Few-Shot Prompting

Introduction

Few-shot prompting means including 2–5 input/output examples in the prompt before asking the model to do the task. It's astonishingly effective — often turning a struggling zero-shot prompt into a reliable one with no fine-tuning required.

Beginner analogy: Showing the intern two completed reports and saying 'now do this one in the same style' — they nail the format every time.

Understanding the topic

Core concepts to understand:

• Provide 2–5 input → output examples that match your real input.
• Examples teach format, tone, edge cases — implicitly.
• Costs more tokens than zero-shot, but improves consistency dramatically.
• Pick diverse examples; don't include only easy cases.

Syntax reference

Visual workflow / architecture:

Few-shot prompt:
  Q: 2 + 2  →  A: 4
  Q: 5 * 3  →  A: 15
  Q: 10 - 7 →  A: 3
  Q: 8 / 2  →  A: ?
                  ▼
                  4

Real-world use

Production classifiers, structured-data extractors and code transformers almost always use few-shot. Tools like DSPy and Promptfoo automate few-shot example selection.

Best practices

• Include 3–5 diverse examples covering edge cases.
• Keep examples short — they consume context window.
• Match example format exactly to expected output.

Common mistakes

• Putting only easy examples → model fails on edge cases.
• Too many examples → context overflow + higher cost.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. What is few-shot prompting?
• Q2. How many examples should you include?
• Q3. When does few-shot beat fine-tuning?