Glossary

Prompting

Zero-Shot Prompting

Asking a model to perform a task with instructions alone, without showing any worked examples of the expected output.

Definition

Zero-shot prompting is the simplest mode of using a language model: describe the task in natural language and let the model respond. No examples, no demonstrations, no template — just instruction. Modern instruction-tuned models perform surprisingly well in zero-shot settings on common tasks because they have effectively been trained on millions of zero-shot interactions during fine-tuning. The trade-off is that the model fills in any unspecified details using its priors, which may or may not match what you actually wanted.

Why it matters

Zero-shot is the default for most chat use because it is fast and conversational. The hidden cost is that the model's priors do the work you did not specify — so the tone, structure, and depth of the response reflects training-data averages rather than your specific preferences. For routine tasks this is fine. For anything where format, voice, or structure matters, zero-shot is the wrong mode and few-shot prompting consistently outperforms it. The practical heuristic: if you can describe the output in a sentence and a generic answer would satisfy you, zero-shot is appropriate. If you need a specific format, voice, or worked structure, switch to few-shot. The cost of adding two or three good examples is small; the quality lift is often dramatic.

Examples

Good zero-shot use

'Summarise this article in three sentences.' The task is unambiguous, the format is conventional, and any reasonable summary is acceptable.

Risky zero-shot use

'Write a LinkedIn post in our voice.' The model has no idea what 'our voice' means and will default to LinkedIn-average phrasing — exactly the result that makes posts feel generic.

Zero-shot with constraint

'Summarise in exactly three sentences, no adjectives.' Adding a hard constraint narrows the space of acceptable outputs and improves consistency without examples.

Frequently asked

When does zero-shot prompting fail?

When the task depends on conventions the model cannot guess — your house style, an unusual format, or a specific reasoning approach.

Is zero-shot worse than few-shot?

Not always. For well-known tasks with conventional formats, zero-shot is faster and equally good. For specific style or structure, few-shot is consistently better.

Can a system prompt do the work of zero-shot examples?

Partially. A well-written system prompt encodes conventions, but worked examples still anchor format and voice more reliably than description alone.

Related terms

Put the concept to work

Open the rewrite engine and apply this principle to a draft of your own.

Try a rewrite