Few-Shot Learning and Examples

Few-shot learning is a powerful prompting technique where you provide examples of the desired input-output pattern before making your actual request. This approach dramatically improves response quality, consistency, and accuracy by showing the AI exactly what you want.

What is few-shot prompting?

Few-shot prompting involves including one or more examples in your prompt to demonstrate the pattern, style, format, or approach you want the AI to follow.

The term “few-shot” refers to the number of examples:

Zero-shot: No examples (just instructions)
One-shot: One example
Few-shot: Multiple examples (typically 2-5)
Many-shot: Many examples (6+)

Why few-shot prompting works

Examples are often more precise than descriptions. Consider these approaches:

Zero-shot (instructions only)
Few-shot (with examples)

Convert this function name to kebab-case.

Function: getUserProfileData

Response may vary in interpretation.

Convert these function names to kebab-case:

Example 1:
Input: getUserProfileData
Output: get-user-profile-data

Example 2:
Input: calculateTotalAmount
Output: calculate-total-amount

Example 3:
Input: validateEmailAddress
Output: validate-email-address

Now convert:
Input: fetchOrderHistoryById
Output:

Response will follow the demonstrated pattern exactly.

Few-shot prompting is especially effective for tasks involving specific formats, consistent styling, or domain-specific conventions.

Basic few-shot structure

Single example (one-shot)

Use one example for simple pattern demonstration:

Transform this error message to be user-friendly.

Example:
Technical: "ERR_CONN_REFUSED: Connection refused at port 5432"
User-friendly: "We're having trouble connecting to our database. Please try again in a moment."

Now transform:
Technical: "TypeError: Cannot read property 'id' of undefined at line 42"
User-friendly:

Multiple examples (few-shot)

Use 2-5 examples for complex patterns:

Generate API endpoint documentation following this pattern:

Example 1:
Endpoint: POST /api/users
Description: Creates a new user account with the provided details.
Authentication: Required (Bearer token)
Rate limit: 10 requests per minute

Example 2:
Endpoint: GET /api/users/:id
Description: Retrieves detailed information for a specific user by their ID.
Authentication: Required (Bearer token)
Rate limit: 100 requests per minute

Example 3:
Endpoint: DELETE /api/users/:id
Description: Permanently deletes a user account and all associated data.
Authentication: Required (Bearer token + Admin role)
Rate limit: 5 requests per minute

Now document this endpoint:
Endpoint: PATCH /api/users/:id/preferences
[Provide details about updating user preferences]

Few-shot prompting by use case

Code generation

Demonstrate coding style, structure, and conventions:

Write React hooks following these examples:

