Zero-shot prompting is the simplest form of prompting where you describe a task using only natural language instructions, without providing any examples of desired input-output behavior. The model relies entirely on its pre-trained knowledge and reasoning capabilities to complete the task.
A zero-shot prompt typically consists of a task description, any constraints or formatting requirements, and the input to process. For example: "Classify the following customer message as positive, negative, or neutral. Respond with only the classification label." followed by the message text.
Zero-shot prompting works well for straightforward tasks that align with common patterns in the model's training data. Tasks like text summarization, sentiment classification, translation, and simple question answering often perform well in zero-shot mode, especially with capable models.
The main advantage of zero-shot prompting is simplicity and efficiency. Without examples consuming tokens in the prompt, you have more context window available for the actual input, which matters for tasks involving long documents. Prompts are also easier to maintain since there are no examples to keep up to date.
The tradeoff is reduced control over output format and behavior. Without examples to anchor expectations, the model may produce outputs in unexpected formats, use different classification labels than intended, or interpret ambiguous instructions differently than desired.
In practice, teams often start with zero-shot prompting during prototyping, then add few-shot examples when they identify cases where the model's default behavior doesn't match requirements. This iterative approach lets you add complexity only where it's needed.