Zero-Shot and Few-Shot

Zero-shot means giving the model an instruction and nothing else. Few-shot means including a small number of worked examples in the prompt so the model can infer the pattern you want.

The striking part is that few-shot examples teach without training. Nothing is learned and no weights change — the examples sit in the context window and shape the answer for that request only. This is often called in-context learning, and it was one of the genuinely surprising properties to emerge from scaling.

The practical rule of thumb has shifted as models improved. Capable models now handle most straightforward instructions zero-shot, so examples earn their place mainly where the desired output has a specific shape — a particular format, a house style, an unusual classification scheme.

This is the cheapest quality lever available and the one most people underuse. When output is nearly right but formatted wrong or pitched wrong, two or three examples usually fix it in seconds, where the instinct is often to reach for fine-tuning — a far larger commitment that teaches the same kind of thing.

Start zero-shot. If the shape of the answer is wrong, add two or three examples that demonstrate the shape rather than describing it. Make them representative rather than ideal — examples that are cleaner than your real inputs teach the model to expect clean inputs.