The analogy

Think of a new colleague on their first day. You tell them "fill out these forms" and let them go: if the form is obvious, that's fine, that's zero-shot. But if the form is strange and has a thousand boxes, you put two or three already-filled-out forms in front of them as samples. They look, they understand the style, the alignments, the abbreviations, and they copy the model. Showing them a finished example is worth more than half an hour of explanations. Few-shot is that little stack of examples you set down beside them.

How it really works

The AI derives the "how it should be done" from the examples you put inside the request. Each example tells it more than an abstract rule: it fixes the format, the tone, the length, and above all the edge cases, which you'd struggle to describe in words. The term "shot" is precisely the example: zero-shot none, few-shot a few. It's not learning permanently: it uses the examples only there, in that conversation, as a model to imitate.

What you can do in practice

  • Start zero-shot: just ask. Often it's already enough, and you save time.
  • If the result is off target in format or style, add two or three before-and-after examples. The operative syntax:
Turn the titles into a clear, catchy but sober version. Examples:
- "Third Quarter Budget Meeting" -> "The Quarterly Budget: the 3 numbers that count"
- "Vacation Policy Update" -> "Vacation: what changes from January"
Now turn this one: "New Warehouse Management System"
  • Keep the examples consistent with each other: if they contradict, the AI gets confused.
  • For classifying, formatting or imitating a style, few-shot is almost always the winning move.

A common misconception

People think the more examples you give, the better. Beyond a handful, though, the examples stop helping: they fill up the conversation and sometimes push the AI to copy them to the letter instead of grasping the criterion and generalizing. Two to four well-chosen examples that cover the variety of cases are worth more than ten identical ones. The quality of the examples counts more than the number.

Frequently asked questions

How many examples are needed?

Generally two to four. They serve to show the variety of cases, not to fill space. If you put in many and all alike, you add weight without adding information.

Is it better to give examples or to explain in words what I want?

It depends. For format and style the examples win, because showing is more effective than describing. For logical rules ("exclude cases under 18") the explanation is clearer. Often the combination of the two is the best thing.

Does it work on all AIs?

Yes, it's a general principle of how these models are steered, not a feature of a specific product. The product changes, not the idea: showing examples of the desired result helps everywhere.

Without examples doesn't the AI understand?

On common tasks it understands perfectly well on its own: asking it for an apology email or a summary doesn't require examples. Examples aren't there to make it "understand," they're there when the result has to come out done your way, with a format or a style that's awkward to explain in words. It's a lever of precision, not a crutch for misunderstanding.