Evidence-first authority page
What Is Few-Shot Learning (and Zero-Shot)?
The difference between zero-shot and few-shot prompting, and when providing examples actually improves output quality.
Direct answer
Zero-shot prompting asks a model to perform a task with no examples, relying entirely on the instruction; few-shot prompting includes a small number of example input-output pairs in the prompt to demonstrate the desired pattern — few-shot generally improves consistency for tasks with a specific format or style the model wouldn't otherwise infer correctly from instructions alone.
Last reviewed 2026-08-29
This page remains noindex while its page-level source receipts are completed and verified. It is available for review, not presented as publication-ready authority.
When examples help more than instructions
Some output formats or stylistic conventions are hard to fully specify in words but easy to demonstrate — a specific tone, an unusual output structure, or a domain-specific convention. In these cases, two or three well-chosen examples in the prompt often produce more consistent results than an equivalently long written instruction trying to describe the same pattern.
Trade-offs of adding examples
- Examples consume context tokens on every call, adding cost — weigh this against the consistency gain for your specific task.
- Poorly chosen or unrepresentative examples can bias the model toward an unwanted pattern more than they help — the examples should genuinely represent the range of expected inputs, not just one easy case.
- For simple, well-specified tasks, zero-shot with a clear instruction is often just as reliable and cheaper.
A practical rule of thumb
Start zero-shot with a clear instruction. If output is inconsistent in format or style across multiple test runs, add 2-3 representative few-shot examples specifically targeting the inconsistency you observed, rather than adding examples preemptively before confirming they're actually needed.