Beginner

Few-Shot Prompting

Provide 2-5 examples to show the AI exactly what you want - pattern matching at its finest.

What Is This Technique?

Few-shot prompting teaches the AI by example. Instead of just describing what you want, you show 2-5 concrete examples of input-output pairs. The AI recognizes the pattern and applies it to new inputs.

This technique is incredibly powerful because it removes ambiguity. Rather than interpreting your description, the AI can directly observe what success looks like and replicate it.

Think of it like training someone: instead of explaining a complex process, you demonstrate it a few times. They see the pattern and can apply it to new situations.

When to Use This Technique

When to Use

✓When you need consistent formatting across outputs
✓For specialized or domain-specific tasks
✓When the desired output structure is complex or unusual
✓To establish a specific tone, style, or pattern

When NOT to Use

✗For simple, well-understood tasks
✗When examples would be longer than the actual task
✗For creative tasks where variety is desired
✗When you want the AI to determine the best approach

Real-World Examples

Consistent API Response Format

Without This Technique

Return user data in a consistent format

With This Technique

Format API responses consistently:

Example 1:
Input: User ID 123
Output: {"success": true, "data": {"id": 123, "name": "John Doe", "email": "john@example.com"}, "timestamp": "2024-10-18T10:30:00Z"}

Example 2:
Input: User ID 456
Output: {"success": true, "data": {"id": 456, "name": "Jane Smith", "email": "jane@example.com"}, "timestamp": "2024-10-18T10:31:00Z"}

Now format for User ID 789

Key Improvement:

Shows exact structure, field names, and data types

+95%Quality

1 hourSaved

Code Comment Style

Without This Technique

Add comments to this code

With This Technique

Add comments following this style:

Example 1:
// ============================================
// USER AUTHENTICATION
// ============================================
// Validates user credentials and returns JWT token
// @param email - User email address
// @param password - Plain text password
// @returns Authentication token or error

Example 2:
// ============================================
// DATA PROCESSING
// ============================================
// Transforms raw CSV data into structured JSON
// @param csvData - Raw CSV string
// @param headers - Column headers array
// @returns Parsed JSON array

Now comment this function: [your function]

Key Improvement:

Establishes exact comment format, structure, and documentation style

+90%Quality

45 minutesSaved

Applied in 777-1

"Applied in 777-1: Few-shot prompting proves essential when Micaela (UI Consistency) reviews applications. By providing examples of consistent component styling, the refined prompts achieve visual coherence that zero-shot prompts miss. The WealthView case study demonstrates this technique for chart configurations. Full case study documentation coming soon."

Best Practices

DO: Provide 2-5 examples (sweet spot is 3)

Two examples establish a pattern, three confirm it. More than five usually provides diminishing returns and uses tokens inefficiently.

DO: Show edge cases in your examples

Include examples with different data types, empty values, or special characters to teach the AI how to handle them.

DO: Make examples progressively complex

Start simple, then add complexity. This helps the AI understand both basic and advanced patterns.

DON'T: Use examples that are too similar

If all examples are nearly identical, the AI might not generalize well. Show variation within the pattern.

DON'T: Mix multiple patterns in examples

Keep examples consistent with each other. Mixing patterns confuses the AI about which pattern to follow.

Common Pitfalls to Avoid

⚠Using too few examples (1-2) which doesn't establish a clear pattern
⚠Examples that are too similar, failing to show the AI how to handle variation
⚠Not annotating why each example is structured that way
⚠Including errors or inconsistencies in examples
⚠Using examples that are overly complex, making the pattern hard to extract

Related Techniques

Combine this technique with others for even better results:

Ready to Apply This Technique?

See Few-Shot Prompting in action in our prompt library