Advanced Prompt Design: Comprehensive Strategies Guide

To help you explore and apply these strategies, Google AI Studio offers an interactive prompt gallery that demonstrates many of the concepts discussed in this guide.

In the sections ahead, you'll uncover the following prompt design techniques:

• Provide clear and specific instructions
• Incorporate few-shot examples
• Add relevant contextual information
• Use prefixes to guide the model
• Let the model complete partial input

Provide Clear and Specific Instructions

Clear and specific instructions are a simple yet powerful way to tailor a model’s response to your needs. To achieve the best results, ensure your guidance is precise and unambiguous. Instructions can range from straightforward step-by-step processes to detailed frameworks that simulate a user's perspective or objective.

Step 1: Define the Task

Clearly outline the task you want the model to perform. This could range from a straightforward, step-by-step list of instructions to a more complex request, such as analyzing a user's mindset or mapping out their experience. Below is an example of a prompt instructing the model to summarize a block of text:

Prompt:

Summarize this text:

Text: A quantum computer exploits quantum mechanical phenomena to perform calculations exponentially faster than any modern traditional computer. At very tiny scales, physical matter acts as both particles and waves, and quantum computing uses specialized hardware to leverage this behavior. The operating principles of quantum devices fall beyond the scope of classical physics. When deployed at scale, quantum computers could be applied in areas such as cybersecurity—by breaking existing encryption methods and helping researchers develop new ones—or meteorology, for better weather forecasting. However, current quantum computers remain largely experimental and impractical for widespread use.

Response:

Quantum computers utilize the wave-particle duality of matter at microscopic scales to perform calculations exponentially faster than traditional computers. They rely on quantum phenomena, which defy classical physics, and promise transformative applications like cybersecurity and weather forecasting. However, today's quantum computers are still in the experimental stage and not yet ready for mainstream use. (gemini-1.5-flash)

Step 2: Specify Constraints

Define any specific constraints for the prompt or the response. For instance, you could limit the length of the response or impose specific rules for its structure. In the previous example, if you want the summary to be even shorter, you can include such a constraint in the prompt:

Prompt:

Summarize this text in two sentences:

Text: A quantum computer exploits quantum mechanical phenomena to perform calculations exponentially faster than any modern traditional computer. At very tiny scales, physical matter acts as both particles and waves, and quantum computing uses specialized hardware to leverage this behavior. The operating principles of quantum devices fall beyond the scope of classical physics. When deployed at scale, quantum computers could be applied in areas such as cybersecurity—by breaking existing encryption methods and helping researchers develop new ones—or meteorology, for better weather forecasting. However, current quantum computers remain largely experimental and impractical for widespread use.

Response:

Quantum computers leverage quantum mechanical phenomena to achieve speeds far beyond traditional computers. Despite their potential in fields like cybersecurity and weather forecasting, they are still in their infancy and not yet practical. (gemini-1.5-flash)

Step 3: Define the Response Format

You can also guide the model to provide its output in a specific format. For example, you can request the response to be structured as a table, bulleted list, elevator pitch, set of keywords, or in paragraph form. Below is an example of a system message defining the format and tone:

System Message:

All questions should be answered comprehensively and with sufficient detail, unless the user explicitly requests a concise response.

Prompt:

What is a smart way to create a business that sells DVDs in 2025?

Response:

Building a successful business around DVDs in 2025 is undoubtedly a challenge. The dominance of streaming platforms, digital downloads, and high-definition formats has greatly reduced the demand for DVDs. However, by adopting a highly focused strategy and understanding niche opportunities, it's still possible to carve out a viable business. Here's how:

I. Recognizing Challenges and Identifying Opportunities

Key Challenges:

• Streaming Services: Major platforms like Netflix, Amazon Prime Video, and Disney+ dominate the market, offering vast content libraries instantly accessible.
• Digital Downloads: Platforms such as iTunes and Google Play make acquiring digital media seamless and convenient.
• Blu-ray and 4K Ultra HD: Advanced formats with higher quality have become the go-to for home theater enthusiasts.

Potential Opportunities:

• Nostalgia Market: Focus on iconic movies, classic TV shows, and collector’s editions that appeal to enthusiasts and collectors.
• Specialized Content: Cater to niche genres or hard-to-find DVDs that aren’t easily available on digital platforms.
• Custom Packaging: Offer unique designs, special edition box sets, or personalized gifts for movie buffs.

II. Strategies for Success

1. Focus on a Niche Audience: Pinpoint and cater to audiences who still value DVDs—such as collectors, retro enthusiasts, or regions with limited access to high-speed internet.

2. Leverage Online Sales Channels: Use platforms like eBay, Amazon, or a dedicated e-commerce website to reach potential customers globally.

3. Partner with Content Creators: Collaborate with independent filmmakers or content creators to distribute their work via physical media.

