What is few-shot prompting?

Few-shot prompting means including a few worked examples in a prompt so that the AI copies their pattern. Each example shows an input and the answer wanted, such as an email and the category it belongs to. AI models follow examples more reliably than written descriptions. A prompt with instructions but no examples is called zero-shot.

Why it matters for a business

Examples are the cheapest way to get consistent output on a repeated task. Describing a house tone in adjectives leaves a wide range of answers; pasting 2 past pieces that got it right leaves a narrow one. The examples travel with every request, so they add a little to the cost of each run, and they work best when they include an awkward case as well as the easy ones. Cases used as examples in the prompt cannot also be used to test it, because a case the model was shown the answer to tests nothing.

Example

Let’s say a team wants inbound emails tagged as billing, delivery or complaint. With 3 tagged examples above the new email, the model returns one word in the same form every time, instead of a sentence explaining its choice.

Go deeper

How to write a good AI prompt, and why prompts matter

Related terms

• Prompt engineering
• System prompt
• Fine-tuning
• Golden dataset

All terms, A to Z