Zero-shot

Overview

Zero-shot prompting provides only an instruction, relying on the model's pre-trained knowledge to infer the expected output format. GPT-4 and Claude 3+ class models perform well zero-shot on a wide range of tasks. The practical workflow is: try zero-shot first, then add few-shot examples or CoT cues if quality is insufficient.

Advantages and limits

Zero-shot prompts are compact and conserve context-window space. For highly domain-specific formats or narrow technical tasks, however, the absence of examples can lead to inconsistent output structure.