Few-Shot Prompting

PreviousNext

Guide the model's output format and behavior by providing example input-output pairs in the prompt.

Few-shot prompting is the technique of including example input-output pairs in your prompt to guide the model's behavior. Instead of describing what you want, you show it.

Why Few-Shot?

Zero-shot (just instructions) works for simple tasks. But when you need:

  • Consistent output format — JSON schemas, specific markdown structures
  • Domain-specific tone — matching your brand voice, technical writing style
  • Complex transformations — input → output mappings that are hard to describe

Few-shot prompting gives the model concrete examples to pattern-match against.

Basic Pattern

import { generateText } from "ai"
import { anthropic } from "@ai-sdk/anthropic"
 
const { text } = await generateText({
  model: anthropic("claude-sonnet-4-20250514"),
  system: `You extract structured data from product descriptions.
 
Here are examples:
 
Input: "Nike Air Max 90, men's running shoe, $129.99, available in black and white"
Output: {"name": "Nike Air Max 90", "category": "running", "price": 129.99, "colors": ["black", "white"]}
 
Input: "Adidas Ultraboost 22, women's training shoe, $189.00, comes in pink"
Output: {"name": "Adidas Ultraboost 22", "category": "training", "price": 189.00, "colors": ["pink"]}`,
  prompt: `Input: "New Balance 574, unisex lifestyle shoe, $79.99, grey and navy"`,
})

How Many Examples?

  • 2-3 examples — enough for most formatting tasks
  • 5-8 examples — for nuanced behavior (tone, edge cases)
  • 10+ — rarely needed; if you need this many, consider fine-tuning instead

Quality matters more than quantity. Pick examples that cover the edge cases you care about.

Tips

Include edge cases. If empty fields are possible, show an example with an empty field. If the input can be ambiguous, show how to handle it.

Keep examples diverse. Don't repeat the same pattern — show the range of inputs the model will see.

Match the real distribution. If 80% of inputs are short and 20% are long, your examples should reflect that ratio.

Order matters slightly. Put the most representative example first and the trickiest edge case last (recency bias helps).

Related Pattern

The Few-Shot Editor pattern (Pro) provides a full interactive UI for building and testing few-shot prompts with live AI feedback.