Give Examples Instead of Explanations

Introduction

🎯 Learning goals

• Understand the principle of “few-shot prompting”
• Learn how many examples is optimal
• Be able to create effective examples for custom instructions

The previous sections covered structuring your instructions and answering the right questions. Now you’ll learn one of the most powerful techniques in prompt engineering: showing instead of telling. A single well-chosen example can replace several sentences of explanation — and give dramatically better results.

You’ve probably experienced it: you carefully explain exactly what you want, but the tone is wrong, the format strange, or the level of detail not at all what you had in mind. There’s a simpler solution — and that’s what we look at now.

Why examples often beat explanations

The secret that the big AI companies use is simple: instead of trying to explain everything in words, show the model examples of exactly what you want the result to look like.

“Examples are your secret weapon shortcut for getting Claude to generate exactly what you need.” — Anthropic, Best Practices for Prompt Engineering

This is called few-shot prompting (or multishot prompting) — and it’s one of the most powerful techniques you can use. Zero-shot means you give no examples at all, one-shot means one example, and few-shot means two or more.

The fact that examples work better than explanations isn’t a coincidence — there are four concrete reasons why the technique is so powerful.

Why does it work so well?

1. Shows instead of tells

Humans (and AI models) learn better from concrete examples than from abstract descriptions. A single good example can clarify things that would take several sentences to explain.

2. Clarifies subtle nuances

There are expectations that are difficult to express in words — such as exactly what tone you want, how formal the text should be, or how detailed an explanation should be. Examples show this directly, without interpretation.

3. Gives consistency

When the model has seen examples of the “correct” format, it follows that pattern much more consistently than if you just described it. Consistency is critical when you’re building an assistant that will be used in practice.

4. Reduces generic AI responses

According to tests from Anthropic, the right examples can eliminate up to 80% of generic, bland AI responses — especially if you combine positive and negative examples (more on that later in this section).

The difference between using and not using examples is easiest to understand with a concrete case. Here’s a direct comparison.

From zero-shot to few-shot: A practical comparison

❌ Zero-shot (without examples)

## TASK
Categorize customer emails into the correct category.

## CATEGORIES
- Technical support
- Billing questions
- Product information
- Complaints
- Other

Result: Works okay for clear-cut cases, but gets uncertain in borderline cases. What’s the difference between “Technical support” and “Complaints” if the customer is both complaining AND needs help?

✅ Few-shot (with examples)

## TASK
Categorize customer emails into the correct category.

## CATEGORIES
- Technical support
- Billing questions
- Product information
- Complaints
- Other

## EXAMPLES

Email: "My app crashes every time I try to log in. Help!"
Category: Technical support

Email: "Why have I been charged an extra $15?"
Category: Billing questions

Email: "Does product X support Bluetooth 5.0?"
Category: Product information

Email: "I'm extremely dissatisfied with your customer service and want compensation!"
Category: Complaints

Email: "I've changed my email address, where do I update it?"
Category: Other

Result: The model now has clear examples of what each category looks like and makes much better judgments — especially in borderline cases where a question could belong to more than one category.

One of the most common questions about few-shot prompting is: how many examples do I actually need? The answer is more nuanced than you might think.

How many examples do you need?

Anthropic’s recommendation: Start with one example

“Start with one example (one-shot). Only add more examples (few-shot) if the output still doesn’t match your needs.”

Practical approach

• Start with 0 examples (zero-shot) — test if the model handles the task anyway
• Add 1 example (one-shot) — this is often enough
• Increase to 2–3 examples if results still vary
• Use 5–10 examples only for complex or niche tasks

The principle of diminishing returns

Research shows that more examples aren’t always better. After a certain number, you get marginal improvement but use more tokens.

Rule of thumb

• 1 example → Simple classification tasks, clear formats
• 2–3 examples → Medium complexity tasks, nuanced tone
• 5+ examples → Complex domains, highly specific requirements

Always test! The optimal number varies depending on the task and model.

Knowing that you need examples is one thing — creating good examples is another. Here’s what actually distinguishes an effective example from an ineffective one.

