Zero-shot prompting is a technique where you ask a language model to perform a task by providing only an instruction — no worked examples, no demonstrations, and no task-specific fine-tuning. The model draws entirely on patterns absorbed during pretraining to interpret the request and produce a useful response. It is the simplest prompting strategy and is often the first thing practitioners try before escalating to few-shot or chain-of-thought approaches.

Think of it like asking a well-read colleague to summarise a contract they have never seen, in a language format they have never been explicitly taught. They rely on general knowledge, reading comprehension, and context clues rather than a step-by-step script you handed them beforehand.

A zero-shot prompt might look like:

Classify the following customer review as Positive, Neutral, or Negative.
Review: "The packaging was damaged but the product works fine."

No labelled examples precede the instruction. Because modern LLMs such as GPT-4 and Claude are trained on enormous text corpora, they have encountered enough sentiment-classification discourse to handle this reliably — even without guidance on what "Positive" or "Negative" should look like in context.

Zero-shot prompting sits at the entry point of prompt engineering and is closely related to zero-shot learning as a capability. When zero-shot results are inconsistent, practitioners often switch to few-shot prompting or chain of thought, which coax the model to reason step by step. Understanding temperature helps here too: lower temperatures make zero-shot outputs more deterministic and reproducible.

You will encounter zero-shot prompting every time you type a plain instruction into ChatGPT, Claude, or a similar interface — it is the default interaction mode for most end users and the baseline against which more elaborate prompting strategies are measured.