Example 1:
```javascript
export const useCounter = (initialValue = 0) => {
  const [count, setCount] = useState(initialValue);
  
  const increment = useCallback(() => setCount(c => c + 1), []);
  const decrement = useCallback(() => setCount(c => c - 1), []);
  const reset = useCallback(() => setCount(initialValue), [initialValue]);
  
  return { count, increment, decrement, reset };
};

Data transformation

Show the exact transformation pattern:

Transform these database records into API responses:

Example 1:
Database record:
{
  user_id: 123,
  first_name: "John",
  last_name: "Doe",
  email_address: "[email protected]",
  created_at: "2024-01-15T10:30:00Z",
  is_active: 1
}

API response:
{
  id: 123,
  name: "John Doe",
  email: "[email protected]",
  createdAt: "2024-01-15T10:30:00Z",
  status: "active"
}

Example 2:
Database record:
{
  user_id: 456,
  first_name: "Jane",
  last_name: "Smith",
  email_address: "[email protected]",
  created_at: "2024-02-20T14:45:00Z",
  is_active: 0
}

API response:
{
  id: 456,
  name: "Jane Smith",
  email: "[email protected]",
  createdAt: "2024-02-20T14:45:00Z",
  status: "inactive"
}

Now transform this record:
{
  user_id: 789,
  first_name: "Bob",
  last_name: "Johnson",
  email_address: "[email protected]",
  created_at: "2024-03-10T09:15:00Z",
  is_active: 1
}

Content writing

Demonstrate tone, style, and structure:

Write product feature announcements in this style:

Example 1:
Feature: Dark Mode
We've added dark mode to reduce eye strain during late-night coding sessions. Toggle between light and dark themes in your settings, and your preference will sync across all your devices. Your eyes (and sleep schedule) will thank you.

Example 2:
Feature: Collaborative Editing
Work together in real-time with your team. See cursor positions, edits, and comments as they happen. No more "did you get my latest changes?" messages cluttering your Slack channels.

Example 3:
Feature: Smart Search
Find anything instantly with our new semantic search. Type natural language queries like "authentication bug from last week" instead of hunting through titles and tags. It's like having a conversation with your project history.

Now write an announcement for:
Feature: Auto-save with Version History

Test case generation

Show test structure and coverage patterns:

Write test cases following these examples:

Example 1:
describe('calculateDiscount', () => {
  it('should apply 10% discount for orders over $100', () => {
    const result = calculateDiscount(150);
    expect(result).toBe(135);
  });
  
  it('should apply no discount for orders under $100', () => {
    const result = calculateDiscount(75);
    expect(result).toBe(75);
  });
  
  it('should handle zero amount', () => {
    const result = calculateDiscount(0);
    expect(result).toBe(0);
  });
});

Example 2:
describe('validateEmail', () => {
  it('should return true for valid email addresses', () => {
    expect(validateEmail('[email protected]')).toBe(true);
    expect(validateEmail('[email protected]')).toBe(true);
  });
  
  it('should return false for invalid email addresses', () => {
    expect(validateEmail('invalid')).toBe(false);
    expect(validateEmail('@example.com')).toBe(false);
  });
  
  it('should return false for empty string', () => {
    expect(validateEmail('')).toBe(false);
  });
});

Now write tests for:
function calculateShipping(weight: number, distance: number): number

Advanced few-shot techniques

Progressive complexity

Start with simple examples and increase complexity:

Extract structured data from text following this pattern:

Example 1 (Simple):
Text: "Meeting with John tomorrow at 2pm"
Output:
{
  "type": "meeting",
  "participants": ["John"],
  "date": "tomorrow",
  "time": "14:00"
}

Example 2 (Multiple participants):
Text: "Team standup with Sarah, Mike, and Lisa on Friday at 9am"
Output:
{
  "type": "meeting",
  "participants": ["Sarah", "Mike", "Lisa"],
  "date": "Friday",
  "time": "09:00"
}

Example 3 (With location):
Text: "Product review meeting with the design team next Monday at 3pm in Conference Room B"
Output:
{
  "type": "meeting",
  "participants": ["design team"],
  "date": "next Monday",
  "time": "15:00",
  "location": "Conference Room B"
}

Now extract from:
Text: "Quarterly planning session with all department heads on March 15th at 10:30am via Zoom"

Contrastive examples

Show both correct and incorrect patterns:

Refactor code to use proper TypeScript types:

❌ Incorrect:
function processUser(user: any) {
  return user.name.toUpperCase();
}

✅ Correct:
interface User {
  name: string;
  email: string;
}

function processUser(user: User): string {
  return user.name.toUpperCase();
}

❌ Incorrect:
const config: any = {
  apiUrl: process.env.API_URL,
  timeout: process.env.TIMEOUT
};

✅ Correct:
interface Config {
  apiUrl: string;
  timeout: number;
}

const config: Config = {
  apiUrl: process.env.API_URL || 'https://api.example.com',
  timeout: parseInt(process.env.TIMEOUT || '5000', 10)
};

Now refactor this code:
function fetchData(endpoint: any, options: any) {
  return fetch(endpoint, options);
}

Domain-specific examples

Provide examples using domain terminology:

Generate SQL queries for an e-commerce database:

