Prompt Engineering Patterns: Zero-Shot vs Few-Shot Prompting Explained for LLM Developers

Zero-shot prompting asks the model to perform a task using only an instruction, with no examples. It relies on the model's pretrained knowledge and its instruction-following ability from RLHF training. Modern instruction-tuned LLMs handle most common tasks zero-shot, making it the simplest and most cost-effective starting point for any prompting task.

Use zero-shot when the task is common and well-defined, when the desired output format is simple, when you want to minimize prompt size and cost, or when you need a baseline before adding complexity. Zero-shot is the right starting point for any new task; only add examples or other techniques if the baseline is insufficient.

Few-shot prompting includes a small number of input-output examples in the prompt before the actual query. The examples teach the model the desired format, reasoning style, or task pattern. Few-shot is especially useful for tasks with specific output formats, ambiguous instructions, or domain-specific patterns the model would not infer from a plain instruction alone.

The optimal number is usually three to eight examples. Fewer than three may not establish a clear pattern, while more than ten brings diminishing returns and increases cost without proportional accuracy gains. For tasks with high variability, use more diverse examples; for narrow tasks, three carefully chosen examples often suffice.

In-context learning is the LLM's ability to learn a task from examples provided in the prompt, without any weight updates. The term comes from the GPT-3 paper, which showed that large models can pick up new tasks from a handful of in-context examples. In-context learning is what makes few-shot prompting possible and is the foundation of most prompt engineering techniques.

Choose examples that are diverse, representative of edge cases, accurate, and formatted exactly as you want the model to respond. Avoid all examples being from one extreme of the input distribution. If your task has multiple categories, include at least one example from each. Order matters too: the most recent example often has the strongest influence on the model.

Random examples come from your training data without selection logic. Curated examples are chosen to be representative, diverse, and high-quality. Curated examples almost always outperform random ones, especially for nuanced tasks. Some advanced systems use embeddings to dynamically select the most semantically similar examples for each query, called dynamic few-shot.

Few-shot fails when the task requires reasoning beyond pattern matching, when the examples introduce subtle biases that the model copies, when the task is too complex for the model regardless of examples, or when the output format the examples show is inherently ambiguous. For complex reasoning tasks, chain-of-thought prompting usually works better than pure few-shot. ---