What is the difference between zero-shot, one-shot, and few-shot?

Zero-shot gives the model only an instruction and no examples. One-shot adds a single example of the task done correctly. Few-shot includes several examples — typically two to eight — before the real input. More examples teach the model the exact format and edge cases you want, at the cost of a longer, more expensive prompt.

When is zero-shot enough?

Zero-shot works well for common tasks the model already understands — summarising, translating, answering general questions, or simple classification with obvious labels. If the model reliably produces the right format and quality from the instruction alone, adding examples just wastes tokens. Start zero-shot and only add examples when you see failures.

How many examples should a few-shot prompt include?

Usually two to eight. The right number is the smallest set that captures the format and the tricky edge cases. Beyond that, returns diminish and you risk overfitting the model to the surface form of your examples. Make sure the examples are diverse and correct — a single wrong example can teach the model the wrong pattern.

Is few-shot the same as fine-tuning?

No. Few-shot examples live inside the prompt and influence only that one request; the model's weights never change. Fine-tuning permanently updates the weights by training on many examples. Few-shot is instant and free to change but adds tokens to every call; fine-tuning has upfront cost but then needs no examples in the prompt.

Zero-Shot, One-Shot, and Few-Shot Learning Explained

This page explains the three main ways to teach a language model a task inside a single prompt — with no examples, with one example, or with several — and gives you an interactive picker to choose the right one for your situation.

What the three approaches mean

Zero-shot prompting gives the model an instruction and nothing else: “Classify this review as positive or negative.” It relies entirely on what the model already learned in training. One-shot adds a single worked example before your real input, showing the model exactly what a correct answer looks like. Few-shot includes several examples — commonly two to eight — which is enough to pin down the output format and demonstrate how to handle edge cases. All three are forms of in-context learning: the model adapts from the prompt without any change to its weights.

When to use each

Reach for zero-shot when the task is common and the model clearly understands it already — summarising, translating, or simple classification with obvious labels. Move to one-shot when the model gets the idea but not the exact format you want; a single example often fixes formatting without much added cost. Use few-shot when the task is unusual, the output format is strict, or there are tricky edge cases the model keeps mishandling. The examples act as a specification the model imitates. The interactive picker below maps your task type and format strictness to a suggested starting point.

Tips that make examples work

The quality of few-shot examples matters more than the quantity. Keep them diverse so the model does not latch onto one narrow pattern, keep them correct because a single wrong example teaches the wrong behaviour, and keep the format identical to what you want back. Order can matter too — models sometimes weight the most recent example most heavily. If a few-shot prompt grows long and expensive and the task is stable, that is the signal it may be worth fine-tuning instead.