Initial commit
This commit is contained in:
Zhongwei Li
@@ -0,0 +1,322 @@
|
||||
# Claude 4.x Best Practices
|
||||
|
||||
Specific best practices for Claude 4.x models (Opus 4, Sonnet 4.5, Haiku 4).
|
||||
|
||||
## What's Different in Claude 4.x
|
||||
|
||||
Claude 4.x models have been trained for **more precise instruction following** than previous generations.
|
||||
|
||||
**Key improvements:**
|
||||
- Better adherence to explicit instructions
|
||||
- Enhanced attention to details and examples
|
||||
- Improved thinking capabilities for complex reasoning
|
||||
- More consistent output formatting
|
||||
- Better handling of structured data
|
||||
|
||||
---
|
||||
|
||||
## Best Practice 1: Be Explicitly Clear
|
||||
|
||||
Claude 4.x responds especially well to clear, explicit instructions.
|
||||
|
||||
**Do this:**
|
||||
```
|
||||
Write a Python function that validates email addresses.
|
||||
|
||||
Requirements:
|
||||
- Function name: validate_email
|
||||
- Input: string
|
||||
- Output: boolean
|
||||
- Use regex for validation
|
||||
- Support international domains
|
||||
- Include docstring with examples
|
||||
```
|
||||
|
||||
**Not this:**
|
||||
```
|
||||
Make an email validator in Python
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Best Practice 2: Provide Context and Motivation
|
||||
|
||||
Explaining *why* helps Claude 4.x understand goals and deliver targeted responses.
|
||||
|
||||
**Good:**
|
||||
```
|
||||
I'm building a user registration system for a SaaS application.
|
||||
I need to validate email addresses to ensure users provide legitimate contact info.
|
||||
|
||||
Write a robust email validation function in Python that:
|
||||
[requirements]
|
||||
```
|
||||
|
||||
**Better understanding** → **Better results**
|
||||
|
||||
---
|
||||
|
||||
## Best Practice 3: Ensure Examples Align
|
||||
|
||||
Claude 4.x pays close attention to examples as part of instruction following.
|
||||
|
||||
**Critical:** Examples must perfectly match desired behavior
|
||||
|
||||
❌ **Misaligned example:**
|
||||
```xml
|
||||
<instructions>
|
||||
Extract dates in ISO format (YYYY-MM-DD)
|
||||
</instructions>
|
||||
|
||||
<example>
|
||||
Input: "Meeting on January 5th, 2024"
|
||||
Output: 01/05/2024 <!-- Wrong format! -->
|
||||
</example>
|
||||
```
|
||||
|
||||
✅ **Aligned example:**
|
||||
```xml
|
||||
<instructions>
|
||||
Extract dates in ISO format (YYYY-MM-DD)
|
||||
</instructions>
|
||||
|
||||
<example>
|
||||
Input: "Meeting on January 5th, 2024"
|
||||
Output: 2024-01-05
|
||||
</example>
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Best Practice 4: Leverage Thinking Capabilities
|
||||
|
||||
Claude 4.x has enhanced thinking capabilities for complex tasks.
|
||||
|
||||
**Use thinking for:**
|
||||
- Multi-step reasoning
|
||||
- Complex analysis
|
||||
- Problem-solving
|
||||
- Reflection after tool use
|
||||
|
||||
**Example:**
|
||||
```xml
|
||||
<task>
|
||||
Analyze this codebase architecture and suggest improvements.
|
||||
</task>
|
||||
|
||||
<instructions>
|
||||
First, think through the analysis in <thinking> tags:
|
||||
1. What patterns are being used?
|
||||
2. What are the strengths?
|
||||
3. What are the weaknesses?
|
||||
4. What improvements would have the most impact?
|
||||
|
||||
Then provide your recommendations.
|
||||
</instructions>
|
||||
|
||||
<codebase>
|
||||
[code here]
|
||||
</codebase>
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Best Practice 5: Guide Initial Thinking
|
||||
|
||||
You can guide Claude's thinking process for better results.
|
||||
|
||||
**Example:**
|
||||
```
|
||||
Solve this algorithm problem. Before coding, think through:
|
||||
|
||||
<thinking_guide>
|
||||
1. What is the core problem?
|
||||
2. What data structures would be efficient?
|
||||
3. What's the time complexity target?
|
||||
4. Are there edge cases to consider?
|
||||
</thinking_guide>
|
||||
|
||||
Then implement the solution.