Example 1 - Find high-value customers:
Query: "customers who spent over $1000 in the last 6 months"
SQL:
SELECT 
  c.customer_id,
  c.email,
  SUM(o.total_amount) as total_spent
FROM customers c
JOIN orders o ON c.customer_id = o.customer_id
WHERE o.order_date >= DATE_SUB(CURRENT_DATE, INTERVAL 6 MONTH)
GROUP BY c.customer_id, c.email
HAVING SUM(o.total_amount) > 1000
ORDER BY total_spent DESC;

Example 2 - Popular products:
Query: "top 10 products by units sold this month"
SQL:
SELECT 
  p.product_id,
  p.product_name,
  SUM(oi.quantity) as units_sold
FROM products p
JOIN order_items oi ON p.product_id = oi.product_id
JOIN orders o ON oi.order_id = o.order_id
WHERE MONTH(o.order_date) = MONTH(CURRENT_DATE)
  AND YEAR(o.order_date) = YEAR(CURRENT_DATE)
GROUP BY p.product_id, p.product_name
ORDER BY units_sold DESC
LIMIT 10;

Now generate SQL for:
Query: "customers who abandoned carts in the last 24 hours with cart value over $50"

Few-shot with structured prompts

Combine few-shot learning with structured techniques:

<task>Generate API error responses</task>

<examples>
Example 1:
Input: User not found
Output:
{
  "error": {
    "code": "USER_NOT_FOUND",
    "message": "The requested user does not exist.",
    "statusCode": 404,
    "details": {
      "userId": "{{userId}}"
    }
  }
}

Example 2:
Input: Invalid authentication token
Output:
{
  "error": {
    "code": "INVALID_TOKEN",
    "message": "The authentication token is invalid or expired.",
    "statusCode": 401,
    "details": {
      "tokenType": "Bearer"
    }
  }
}

Example 3:
Input: Rate limit exceeded
Output:
{
  "error": {
    "code": "RATE_LIMIT_EXCEEDED",
    "message": "Too many requests. Please try again later.",
    "statusCode": 429,
    "details": {
      "limit": 100,
      "window": "1 hour",
      "retryAfter": 3600
    }
  }
}
</examples>

<your_task>
Generate error responses for:
1. Insufficient permissions
2. Resource conflict
3. Validation failure
</your_task>

Optimizing example quality

Use diverse examples

Cover different variations and edge cases in your examples to show the full range of expected patterns.

Keep examples realistic

Use realistic, production-quality examples rather than overly simplified ones.

Annotate when helpful

Add comments or explanations to examples when the pattern might not be obvious:

Example:
Input: calculateTotalPrice(items)
Output: items.reduce((sum, item) => sum + item.price * item.quantity, 0)
// Note: Always handle quantity multiplier and use reduce for summing arrays

Balance quantity vs. quality

2-3 high-quality examples usually work better than 10 mediocre ones. More examples aren’t always better.

When to use few-shot prompting

✅ Ideal use cases

Generating code in a specific style or framework
Data transformation with consistent formatting
Content creation matching a particular tone
Test case generation with consistent structure
API responses following a standard format
Documentation with specific templates
Naming conventions and style guides

⚠️ May not be necessary

Well-known standard formats (JSON, CSV)
Common programming tasks with clear conventions
Simple, straightforward requests
When zero-shot instructions are sufficient

Token efficiency considerations

Examples consume tokens. Balance thoroughness with efficiency:

Token management tips:

Use 2-3 examples for most cases (sweet spot for quality vs. tokens)
Reference documentation or specifications for very complex patterns rather than including many examples
Consider creating an assistant with embedded examples for repeated use
See tokens and limits for more optimization strategies

Combining with other techniques

Few-shot prompting works well with other approaches:

Few-shot + Chain-of-thought
Few-shot + Role-based

Solve these debugging problems step by step:

Example 1:
Bug: "React component not re-rendering when state changes"

Analysis:
1. Check if state is being mutated directly (mutation won't trigger re-render)
2. Verify useState is being used correctly
3. Confirm component isn't memoized incorrectly

Solution: Replace `items.push(newItem)` with `setItems([...items, newItem])`
Explanation: Direct array mutation doesn't trigger React re-renders. Use immutable updates.

Example 2:
Bug: "API returns 401 but user is logged in"

Analysis:
1. Check if token is being sent in request
2. Verify token hasn't expired
3. Confirm token is in correct format

Solution: Include token in Authorization header: `Authorization: Bearer ${token}`
Explanation: Token must be sent with each API request using proper header format.

Now debug:
Bug: "Form submission triggers twice on button click"

As a senior code reviewer, provide code review feedback:

Example 1:
Code: `if (user.isActive == true)`
Feedback: ⚠️ **Unnecessary comparison** - `isActive` is already boolean. Use `if (user.isActive)` instead.
Impact: Minor - readability
Suggestion: `if (user.isActive)`

Example 2:
Code: `const data = JSON.parse(response);`
Feedback: 🔴 **Missing error handling** - JSON.parse can throw. Wrap in try-catch.
Impact: High - potential runtime error
Suggestion:
```javascript
try {
  const data = JSON.parse(response);
} catch (error) {
  console.error('Invalid JSON:', error);
}

