Few-Shot Prompting

Few-shot prompting is a technique where you include several examples of the desired input-output pattern in your prompt before presenting the actual task. By showing the AI what good output looks like, you dramatically improve the consistency and quality of its responses — without any model training or fine-tuning.

How few-shot prompting works

Instead of just describing what you want, you show the AI examples:

Without few-shot (zero-shot): "Classify this customer feedback as positive, negative, or neutral: 'The delivery was fast but the packaging was damaged.'"

With few-shot (3 examples): "Classify customer feedback as positive, negative, or neutral.

Example 1: 'Absolutely love this product, will buy again!' → Positive Example 2: 'Terrible quality, broke after one day.' → Negative Example 3: 'It arrived on time and works as expected.' → Neutral

Now classify: 'The delivery was fast but the packaging was damaged.'"

The few-shot version produces more reliable and consistent results because the AI has concrete examples of how you want the task performed.

Why few-shot prompting works

LLMs are exceptional pattern matchers. When you provide examples, the model identifies the pattern (input format → output format, classification rules, writing style) and applies it to the new input. This works because:

• Examples remove ambiguity about what you want
• The model adapts its behaviour to match the demonstrated pattern
• Consistency improves because every response follows the same template

How many shots?

The "few" in few-shot typically means 2-5 examples:

• Zero-shot: No examples. Relies entirely on instructions.
• One-shot: A single example. Better than zero but limited.
• Few-shot (2-5): The sweet spot for most tasks. Enough examples to establish a pattern without consuming too much of the context window.
• Many-shot (10+): Sometimes useful for complex or nuanced tasks, but watch context window limits.

More examples is not always better. After 5-7 examples, additional examples typically provide diminishing returns and consume valuable context window space.

When to use few-shot prompting

Few-shot prompting is especially effective for:

• Classification: Teaching the AI your specific categories and criteria
• Data extraction: Showing the exact format you want information pulled into
• Content formatting: Demonstrating your preferred writing style, structure, or tone
• Translation of concepts: Showing how to convert technical language to plain English (or vice versa)
• Consistent labelling: Ensuring the AI applies the same standards across all items

Few-shot prompting best practices

• Choose representative examples: Include examples that cover the range of inputs you expect — easy cases, hard cases, and edge cases.
• Be consistent: All examples should follow exactly the same format. Inconsistent examples produce inconsistent output.
• Include edge cases: Show the AI how to handle ambiguous or unusual inputs.
• Order matters: Put simpler examples first and more complex ones later. End with the most representative example.
• Match your actual data: Use examples that look like real inputs, not artificial ones.

Few-shot vs fine-tuning

Few-shot prompting and fine-tuning both improve AI output on specific tasks, but they differ:

• Few-shot: No training required. Instant. Examples live in the prompt. Limited by context window.
• Fine-tuning: Requires training data and compute. Permanent behaviour change. Not limited by context window but less flexible.

For most business tasks, few-shot prompting should be tried first. It is faster, cheaper, and often sufficient.