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Best Practice 6: Use Structured Prompts
|
||||
|
||||
Claude 4.x excels with well-organized prompts.
|
||||
|
||||
**Template:**
|
||||
```xml
|
||||
<role>
|
||||
You are a senior software engineer reviewing code.
|
||||
</role>
|
||||
|
||||
<task>
|
||||
Review the following pull request for security vulnerabilities.
|
||||
</task>
|
||||
|
||||
<code>
|
||||
[code here]
|
||||
</code>
|
||||
|
||||
<focus_areas>
|
||||
- SQL injection risks
|
||||
- XSS vulnerabilities
|
||||
- Authentication issues
|
||||
- Input validation
|
||||
</focus_areas>
|
||||
|
||||
<output_format>
|
||||
For each issue found:
|
||||
1. Severity: Critical/High/Medium/Low
|
||||
2. Location: File and line number
|
||||
3. Description: What's wrong
|
||||
4. Recommendation: How to fix
|
||||
</output_format>
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Best Practice 7: Specify Precision Level
|
||||
|
||||
Claude 4.x can adjust precision based on your needs.
|
||||
|
||||
**For high accuracy:**
|
||||
```
|
||||
Analyze this data with high precision. Double-check all calculations.
|
||||
If uncertain about anything, state your confidence level.
|
||||
```
|
||||
|
||||
**For creative tasks:**
|
||||
```
|
||||
Generate creative marketing slogans. Prioritize originality and impact.
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Best Practice 8: Test and Iterate
|
||||
|
||||
Claude 4.x's consistency makes testing more reliable.
|
||||
|
||||
**Recommended approach:**
|
||||
1. Create test cases covering:
|
||||
- Happy path
|
||||
- Edge cases
|
||||
- Error conditions
|
||||
- Boundary values
|
||||
|
||||
2. Run prompt against all test cases
|
||||
|
||||
3. Identify patterns in failures
|
||||
|
||||
4. Refine prompt based on failures
|
||||
|
||||
5. Repeat until quality threshold met
|
||||
|
||||
---
|
||||
|
||||
## Model Selection Guide
|
||||
|
||||
Choose the right Claude 4.x model for your use case:
|
||||
|
||||
### Opus 4
|
||||
**Best for:**
|
||||
- Highest accuracy requirements
|
||||
- Complex reasoning tasks
|
||||
- Mission-critical applications
|
||||
- Creative writing
|
||||
- In-depth analysis
|
||||
|
||||
**Example use cases:**
|
||||
- Legal document analysis
|
||||
- Complex code generation
|
||||
- Research synthesis
|
||||
- Strategic planning
|
||||
|
||||
### Sonnet 4.5
|
||||
**Best for:**
|
||||
- Balance of performance and cost
|
||||
- Most production applications
|
||||
- Real-time interactions
|
||||
- General-purpose tasks
|
||||
|
||||
**Example use cases:**
|
||||
- Chatbots
|
||||
- Content generation
|
||||
- Code review
|
||||
- Data extraction
|
||||
|
||||
### Haiku 4
|
||||
**Best for:**
|
||||
- Speed-critical applications
|
||||
- High-volume processing
|
||||
- Simpler tasks
|
||||
- Cost optimization
|
||||
|
||||
**Example use cases:**
|
||||
- Classification
|
||||
- Simple extraction
|
||||
- Quick summaries
|
||||
- Moderation
|
||||
|
||||
---
|
||||
|
||||
## Performance Optimization
|
||||
|
||||
### Token Efficiency
|
||||
|
||||
**Minimize unnecessary tokens:**
|
||||
```xml
|
||||
<!-- Inefficient -->
|
||||
<instructions>
|
||||
I would like you to please analyze the following text
|
||||
and then extract any email addresses that you find
|
||||
in the text and return them to me in a list format.
|
||||
</instructions>
|
||||
|
||||
<!-- Efficient -->
|
||||
<instructions>
|
||||
Extract all email addresses from the text below.
|
||||
Return as a JSON array.
|
||||
</instructions>
|
||||
```
|
||||
|
||||
### Batching
|
||||
|
||||
**For multiple similar tasks:**
|
||||
```xml
|
||||
<instructions>
|
||||
Classify each review as positive/negative/neutral.
|
||||
</instructions>
|
||||
|
||||
<reviews>
|
||||
<review id="1">[text]</review>
|
||||
<review id="2">[text]</review>
|
||||
<review id="3">[text]</review>
|
||||
</reviews>
|
||||
|
||||
<output_format>
|
||||
[
|
||||
{"id": "1", "sentiment": "positive"},
|
||||
{"id": "2", "sentiment": "negative"},
|
||||
...
|
||||
]
|
||||
</output_format>
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Production Readiness Checklist
|
||||
|
||||
Before deploying prompts in production:
|
||||
|
||||
- [ ] Tested on diverse inputs
|
||||
- [ ] Edge cases handled
|
||||
- [ ] Output format strictly controlled
|
||||
- [ ] Examples align with instructions
|
||||
- [ ] Context minimized to essentials
|
||||
- [ ] Error handling specified
|
||||
- [ ] Token usage optimized
|
||||
- [ ] Model version specified
|
||||
- [ ] Success metrics defined
|
||||
- [ ] Fallback behavior defined
|
||||
Reference in New Issue
Block a user