Definition
Zero-shot prompting asks a model to perform a task without supplying worked input-output examples in the prompt. The prompt can still contain instructions, background information, constraints, or an output schema. “Zero-shot” only describes the number of demonstrations: zero.
The shot count covers the complete model-visible input. Examples in a system message or retained conversation history still count as demonstrations.
The model may have encountered similar tasks during training, and the request can still provide plenty of context.
Simple example
A support application needs to classify a message as billing, technical, or account. It sends the allowed labels, a short definition of each one, and the new message. It does not show any previously classified messages.
The model must interpret the instruction and apply what it learned during training. If the prompt included reviewed input-output pairs before the new message, it would be one-shot or few-shot prompting instead.
Why it matters
Zero-shot prompting is often a useful baseline. Without demonstrations, the prompt generally uses fewer tokens than an otherwise equivalent few-shot prompt. This saves context-window capacity and may reduce input cost. It also avoids introducing accidental patterns from a small or unrepresentative example set.
Without demonstrations, the output depends on the model’s learned capabilities and the instructions and other task information in the prompt. Labels that sound similar, unusual output formats, or domain-specific rules may produce inconsistent results. Test the prompt with representative inputs before deciding that examples are unnecessary.
One important nuance
Zero-shot prompting and in-context learning are related, but their relationship depends on how in-context learning is defined. Zero-shot tells you that the prompt contains no demonstrations. In-context learning often refers to adapting model behavior using demonstrations supplied in the input, although broader definitions also include zero-shot instruction following.
Under that broader definition, a zero-shot request can also be considered an instance of in-context learning. Adding examples changes the prompting setup while the model’s weights and training history remain unchanged.