Creating effective examples

1. Representative of real use cases

Your examples should reflect the variation you expect in practice — not constructed textbook cases.

❌ Bad example (too simple and constructed)

Question: "What is Python?"
Answer: "Python is a programming language."

✅ Good example (realistic and with the right level of detail)

Question: "I just started learning Python and don't understand the
difference between lists and tuples. Can you explain?"

Answer: "Great question! Both lists and tuples store multiple values,
but the big difference is that lists can be changed (mutable)
while tuples are fixed (immutable).

Think of it this way: A list is like a to-do list — you can add,
remove, or change items. A tuple is like a birthday — once it's
set you can't change the date.

  # List (can be changed)
  fruit = ['apple', 'banana']
  fruit.append('orange')  # OK!

  # Tuple (can't be changed)
  date = (2024, 3, 15)
  date[0] = 2025  # ERROR! Raises an exception

Use lists when data can change, tuples when data should be constant."

The second example is better because it shows how a real user asks the question, demonstrates the desired tone (educational, with analogies), and shows exactly the level of detail and structure you want.

2. Covers different scenarios

If you use multiple examples, make sure they show variation in tone and situation — not just easy, clear-cut cases.

## EXAMPLES

Scenario 1: Simple technical question
Customer: "How do I reset my password?"
Assistant: "Click 'Forgot password' on the login page.
You'll receive an email with a reset link within 5 minutes.
Also check your spam folder! Need more help? 😊"

Scenario 2: Frustrated customer
Customer: "This is the third time I've contacted you about the same problem!"
Assistant: "I completely understand your frustration, and I apologize
that we haven't resolved this earlier. I'll prioritize your issue
right now. Can you tell me what error number you received so I can
escalate directly to our technical team?"

Scenario 3: Question outside knowledge
Customer: "Do you have plans to release a Mac version?"
Assistant: "Unfortunately I don't have information about future
product releases. I recommend following our blog at company.com/news.
Is there anything else I can help you with today?"

The examples cover different emotional states and show how the tone should adapt without compromising professionalism.

3. Clear structure that’s easy to follow

Use consistent formatting in your examples so the model clearly sees the pattern.

## EXAMPLES

Input: [user's actual input]
Output: [exact format you want]

Input: [next example]
Output: [next response]

4. Negative examples

A powerful technique is to show both what you WANT and what you DON’T want. According to Anthropic’s tests, this can eliminate up to 80% of generic AI responses.

## EXAMPLES OF GOOD RESPONSES

Question: "What is machine learning?"

✅ GOOD RESPONSE:
"Machine learning is when computers learn patterns from data
instead of following hardcoded rules. Think of it this way:
instead of programming 'if email contains the word Viagra → spam'
you train the model on thousands of examples.
It learns on its own what patterns are typical for spam.

Example: Netflix recommendations use machine learning
to learn what you like based on what you've watched before."

❌ BAD RESPONSE (too technical):
"Machine learning is a subset of artificial intelligence
that uses statistical techniques to give computer systems
the ability to 'learn' from data without being explicitly programmed."
Why bad: Technical jargon without explanation, no
analogy or concrete example.

❌ BAD RESPONSE (too vague):
"Machine learning is when computers get smarter."
Why bad: Too superficial, doesn't give the user
real understanding.

Key takeaways

Few-shot prompting is one of the most powerful techniques in prompt engineering — and easier to get started with than you think.

• Examples beat explanations — they show the model exactly what you want, including tone, format, and level of detail that are hard to describe in words.
• Start with 1 example — only add more if the output still doesn’t match your expectations.
• 2–3 examples are usually enough — more examples give diminishing returns and increase cost without proportional improvement.
• Make examples realistic — reflect real usage and variation, not constructed textbook cases.
• Show negative examples — demonstrating what you don’t want eliminates up to 80% of generic responses.
• Consistent formatting — use the same structure (Input/Output) across all examples so the model clearly sees the pattern.

Test your knowledge

5 questions · 100% correct to pass · Review your answers when done

Start quiz →