What is few-shot prompting?

Few-shot prompting is a technique where you include a few examples of the task in your prompt, so the AI can see exactly what you want before doing it. "Shot" here simply means example.

There are three basic approaches:

Zero-shot: you give instructions without examples. "Classify this customer message as Complaint, Query or Praise."
One-shot: you include one example.
Few-shot: you include several examples, typically two to five.

Here is a few-shot example for classifying customer messages:

Message: "My order arrived damaged and nobody replies to my emails." Category: Complaint
Message: "Do you deliver to Gorakhpur?" Category: Query
Message: "The saree quality is amazing, thank you!" Category: Praise
Message: "When will the blue kurta be back in stock?" Category:

The AI sees the pattern and continues it, answering "Query."

Why does it work? Examples communicate things that are hard to describe in words: the exact format, tone, level of detail, style and edge cases. Instead of explaining "write short, friendly product titles with the key feature first," you show three titles you like, and the AI imitates them.

Few-shot prompting is especially useful for:

1. Classification tasks with your own categories.
2. Consistent formatting, such as converting notes into a fixed template.
3. Matching a writing style or brand voice.
4. Data extraction, such as pulling names, dates and amounts from text into a table.
5. Tasks with unusual rules that are easier to show than explain.

Tips for writing good examples:

Make examples realistic, similar to the actual inputs the AI will receive.
Show variety, covering different types of cases, including tricky ones.
Keep the format identical across examples.
Make sure examples are correct; the AI will copy mistakes too.
Balance categories, so the AI does not lean toward one answer because most examples show it.
Clearly separate examples from the actual task, using labels or tags.

There is one caution. AI can copy examples too closely, repeating their wording or length. If outputs look too similar to your examples, add more variety or say "use these as a guide to style, not as templates to copy."

Modern AI models are good at zero-shot tasks, so you do not always need examples. But when results are inconsistent or the format keeps going wrong, adding two or three good examples is often the fastest fix.