Zero-shot prompting
Giving a model a task with only instructions, no examples, and expecting it to figure it out on the first try.
Think of it like
Asking a smart stranger for directions to a place they've never been: they reason it out from first principles.
Example
You ask a language model "Translate this to French:" and paste text, without showing it any translation examples first.
How it actually works
Zero-shot works because modern LLMs are instruction-tuned — they learn to follow a wide variety of commands during training. It's not magic: the model has seen countless examples of translations in its pretraining data and learned the pattern. Failure happens when the task is truly novel or requires specialized knowledge.
For product teams
The easiest path to new use cases: no labeling, no examples to curate. Works for tasks similar to the model's training distribution.
For engineers
No prompt engineering needed, just clear instructions. Baseline performance is often decent; few-shot or fine-tuning helps only if performance gaps are large.
Related
- Few-shot — Contrasts with few-shot and one-shot, where examples improve performance.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome