Suppose a restaurant owner asks a language model to mark every meat-free dish on a pasted menu. The menu supplies items to work on, but no example of a correctly marked dish. This is a zero-shot task: the model must follow the instruction without a task demonstration in the prompt. section 2.14
In zero-shot prompting, a direct instruction draws on what the model learned during pretraining. “Zero-shot” describes the examples supplied for this request; it does not mean the model has never encountered related language or ideas. Few-shot prompting instead includes a small number of demonstrations to show the desired pattern or format. section 7.3 section 2.14
Table 7.2 Foundational prompting techniques
| Technique |
Description |
Example use case |
Example prompt |
| Zero-shot |
Provide an instruction without examples |
Language translation, Q&A |
Translate the following sentence into Spanish: “Good morning, how are you?” |
| Few-shot |
Include a small set of examples in the prompt |
Summarization, classification, language translation, code generation |
Translate: “Hello” → “Hola”; “Goodbye” → “Adiós”; Now translate: “How are you?” |
| Multi-turn |
Carry information forward across multiple exchanges |
Customer support, tutoring agents |
User: My internet isn’t working. Model: What device are you using? User: MacBook Pro. Model: Let’s run a few tests. |
Compare a direct translation instruction with a translation prompt that supplies example pairs before asking for a new answer. section 7.3
A direct request may work well, but examples can help when the response does not follow the intended format. The term also appears outside prompting: DetectGPT was described as a zero-shot method for detecting AI-generated text because it did not require labeled examples. Its reported performance fell when the model used to assess text differed from the one that generated it. Zero-shot therefore describes what examples a method does without, not a guarantee of accuracy. section 2.14 section 6.9
