Zero-Shot Prompting

Zero-shot prompting means asking a model to do a task using only instructions, with no worked examples. Modern instruction-tuned models handle many tasks zero-shot, so it is the natural place to start.

Zero-Shot Classification

Classify the sentiment of this review as positive, negative or neutral.
Reply with one word only.

Review: "Delivery was fast but the screen arrived cracked."
Sample response:
negative

The prompt names the labels, restricts the answer to one word, and supplies the input. Listing the allowed labels is the most important part; otherwise the model might invent "mixed" or write a paragraph.

Zero-Shot Summarisation

Summarise the following meeting notes in one sentence for a busy manager.

Notes: The team agreed to delay the mobile launch by two weeks because
payment testing found bugs. Marketing will move the campaign. QA needs
one extra tester until the 15th.
Sample response:
The mobile launch is delayed two weeks due to payment bugs, marketing will reschedule, and QA needs one extra tester until the 15th.

Zero-Shot Transformation

Convert this sentence to formal British English:
"gonna need those reports asap, thx"
Sample response:
I will need those reports as soon as possible. Thank you.

When Zero-Shot Is Not Enough

  • The output format is unusual or strict.
  • Labels are subtle or company-specific ("priority P2 vs P3").
  • The model keeps choosing the wrong style or tone.
  • Results are inconsistent across similar inputs.

In those cases, add examples (few-shot prompting, next chapter) or clearer definitions of each label.

Frequently asked questions

What does zero-shot mean in AI?

Performing a task without being shown any examples of it in the prompt, relying only on instructions and prior training.

Is zero-shot or few-shot better?

Few-shot is usually more consistent for specific formats, but zero-shot is simpler, cheaper and often good enough. Start zero-shot and add examples if needed.

How do I make zero-shot outputs consistent?

List the allowed answers, specify the exact format, and use a low temperature.