What is Zero-shot Prompting?
Also called zero-shot.
Zero-shot prompting asks a model to perform a task from instructions alone, with no worked examples included in the prompt. It relies on abilities acquired during pre-training and instruction tuning. Modern instruction-tuned models handle a wide range of common tasks this way, which keeps prompts short and cheap, though results vary more on unusual output formats.
The approach works because instruction tuning exposed the model to a very wide variety of described tasks, so it generalizes to phrasings it never encountered during training. Summarize, classify, translate, extract, rewrite, and explain are all well covered by that experience, and adding examples for tasks in this group often yields no measurable improvement at all.
Its advantages are practical rather than theoretical: fewer tokens per request, less prompt text to maintain, and no risk that a stale example teaches outdated behavior long after the policy changed. Its weakness is precision. Without demonstrations, the model picks its own formatting, verbosity, and edge case handling, which is fine for prose and unreliable for parsed output.
The label is used loosely across the field. A prompt that names the allowed categories, or sketches a required output shape, is sometimes still called zero-shot because it contains no complete input and output pairs. Strictly, zero-shot means no demonstrations of any kind, and some writers count any illustrative fragment as a shot, so definitions genuinely differ.
The practical rule is to start zero-shot, measure the results, and add examples only where measurement shows a real gap. That ordering keeps prompts small and makes it obvious which examples are earning their token cost. Instructions to reason step by step can also be applied zero-shot, with no worked examples attached.
Key points
- Instructions only, with no worked examples in the prompt.
- Works well for common tasks after instruction tuning.
- Shorter and cheaper, but output formatting is less predictable.
- The term is used loosely; definitions of a shot vary.
In practice
Send a model the instruction 'Translate the following into Spanish, keeping product names in English' followed by a paragraph. No examples are provided, and the translation comes back correct with the product names intact. The instruction alone was sufficient, because translating under a stated constraint is a task type the model encountered many times during training.