Zero-Shot vs Few-Shot Prompting, Explained (2026)

Zero-shot means you describe the task and ask for the result without showing any examples. The model relies entirely on its instructions and what it learned in pretraining. Modern models are strong enough that zero-shot works for most common, well-understood tasks.

A zero-shot prompt looks like this:

``` Classify the sentiment of the review as positive, negative, or neutral. Review: "The battery lasts forever but the screen scratches easily." Sentiment: ```

It is short, fast, and cheap. It also leaves the model free to interpret edge cases its own way — which is exactly where it can drift from what you wanted.

Few-shot means you include a handful of examples — each showing an input and the exact output you want — before the real input. The model infers the pattern and continues it. One-shot is the special case with exactly one example.

The same sentiment task as few-shot, pinning down a specific output format:

``` Classify the sentiment. Respond with only one word. Review: "Shipping was fast and the quality is great." Sentiment: positive Review: "It broke after two days. Total waste of money." Sentiment: negative Review: "It works as described, nothing special." Sentiment: neutral Review: "The battery lasts forever but the screen scratches easily." Sentiment: ```

The examples do the teaching: they fix the label set, the one-word format, and how to handle mixed sentiment. This is in-context learning — the model adapts to the task from the examples without any change to its weights.

Start zero-shot. It is cheaper (fewer input tokens), faster to write, and usually enough for standard tasks. Only add examples when zero-shot output misses on format, edge cases, or consistency.

Few-shot wins when the task has a precise output shape, an unusual or domain-specific convention, subtle classification boundaries, or a tone the model keeps getting wrong. Examples communicate 'do it exactly like this' far more reliably than prose instructions. The cost is more input tokens per call and the effort of curating good, representative examples — and bad or skewed examples can actively hurt, since the model will copy their mistakes.

For multi-step reasoning, a third option often beats both: chain-of-thought prompting, which asks the model to reason through steps before answering. It was introduced by Wei et al. 2022 (arXiv:2201.11903) and pairs well with either zero-shot or few-shot.

Use zero-shot when: The task is common and well-understood, you want minimal tokens and fast iteration, and default output quality is already acceptable.
Use few-shot when: You need a precise format, the task has subtle edge cases or domain conventions, or zero-shot keeps drifting on tone or structure.

More is not always better. Two to five well-chosen, diverse examples usually capture the pattern; beyond that you mostly add token cost and risk overfitting the model to your specific phrasings. The examples should cover the variety you expect, including a tricky edge case, and should be flawless — the model imitates everything it sees, errors included.

Watch ordering and balance too: in classification, skewing examples toward one label can bias the output. Keep the format identical across examples and identical to what you ask for in the final input, so the model has one unambiguous pattern to follow.