Now review this code:

async function updateUser(id, data) {
  await fetch(`/api/users/${id}`, {
    method: 'POST',
    body: data
  });
}

```

Saving and reusing examples

Make few-shot patterns reusable:

Custom instructions: Add frequently used examples to custom instructions
Assistants: Create specialized assistants with embedded examples for specific tasks
Prompt library: Maintain a collection of effective few-shot prompts

Next steps

Custom instructions

Save few-shot patterns for reuse

Structured techniques

Combine with structured prompting

Create assistants

Build assistants with embedded examples

Token optimization

Manage example token usage

Overview

Getting Started

Chat Interface

Projects & Teams

Tools & Extensions

Assistants

Prompts & Memory

Integrations & Connectors

Models & Providers

Account & Billing

Troubleshooting

Few-Shot Learning and Examples

What is few-shot prompting?

Why few-shot prompting works

Basic few-shot structure

Single example (one-shot)

Multiple examples (few-shot)

Few-shot prompting by use case

Code generation

Data transformation

Content writing

Test case generation

Advanced few-shot techniques

Progressive complexity

Contrastive examples

Domain-specific examples

Few-shot with structured prompts

Optimizing example quality

When to use few-shot prompting

Token efficiency considerations

Combining with other techniques

Saving and reusing examples

Next steps

Custom instructions

Structured techniques

Create assistants

Token optimization

Overview

Getting Started

Chat Interface

Projects & Teams

Tools & Extensions

Assistants

Prompts & Memory

Integrations & Connectors

Models & Providers

Account & Billing

Troubleshooting

​What is few-shot prompting?

​Why few-shot prompting works

​Basic few-shot structure

​Single example (one-shot)

​Multiple examples (few-shot)

​Few-shot prompting by use case

​Code generation

​Data transformation

​Content writing

​Test case generation

​Advanced few-shot techniques

​Progressive complexity

​Contrastive examples

​Domain-specific examples

​Few-shot with structured prompts

​Optimizing example quality

​When to use few-shot prompting

​Token efficiency considerations

​Combining with other techniques

​Saving and reusing examples

​Next steps

Custom instructions

Structured techniques

Create assistants

Token optimization

What is few-shot prompting?

Why few-shot prompting works

Basic few-shot structure

Single example (one-shot)

Multiple examples (few-shot)

Few-shot prompting by use case

Code generation

Data transformation

Content writing

Test case generation

Advanced few-shot techniques

Progressive complexity

Contrastive examples

Domain-specific examples

Few-shot with structured prompts

Optimizing example quality

When to use few-shot prompting

Token efficiency considerations

Combining with other techniques

Saving and reusing examples

Next steps