Few-Shot Prompting

Introduction

Few-shot provides 2–5 examples to teach the model a pattern. It's the workhorse for classification, extraction, and any task with subtle format rules.

Beginner analogy: Showing a new hire three or four sample emails before they write one. They infer tone, structure and word choice from the variety.

Understanding the topic

Core concepts to understand:

• Format: instruction + N examples + new input.
• N = 2–5 is the sweet spot for most tasks (more = diminishing returns).
• Examples should be diverse and cover edge cases.
• Order matters — recent examples are weighted more.
• Place examples BETWEEN instruction and target for best effect.

Syntax reference

Visual workflow / architecture:

Workflow

Instruction
   │
   ▼
Example 1 (input → output)
Example 2 (input → output)
Example 3 (input → output)
   │
   ▼
NEW input → ?

Real-world use

Customer-support classification, invoice extraction, sentiment scoring — all routinely run on 3–5 carefully chosen few-shot examples.

Best practices

• Pick examples that span the variety of inputs the model will see.
• Keep examples short — long ones eat the window.
• Never use examples that contradict each other.

Common mistakes

• Including biased examples that the model amplifies.
• Re-using the same trivial example three times — no extra signal.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. How many examples is 'few'?
• Q2. Why does example order matter?
• Q3. How do you choose few-shot examples?
• Q4. What's the downside of too many examples?