← Back to glossary
+Suggest a term
Concept·Patterns & Practices·Added 1 month ago

Zero-shot prompting

Also known as: zero-shot inference, zero-shot

Asking a model to do something with no worked examples in the prompt at all. You rely entirely on the model's pre-trained knowledge and instruction-following ability to get the right output.

Zero-shot prompting is the simplest form of prompting: give the model a task description and let it figure out how to execute it. No examples, no demonstrations. The model draws on everything it learned during training to produce a response.

Modern instruction-tuned models (models fine-tuned specifically to follow directions) handle zero-shot prompting surprisingly well across a wide range of tasks. For straightforward requests, adding examples often doesn't improve results much and just adds cost by consuming more tokens in the context window.

Where zero-shot struggles is with unusual output formats, niche domains, or tasks that require the model to adopt a very specific style. In those cases, a few examples go a long way. Most builders end up treating zero-shot as their default and layering in examples only when quality drops.

This definition is AI-generated and refreshed weekly. It may contain inaccuracies. Use your own judgment, especially for production decisions.
Related terms
Few-shot promptingPrompt engineeringChain-of-thoughtSystem prompt