Few-Shot Prompting

Few-shot prompting provides an AI model with 2-5 examples of input/output pairs before your actual request. The model learns the desired pattern — tone, format, classification logic, extraction structure — from the examples rather than from a verbal description. It's one of the highest-impact prompting techniques for tasks where you can demonstrate the pattern more easily than you can describe it.

Three examples is the sweet spot for most tasks. One example (one-shot) often works for simple patterns but can be brittle. Two to three examples covers the pattern reliably. More than five rarely improves results and increases token usage. The exception is classification tasks with many categories — provide at least one example per category. For chain-of-thought few-shot prompting, even one worked example showing the reasoning process is highly effective.

Zero-shot prompting describes what you want in words, without examples. Few-shot prompting demonstrates what you want with 2-5 examples. Zero-shot is faster to write and works well for straightforward tasks. Few-shot is more work upfront but produces more consistent and accurate results for tasks involving specific formats, styles, or classification logic. Use zero-shot first; switch to few-shot when zero-shot produces inconsistent results.

For most business use cases, few-shot prompting achieves 70-90% of fine-tuning performance at near-zero cost. Fine-tuning bakes the pattern into the model weights permanently and is faster at inference, but requires hundreds or thousands of examples and ongoing maintenance. Few-shot prompting is the practical choice for most teams — start here. Only invest in fine-tuning if you have a clear quality gap, a high-volume use case, and quality training data.

Highest-benefit tasks: format conversion (converting one data structure to another), tone matching (rewriting in a specific brand voice), classification (tagging content into predefined categories), structured data extraction (pulling fields from unstructured text), and code translation (converting code between languages or patterns). Tasks that benefit less: open-ended generation, research, and tasks where the desired output is fully described by a verbal instruction.