Few-Shot Learning
Learning from examples
The Problem: Sometimes the AI doesn't understand exactly what format or style you want. How can you show it precisely what you need without lengthy explanations?
The Solution: Show, Don't Tell
Few-shot prompting means giving the model a few worked examples of the task inside the prompt, right before asking it to handle a new input. Instead of describing the rules in words, you show the model two to five input→output pairs and let it infer the pattern. It's like onboarding a new employee with sample work instead of a 10-page manual: people, and language models, often learn faster from demonstrations than from abstract instructions. Unlike Zero-Shot (no examples at all), few-shot teaches by demonstration.
The mechanism behind this is called in-context learning: the model is not retrained or updated, it simply conditions its next prediction on everything already in the prompt, including your examples. Because the examples sit in the context window, they steer the model toward the right format, tone, and decision boundary for a single request and then vanish, they have no lasting effect on the model's weights. This makes few-shot ideal when you need a very specific output shape (a JSON schema, a fixed label set, a house writing style) that is hard to specify precisely in prose. It also composes well with other techniques: for multi-step reasoning, pair it with Chain-of-Thought so each example also shows the reasoning, not just the final answer.
Tradeoffs and a worked example
The main costs are tokens and bias. Every example you add consumes context-window space and increases latency and price, and models are sensitive to example order and class balance, listing four positive reviews then one negative can skew predictions toward "positive". The fix is curation, not volume: 3-5 diverse, correctly labeled examples that cover the tricky edge cases usually beat 20 random ones. Concretely, to classify support tickets you might show: "My card was charged twice" → Billing, "The app crashes on login" → Bug, and "Can you add dark mode?" → Feature request. Faced with a new ticket like "I was billed after I cancelled", the model now reliably outputs Billing in the exact one-word format your examples established, no lengthy instructions required.
Think of it like training a new employee:
- 1. Example 1: "When a customer says 'I'm angry', we respond: 'I understand your frustration...'"
- 2. Example 2: "When they say 'This is broken', we respond: 'I'm sorry to hear that...'"
- 3. Now you try: "Customer says 'I want a refund'..."
- 4. Employee learns: The pattern of empathetic, helpful responses
Where Is This Used?
- Custom Formatting: Specific output formats, styles, or structures
- Domain-Specific Tasks: When you need industry-specific language
- Classification: Teaching categories through examples
- Data Transformation: Converting data from one format to another
Fun Fact: Research shows that just 3-5 well-chosen examples often work better than 20+ random ones. Quality over quantity! The examples should cover different cases and edge scenarios.
Try It Yourself!
Use the interactive example below to see how adding examples changes the AI's output. Try adding more examples and see how the pattern recognition improves.
📚 Few-shot — a technique where we show the model several examples BEFORE the task. More examples = better format understanding!
This expresses mixed feelings with slight disappointment...
| Mode | Examples | Quality | When to use |
|---|---|---|---|
| Zero-shot | 0 | ⭐⭐ | Simple tasks |
| One-shot | 1 | ⭐⭐⭐ | Show format |
| Few-shot | 2-5 | ⭐⭐⭐⭐⭐ | Complex/unusual tasks |
Ready-to-use templates for copying. Replace {review} with your text.
Zero-shotNo examples
Classify the review as positive or negative:
"{review}"Suitable for simple tasks where the response format is clear from the instruction.
One-shotSingle example
Classify the review as positive or negative.
Example:
Review: "Great product, highly recommend!"
Answer: positive
Now classify:
Review: "{review}"
Answer:Shows the exact response format. Sufficient for tasks with a clear pattern.
Few-shot (5)Multiple examples
Classify the review as positive or negative.
Examples:
Review: "Great product!"
Answer: positive
Review: "Terrible quality, waste of money"
Answer: negative
Review: "Fast delivery, everything works"
Answer: positive
Review: "Don't recommend, disappointed with purchase"
Answer: negative
Review: "Price/quality ratio is excellent"
Answer: positive
Now classify:
Review: "{review}"
Answer:Ideal for complex tasks. Diverse examples cover more edge cases.
- Start with zero-shot, add examples only if results are imprecise
- 3-5 examples are usually enough, more = more tokens without guaranteed better quality
- Examples should be diverse and cover edge cases
- Use consistent format for all examples (Input/Output, Question/Answer, etc.)
Few-shot examples work like "calibration" — they show the model the EXACT format you expect. 3-5 diverse examples are usually enough. More examples = more tokens, but not always better quality.
Frequently asked questions
How many examples should I use in few-shot prompting?
Typically 2-5 examples work best. Too few examples may not establish the pattern, while too many waste tokens and can confuse the model. Start with 3 diverse examples and adjust based on output quality.
When should I use few-shot instead of zero-shot prompting?
Use few-shot when: the output format is unusual or specific, the task requires domain-specific knowledge, zero-shot results are inconsistent, or you need the model to match a particular style or tone.
How do I choose good examples for few-shot prompts?
Pick diverse examples that cover edge cases, use consistent formatting across all examples, include both simple and complex cases, and order them from simple to complex. Avoid redundant examples that show the same pattern.
Does few-shot prompting work with all LLMs?
Yes, few-shot prompting works with all modern LLMs including GPT-4, Claude, Gemini, and open-source models. Larger models generally learn better from fewer examples, while smaller models may need more examples.
Try it yourself
Interactive demo of this technique
Classify this review as positive, negative, or neutral: "The food was okay, but the service leaves much to be desired."
This is a mixed review. The customer notes both positive and negative aspects.
negative
Few-shot examples showed the model the exact answer format (one word) and established a classification standard. Without examples, the model answers "in its own way".
Create a free account to solve challenges
4 AI-verified challenges for this lesson
This lesson is part of a structured LLM course.
My Learning Path