How can I effectively use examples in my prompts?

The Art of Using Examples in AI Prompts

Examples are powerful tools that can dramatically improve the quality, consistency, and accuracy of AI responses. When strategically incorporated into your prompts, examples act as clear demonstrations that guide AI models toward producing exactly the type of output you're seeking. Let's explore how to effectively use examples in your prompts to get optimal results.

Why Examples Are So Effective

Examples serve multiple crucial functions in prompt engineering:

They reduce misinterpretation by showing rather than telling what you want
They enforce consistent structure and style in the AI's response
They significantly boost performance on complex or specialized tasks
They establish patterns that the AI can recognize and follow
They clarify ambiguous instructions by providing concrete illustrations

As Anthropic notes, examples are your "secret weapon shortcut" for getting AI models to generate exactly what you need, particularly for tasks requiring structured outputs or adherence to specific formats.

Best Practices for Using Examples

Include 3-5 Diverse Examples

For optimal results, aim to include 3-5 examples in your prompt. This range provides sufficient guidance without overwhelming the model. Anthropic specifically recommends this number as a sweet spot, noting that "more examples = better performance, especially for complex tasks."

Ensure Examples Are Relevant

Your examples should closely mirror your actual use case. If you're asking for product descriptions, provide sample product descriptions rather than unrelated content types. The closer your examples match your desired output, the better the AI will understand your expectations.

Diversify Your Examples

Include examples that cover different scenarios, edge cases, and potential challenges. This diversity helps the AI understand the full range of what you're looking for and prevents it from picking up on unintended patterns. For instance, if you're asking for customer service responses, include examples handling different types of customer inquiries.

Format Examples Clearly

Clearly distinguish examples from instructions by using formatting techniques like:

Wrapping examples in tags (e.g., <example>...</example>)
Using markdown formatting (bold, italics, code blocks)
Separating examples with clear delimiters (###, ---, etc.)

This structural clarity helps the AI understand where examples end and new instructions begin.

Show Input-Output Pairs

When possible, structure your examples as input-output pairs to demonstrate the transformation you want the AI to perform:

Input: [sample input]
Output: [desired output format]

This format is particularly effective for classification, transformation, or generation tasks.

Example Structures That Work Well

For Classification Tasks

Classify the following customer feedback into categories (UI/UX, Performance, Feature Request):

Example 1:
Input: "The app crashes whenever I try to upload photos."
Classification: Performance

Example 2:
Input: "I love the interface, but I wish the buttons were larger."
Classification: UI/UX

Now classify this feedback:
"Your product would be perfect if it could integrate with Salesforce."

For Content Generation

Write a product description for the following items:

Example 1:
Product: Wireless earbuds
Description: Experience crystal-clear sound without the tangle of wires. These lightweight, ergonomic earbuds deliver 8 hours of playback on a single charge, with the pocket-sized charging case providing an additional 24 hours of battery life.

Example 2:
Product: Stainless steel water bottle
Description: Stay hydrated in style with this durable, leak-proof water bottle. Made from premium-grade stainless steel, it keeps cold drinks chilled for 24 hours and hot beverages warm for 12 hours, making it perfect for any adventure.

Now write a description for:
Product: Bamboo cutting board

For Format Adherence

Create a weekly meal plan following this format:

Example:
Monday:
- Breakfast: Overnight oats with berries (Prep time: 5 min)
- Lunch: Mediterranean salad with chickpeas (Prep time: 15 min)
- Dinner: Baked salmon with roasted vegetables (Prep time: 30 min)

Please create a meal plan for a vegetarian diet.

Advanced Techniques

Multishot Prompting

This technique involves providing multiple examples (also called "shots") to help the AI learn patterns. According to Anthropic's documentation, multishot prompting is particularly effective for tasks requiring structured outputs or adherence to specific formats.

Contrast Examples

Sometimes showing what you don't want is as helpful as showing what you do want:

I need email responses that are friendly but professional.

Good example:
"Thank you for reaching out! I've reviewed your request and can help you with this matter. Please expect a solution by Friday."

Bad example:
"Got your email. Will look into it. Back to you soon."

Please write a response to this customer inquiry about a delayed shipment.

Graduated Examples

For complex tasks, consider providing examples that gradually increase in difficulty or complexity:

Translate the following English idioms into Spanish:

Simple example:
"It's raining cats and dogs" → "Está lloviendo a cántaros"

Moderate example:
"Don't put all your eggs in one basket" → "No pongas todos los huevos en la misma canasta"

Complex example:
"The pot calling the kettle black" → "El burro hablando de orejas"

Now translate: "Barking up the wrong tree"

Common Pitfalls to Avoid

Unintentional Patterns

Be careful not to include unintended patterns across your examples that the AI might pick up on. For instance, if all your examples happen to be in past tense, the AI might assume all outputs should use past tense.

Insufficient Diversity

If your examples are too similar, the AI may struggle with variations or edge cases. Ensure your examples cover different scenarios and edge cases.

Overwhelming with Too Many Examples

While examples are helpful, too many can dilute the focus of your prompt. Stick to 3-5 well-crafted examples rather than providing an exhaustive list.

Iterative Refinement

As with all prompt engineering, using examples effectively is an iterative process. If you don't get the desired results:

Analyze what went wrong
Adjust your examples to better illustrate your expectations
Try again with the refined examples
Continue refining until you achieve the desired output

Remember that the quality of your examples directly impacts the quality of the AI's response. Taking time to craft clear, diverse, and relevant examples will pay dividends in the accuracy and usefulness of the AI's output.

FAQ

How many examples should I include in my prompt?

The optimal number is typically 3-5 examples. This provides sufficient guidance without overwhelming the model or consuming too much of your token limit.

Should my examples be real or can I make them up?

Both approaches work. What matters most is that your examples accurately represent the pattern, format, or style you want the AI to follow.

How do I format examples to make them clear to the AI?

Use consistent formatting with clear delimiters like tags (<example>...</example>), markdown formatting, or separators (###) to distinguish examples from instructions.

What if I only have one good example?

Even a single well-crafted example is better than none. However, try to create additional examples that showcase different aspects or variations of what you're looking for.

Do examples work for creative writing prompts?

Absolutely! Examples can be particularly effective for creative tasks by establishing tone, style, and format expectations.

Should examples come before or after my main instructions?

Generally, it's best to provide your instructions first, followed by examples, and then the specific request. This structure helps the AI understand the context before seeing the examples.

How detailed should my examples be?

Your examples should match the level of detail you want in the AI's response. If you want comprehensive outputs, provide detailed examples.

What's the difference between few-shot and zero-shot prompting?

Zero-shot prompting means giving instructions without examples. Few-shot (or multishot) prompting involves providing examples to guide the AI's understanding.

Can examples help with specialized technical content?

Yes, examples are particularly valuable for specialized or technical content where precision and domain-specific formatting are important.

How do I know if my examples are effective?

The proof is in the output. If the AI's response closely matches the pattern, style, and quality of your examples, they're working well. If not, refine your examples and try again.