Few-Shot Learning

In plain English

A few-shot prompt pairs your instruction with two to five demonstrations — input, then the ideal output — followed by the real input you want handled. The model infers the pattern from the examples: the format, the tone, the level of detail, the edge-case handling. This is remarkably powerful because it requires no fine-tuning, no dataset, and no training run — you steer behavior with a handful of examples right in the prompt. Few-shot shines for specific output formats (JSON schemas, classification labels, a particular writing style) and for tasks where zero-shot output drifts from what you mean. The costs are longer prompts (more tokens, more expense) and the risk that badly chosen examples bias the model. When you find yourself pasting the same examples across thousands of calls, that's the point where fine-tuning may become cheaper than paying for the example tokens every time.