Course contentsShow
← AI Engineering
Lesson 126 of 1,8864. Prompt Engineering FoundationsFree lesson

Few-Shot Learning with Examples

Provide 2-5 example input-output pairs to teach the model the desired format and behavior through demonstration.

Few-Shot Learning with Examples

What You'll Learn

You'll learn how to teach an LLM the exact behavior you want by showing it 2-5 concrete examples in your prompt, rather than just describing what to do.

The Core Idea

After learning zero-shot prompting (asking without examples), few-shot learning is the next level up: showing the model what you want through demonstration. Instead of describing your desired format or behavior in words, you provide a handful of input-output pairs that serve as a template.

Think of it like training a new employee. Rather than saying "format customer emails professionally," you'd show them 3 examples of good customer emails. The pattern becomes instantly clear.

Why Examples Work Better Than Instructions

LLMs are pattern-matching machines. When you provide examples, you're giving the model a concrete pattern to follow rather than abstract instructions to interpret. This dramatically improves consistency and format adherence.

Structure of a few-shot prompt:

  1. Optional: Brief instruction or context
  2. Example 1: Input → Output
  3. Example 2: Input → Output
  4. Example 3: Input → Output (optional but helpful)
  5. Your actual input (the real task)

How Many Examples?

2-3 examples usually suffice for simple formatting tasks. 4-5 examples help with more complex or nuanced behavior. More isn't always better—it uses more tokens and can sometimes confuse the model.

When to Use Few-Shot vs Zero-Shot

Use few-shot when:

  • Output format must be precise and consistent
  • The task is ambiguous or has multiple valid interpretations
  • Zero-shot results were inconsistent
  • The pattern is easier to show than describe

Key Takeaway: Few-shot learning uses 2-5 concrete input-output examples to demonstrate the exact format and behavior you want, making the model's job easier and outputs more consistent than describing requirements in words alone.