do 7 apples cost?
</problem>
Think through this step by step, showing your work.
Structured thinking:
Analyze this code for bugs. Use the following process:
<thinking>
1. Read and understand the code structure
2. Identify potential issues
3. Assess severity of each issue
4. Recommend fixes
</thinking>
<code>
[code here]
</code>
4. Prefilling Claude's Response
What it is: Provide the beginning of Claude's response to guide output format and style
Why it works: Immediately directs the response in the desired direction
When to use:
- Forcing specific output formats (especially JSON)
- Controlling tone and style
- Ensuring responses start correctly
- Avoiding preambles
Examples:
JSON output:
User: Extract the name, email, and phone from this text: [text]