Few-Shot Prompting: Teaching Models by Example

The Core Idea

Few-shot prompting is a technique where you include one or more worked examples in your prompt before presenting the actual task. The model reads the examples, infers the pattern, and applies it to the new input. It is one of the most reliable ways to improve output quality without modifying the model itself.

The name comes from machine learning terminology: "zero-shot" means no examples, "one-shot" means one example, and "few-shot" means a small number (typically 2–8). In practice, even a single well-chosen example can substantially improve output consistency.

Why It Works

LLMs are trained to predict the next token given everything that came before it. When you include examples in the prompt, you are effectively showing the model what the "correct" next token sequence looks like for inputs similar to the one you are about to provide. The model uses those examples as a template.

This is particularly powerful for tasks where the desired output format is unusual, where the task requires domain-specific conventions, or where the model's default behavior produces output that is technically correct but not quite what you need.

Anatomy of a Few-Shot Prompt

A few-shot prompt has three parts:

1. Task instruction — A brief description of what the model should do.
2. Examples — One or more input/output pairs that demonstrate the desired behavior.
3. Query — The actual input you want the model to process.

Here is a minimal example for sentiment classification:

Classify the sentiment of each customer review as POSITIVE, NEGATIVE, or NEUTRAL.

Review: "The battery lasts all day and the sound quality is excellent."
Sentiment: POSITIVE

Review: "Arrived damaged and customer service was unhelpful."
Sentiment: NEGATIVE

Review: "It works as described. Nothing special."
Sentiment: NEUTRAL

Review: "{{review_text}}"
Sentiment:

Review: "The battery lasts all day and the sound quality is excellent."
Sentiment: POSITIVE

Review: "Arrived damaged and customer service was unhelpful."
Sentiment: NEGATIVE

Review: "It works as described. Nothing special."
Sentiment: NEUTRAL

Review: "{{review_text}}"
Sentiment:

Notice that the examples establish the exact output format (a single capitalized label), the vocabulary (POSITIVE/NEGATIVE/NEUTRAL), and the level of granularity (no explanations, just the label). The model will follow all of these conventions when processing the query.

Choosing Good Examples

The quality of your examples matters more than the quantity. A few principles:

Cover the output space. If your task has multiple possible outputs, include at least one example of each. This prevents the model from defaulting to the most common class.

Use representative inputs. Examples that are too simple or too similar to each other may not generalize well. Choose examples that span the range of inputs you expect to encounter.

Match the difficulty level. If your real inputs are complex, use complex examples. Simple examples may lead the model to underestimate the depth of analysis required.

Keep examples consistent. All examples should follow exactly the same format. Any inconsistency in the examples will introduce inconsistency in the outputs.

Few-Shot Prompting for Batch Processing

Few-shot prompting pairs exceptionally well with batch processing. The pattern is:

1. Write a few-shot prompt with 2–4 examples that demonstrate the desired output format.
2. Use {{variable}} as the placeholder for the actual input in the query section.
3. Run the prompt against your full dataset.

Because the examples are part of the prompt template, they apply consistently to every row in your batch. You get the benefit of few-shot guidance across hundreds of inputs without any additional overhead.

Zero-Shot vs. Few-Shot: When to Use Each

Zero-shot prompting works well for straightforward tasks where the model's default behavior is already close to what you want — summarization, translation, and basic question answering.

Few-shot prompting is worth the extra token cost when the task requires a specific output format, domain-specific conventions, or a level of precision that zero-shot prompting does not reliably achieve. The examples act as a specification that the model can follow precisely.