In plain terms
You describe the job and the model does it, relying on what it learned in training. “Translate this into German.” “Is this review positive or negative?” No samples, no demonstrations. It works because modern models have seen so many tasks that most ordinary ones need no introduction.
Why it matters
Zero-shot is the default way people use AI, and its strength is the reason AI became broadly useful: no training data, no setup, results in seconds. Knowing where it runs out is practical knowledge. When the task has a house style, unusual categories or a strict format, instructions alone leave too much to guesswork and examples are needed.
Example
A marketing team asks a model to tag incoming leads as “enterprise”, “mid-market” or “small business” from the company description. Zero-shot gets most of them right. It stumbles on holding companies and franchises, where the team's own convention is not something the model could have guessed.
Most often confused with
Zero-shot Prompting vs. Few-shot Prompting
The “shots” are examples. Zero-shot gives none and relies on the description; few-shot adds a handful of input–output pairs so the model can copy the pattern. Start with zero-shot because it is simplest, and add examples when the output is close but inconsistent.
Under the hood
Zero-shot performance depends on how clearly the task is specified and how common it was in training. Ways to strengthen it without examples: state the purpose and the audience, define each category or criterion, describe the output format exactly, and allow reasoning before the final answer (zero-shot chain of thought). Reasoning models narrow the gap with few-shot on many tasks. The term comes from machine-learning research, where it meant handling classes never seen in training; in prompting it simply means no examples in the prompt.