In zero-shot prompting the model is given only a description of the task. Brown et al. (2020) demonstrated that sufficiently large models do this surprisingly well, and Wei et al. (2022) showed that instruction tuning — fine-tuning on many tasks phrased as instructions — sharply improves zero-shot performance on unseen tasks.
Zero-shot is fast and clean but less reliable for unusual formats or edge cases, where a few examples help.