What is few-shot prompting?

Few-shot prompting means giving the AI a few examples of exactly what you want — an input paired with the ideal output — right inside your prompt, before you ask it to do the same thing for your actual input. It's called "few-shot" because you're showing it a handful of shots, usually somewhere between 2 and 5.

Compare that to zero-shot prompting, where you just describe the task in plain instructions and give zero examples. Zero-shot works fine for simple, common requests. But for anything with a specific structure — sorting messages into exact categories, pulling data into a particular format, matching a particular tone — showing the model a couple of real examples usually beats describing the format in words.

The model picks up the pattern from the examples themselves, which cuts down on guessing and gives you more consistent, correctly-formatted results.

What it looks like

Say you want the model to tag customer messages as positive or negative. You'd feed it something like:

• Input: "This laptop is amazing!" -> Positive
• Input: "Battery died in two hours." -> Negative
• Input: "Shipping was fine, nothing special." -> ?

Give the model those first two example pairs, then hand it the third line. It copies the pattern it just saw instead of inventing its own rules for what counts as "positive."

Few-shot prompting isn't a special mode you switch on somewhere — it's just a way of writing your prompt, and it works with any LLM: ChatGPT, Claude, Gemini, whatever you're using. If your outputs keep coming back inconsistent or in the wrong format, adding two or three examples is usually the fastest fix.