Prompting a model to perform a task without providing examples, relying entirely on the model's pre-existing capabilities. Zero-shot performance varies significantly by task; testing often verifies the model can reliably perform intended uses without examples.