Zero-shot Prompting

What Does “Zero-shot” Mean?

In a zero-shot prompt, you ask the model to complete a task without showing task-specific examples. “Zero-shot” refers to the number of examples supplied for that task: zero.

Classify this support message as billing, technical, shipping, or account.

“My card was charged twice” → billing
“Where is my package?” → shipping

Now classify: “I cannot reset my password.”

Zero-shot vs. Few-shot

Both approaches can solve the same task. The main difference is whether examples are included in the prompt.

Extract the order ID from this message.
Return only the ID.

“Order #A102” → A102
“Ref B908” → B908

Extract the ID from: “Order #C441”

How a Zero-shot Prompt Works

The model already has broad capabilities from training. Your prompt specifies what you want it to do with the current input.

For example: Classify the message into exactly one of four labels. Return only the label. followed by the customer message is a zero-shot classification prompt.

Example 1: Zero-shot Classification

Classification is a common zero-shot use case because the labels can be stated directly in the instruction.

Production tip: If the labels are similar, define each label in the prompt. Clear label definitions are still zero-shot as long as you do not provide task-specific input/output demonstrations.

Example 2: Zero-shot Extraction

You can ask a model to extract information without demonstrating the extraction on previous examples.

Run a Real Zero-shot Experiment in Python

Try the same customer messages with a clear zero-shot classification instruction. The code supplies no task-specific examples.

import os
from openai import OpenAI

client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])

messages = [
    "My card was charged twice for one order.",
    "The tracking page says delivered, but my package is missing.",
]

instruction = """Classify each customer message as exactly one of:
billing, technical, shipping, account.
Return only the label. Do not add an explanation."""

for message in messages:
    response = client.responses.create(
        model="gpt-5-mini",
        input=[
            {"role": "system", "content": instruction},
            {"role": "user", "content": message},
        ],
    )
    print(message, "->", response.output_text)

Important: This is zero-shot because there are no examples such as “message → label” in the prompt. Keep the API key in an environment variable and validate the returned label in production code.

Clear Instructions Matter More in Zero-shot

When you remove examples, the instruction carries more of the burden. Compare a vague request with a precise one.

Classify this.

Classify the message as billing, technical, shipping, or account. Return exactly one label and nothing else.

When Zero-shot Works Well

Zero-shot is especially useful when the task is familiar, the instruction is unambiguous, and the desired output is easy to describe.

When Zero-shot May Struggle

Zero-shot is not automatically worse than few-shot. The right choice depends on task complexity, consistency requirements, prompt length, cost, and evaluation results.

Zero-shot Is Not “No Context”

Zero-shot describes the absence of task-specific examples, not the absence of useful information.

For example, “Using this return policy, decide whether this order is eligible. Return yes or no.” can be zero-shot even though the policy is included. There are still zero task-specific demonstrations.

Zero-shot + Context

This is where the previous lessons connect. You can provide relevant context and still use a zero-shot prompt.

Mini-Project: Zero-shot Support Router

Build a small support-routing workflow that sends each incoming message to the right team.