Few-Shot Prompting

Few-Shot Prompting uses several examples to show the AI what format or type of response you expect. This technique significantly improves results, especially for specific or complex tasks.

How does Few-Shot work?

You provide 2-5 examples before your actual request. The AI learns from these patterns and applies them to your task.

Basic Structure

Example 1: [Input] → [Output]
Example 2: [Input] → [Output]
Example 3: [Input] → [Output]
Your request: [Input] → ?

Practical Applications

Sentiment Analysis

Text: "The service was fast and friendly"
Sentiment: Positive

Text: "The product arrived damaged"
Sentiment: Negative

Text: "The delivery was on time"
Sentiment: Neutral

Text: "I'm thrilled with the quality and price"
Sentiment:

Output: Positive

Data Formatting

Input: Max Mustermann, Berlin, 35 years
Output: {"name": "Max Mustermann", "city": "Berlin", "age": 35}

Input: Anna Schmidt, Munich, 28 years
Output: {"name": "Anna Schmidt", "city": "Munich", "age": 28}

Input: Tom Weber, Hamburg, 42 years
Output:

Output: {"name": "Tom Weber", "city": "Hamburg", "age": 42}

Email Categorization

Email: "When will my order be delivered?"
Category: Delivery inquiry

Email: "I want to cancel my order"
Category: Cancellation

Email: "The product has a defect"
Category: Complaint

Email: "Do you have this product in blue?"
Category:

Output: Product inquiry

Advantages of Few-Shot

Best Practices for Few-Shot

1. Quality of Examples

✅ Good examples:
- Diverse and representative
- Clearly structured
- Error-free

❌ Bad examples:
- Too similar to each other
- Inconsistent format
- Contains errors

2. Number of Examples

• 2-3 examples: For simple tasks
• 3-5 examples: For more complex patterns
• 5+ examples: Rarely necessary, can be counterproductive

3. Order

Tip: Arrange examples from simple to complex
1. Clear-cut case
2. Typical case
3. Edge case

Advanced Techniques

Negative Examples

Show what is NOT desired:

Correct: "Dear Mrs. Schmidt,"
Incorrect: "Hey Schmidt,"

Correct: "Best regards"
Incorrect: "BR"

Few-Shot with Explanations

Text: "The food was cold"
Sentiment: Negative
Explanation: Cold food indicates poor quality

Text: "The price is fair"
Sentiment: Neutral
Explanation: "Fair" is neither particularly positive nor negative

Chain-of-Thought in Few-Shot

Question: "If Peter has 3 apples and buys 2 more, how many does he have?"
Thought process: Peter starts with 3 apples. He buys 2 more. 3 + 2 = 5
Answer: 5 apples

Question: "Maria has 10€ and spends 4€. How much is left?"
Thought process: Maria starts with 10€. She spends 4€. 10 - 4 = 6
Answer: 6€

Avoiding Common Mistakes

When to Use Few-Shot

✅ Ideal for:

• Specific formatting requirements
• Company-specific classifications
• Consistent outputs across multiple requests
• New or unusual tasks

❌ Less suitable for:

• Simple, standardized tasks
• When you don't have good examples
• Very creative or open-ended tasks

Practical Template

# Few-Shot Template for [Your Task]

## Example 1
Input: [Sample input]
Output: [Desired output]

## Example 2
Input: [Sample input]
Output: [Desired output]

## Example 3
Input: [Sample input]
Output: [Desired output]

## Your Request
Input: [Your actual input]
Output:

Exercise

Create a Few-Shot prompt for:

1. Product descriptions in your company style
2. Categorization of customer inquiries
3. Converting technical language into simple language