Few-shot prompting

No, they are very different. Few-shot examples live in the prompt and teach the model on-the-fly without changing its weights. Fine-tuning permanently updates the model weights with training data. Few-shot is faster, cheaper, and requires no infrastructure since you just add examples to the prompt.

There are three levels: Zero-shot: No examples. "Classify this email as spam or not spam." One-shot: One example. "Here is an example: 'Win a free iPhone!' -> Spam. Now classify: 'Meeting at 3pm tomorrow'" Few-shot: 2-5 examples that establish a clear pattern. The model picks up on format, style, and decision logic from the examples. More examples means more reliable pattern matching, but also costs more tokens.

For simple tasks (classification, formatting), 2-3 examples are usually enough. For complex patterns (style transfer, nuanced decisions), 3-5 examples work better. Beyond 5, you get diminishing returns and higher token costs. The key is covering edge cases, so pick diverse examples that show the boundaries of each category.