Zero-Shot Prompting

Introduction

Zero-shot means asking the model to do a task with no examples — just the instruction. It's the simplest prompt style and surprisingly powerful for well-known tasks (summarise, translate, classify).

Beginner analogy: Asking a barista "flat white please" — no recipe needed because they already know. Zero-shot works when the model already 'knows' the task from training.

Understanding the topic

Core concepts to understand:

• Format: just the instruction + the input.
• Best for common tasks: translation, classification, summarisation.
• Cheap on tokens — no examples to ship.
• Fails when the task is novel, domain-specific, or has a unique format.
• Foundation for all other techniques.

Syntax reference

Visual workflow / architecture:

┌──────────────────────────────┐
│  Instruction: "Translate to  │
│  French: 'Hello world'"      │
└──────────┬───────────────────┘
           ▼
        ┌─────┐
        │ LLM │
        └──┬──┘
           ▼
    "Bonjour le monde"

Real-world use

ChatGPT's default chat is zero-shot 95% of the time. Most successful prompts in production are zero-shot with great instructions.

Best practices

• Try zero-shot first — only add examples if it fails.
• State the format explicitly even with no examples.
• Use action verbs: classify, summarise, extract, translate.

Common mistakes

• Using zero-shot for tasks the model has never seen and being surprised.
• Vague verbs like 'analyse' that the model interprets differently every time.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Define zero-shot prompting.
• Q2. When does zero-shot fail?
• Q3. Why is zero-shot cheaper than few-shot?
• Q4. How do you improve a zero-shot prompt without adding examples?