Zero-Shot Prompting

Introduction

Zero-shot means asking the model with no examples — just the instruction. Modern LLMs are strong zero-shot for common tasks (classify, summarise, translate). When zero-shot works, ship it: it's the cheapest, simplest prompt.

Beginner analogy: asking an experienced bartender for a Negroni — no recipe needed.

Understanding the topic

Core concepts:

• Pure instruction, no examples.
• Works best for well-known tasks (translation, summarisation, common classification).
• Cheapest tokens — no examples to repeat each call.
• Falls over on niche tasks or custom output schemas.
• Always try zero-shot first; add examples only if quality is low.

Syntax reference

Visual workflow / architecture:

Prompt: "Translate to French: 'Hello, world!'"
            │
            ▼
        LLM (no examples)
            │
            ▼
       "Bonjour le monde !

Real-world use

Most product chat features (summarise this PR, classify this email) use zero-shot first.

Best practices

• Start every prompt zero-shot — measure before adding shots.
• Add format hints to fix zero-shot output shape.

Common mistakes

• Defaulting to few-shot for trivial tasks — wasted tokens.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. When does zero-shot beat few-shot?
• Q2. Why is zero-shot cheaper?
• Q3. Failure mode unique to zero-shot?