Few-Shot Prompting

Introduction

Few-shot means including 1-5 worked examples in the prompt. It dramatically improves quality on niche tasks, custom formats, and rare labels. Cost goes up because examples repeat every call.

Beginner analogy: showing a new hire two solved tickets before asking them to handle a third.

Understanding the topic

Core concepts:

• Examples teach format and style.
• 3-5 diverse examples usually beat 1.
• Order matters — put the most-similar example last.
• Use a consistent delimiter (`---`, `###`).
• Cache prompts to amortise example token cost.

Syntax reference

Visual workflow / architecture:

System: You are a JSON intent classifier.
Examples:
Q: cancel flight to NYC → {"intent":"cancel","target":"NYC"}
Q: book a table for 4 → {"intent":"book","party":4}
---
Q: refund my order #842
A: {"intent":"refund","order":842}

Real-world use

Used in classification, structured extraction, code translation, agent tool-selection prompts.

Best practices

• Pick diverse, high-quality examples.
• Pin the prompt + examples in cached input (Anthropic / OpenAI).

Common mistakes

• Stale examples — keep updating as your data drifts.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. How many shots is usually enough?
• Q2. Why does example order matter?
• Q3. How do you control few-shot cost?