4. Build a Brand Community: Engage with your audience through social media, forums, or newsletters, creating a loyal fan base around your offerings.

While the DVD market is undeniably shrinking, there’s room to thrive with creativity, strategy, and a focus on niche opportunities. Success hinges on how well you adapt to the remaining demand and deliver value to a passionate audience.

Prompt:

What is a smart way to create a lemonade business?

Response:

Launching a Successful Lemonade Business: Smart Strategies

Starting a lemonade business—whether it’s a classic roadside stand, a mobile food truck, or a larger operation—requires thoughtful planning and a clear strategy. Here’s how to get started:

I. Define Your Business Vision

• Determine the Scope: Will this be a small, seasonal setup, a permanent storefront, or a delivery-based business?
• Identify Your Target Audience: Are you serving kids, office workers, tourists, or health-conscious consumers?
• Develop a Unique Selling Proposition (USP): What makes your lemonade special? Organic ingredients, creative flavors, or a secret family recipe?

II. Plan Your Operations

• Pricing: Research competitors and balance affordability with profitability.
• Location Strategy: Choose high-traffic areas such as parks, busy streets, or community events. Ensure compliance with local permits and regulations.
• Marketing: Use social media, local advertising, and promotions to spread the word. Create an engaging, recognizable brand identity.

III. Sustainability and Growth

• Eco-Friendly Practices: Use biodegradable cups and straws to attract environmentally conscious customers.
• Expand Offerings: Include add-ons like snacks, merchandise, or seasonal beverages to increase revenue streams.
• Customer Feedback: Regularly gather feedback to improve your product and service based on customer preferences.

By focusing on a clear vision, understanding your audience, and positioning your lemonade business creatively, you can stand out in the market and build a sustainable brand.

Summary Guidelines:

• Provide detailed, clear, and actionable responses unless a concise answer is specifically requested.
• Ensure instructions are easy to follow and include specific examples for better understanding.
• When applicable, offer step-by-step guidance to implement the suggested strategies effectively.

Including examples in your prompts demonstrates what “getting it right” looks like, helping the model understand patterns and relationships. By analyzing these examples, the model applies the identified patterns when generating responses. Prompts with a few examples are called "few-shot prompts," while those without examples are referred to as "zero-shot prompts." Few-shot prompts are particularly useful for guiding the model's formatting, phrasing, scope, or general response patterns. To achieve more accurate results, use specific, varied examples to help the model focus on the desired task.

We strongly recommend incorporating few-shot examples into your prompts whenever possible. Prompts without such examples are often less effective. Clear, illustrative examples can even replace detailed instructions, as they inherently demonstrate the task to the model without the need for further clarification.

Zero-shot vs. Few-shot Prompts

Here’s an example of a zero-shot prompt where the model is asked to select the best explanation:

Prompt:

Please choose the best explanation to the question:

Question: How is snow formed?

Explanation 1: Snow is formed when water vapor in the air freezes into ice crystals in the atmosphere. These ice crystals can combine and grow into snowflakes as they fall through the atmosphere, eventually accumulating on the ground.

Explanation 2: Water vapor freezes into ice crystals, forming snow.

Answer:

Response: Explanation 1 is the best explanation. (gemini-1.5-flash)

To generate concise responses, you can refine your prompt by including examples that emphasize brevity. By clearly structuring examples with a preference for shorter explanations, you guide the model to produce responses that better align with your specific needs.

Consider the following prompt, which provides two examples demonstrating a preference for concise explanations. As shown in the responses, the examples guide the model to select the shorter explanation (Explanation2) over the longer one (Explanation1), ensuring a more succinct output.

Prompt:

Below are examples using a question, explanation, and answer format:

Example 1:

Question: Why is the sky blue?

Explanation1: The sky appears blue because of Rayleigh scattering, which causes shorter blue wavelengths of light to scatter more than longer red wavelengths, giving the sky its blue color.

Explanation2: Due to the Rayleigh scattering effect.

Answer: Explanation2

Example 2:

Question: What causes earthquakes?

Explanation1: A sudden release of energy in the Earth's crust.

Explanation2: Earthquakes occur when tectonic plates suddenly slip or break apart, releasing energy that creates seismic waves, shaking the ground and causing potential damage.

Answer: Explanation1

Your Turn:

Now, answer the following question using the example formats above:

Question: How is snow formed?

Explanation1: Snow forms when water vapor in the air freezes into ice crystals in the atmosphere. These ice crystals combine and grow into snowflakes as they fall, eventually accumulating on the ground.

Explanation2: Water vapor freezes into ice crystals, forming snow.

Answer:

Response:

Answer: Explanation2

(gemini-1.5-flash)

Find the Optimal Number of Examples

To achieve the best results, experiment with the number of examples you include in the prompt. Models like Gemini can often identify patterns with just a few examples, but finding the right balance is key. Including too many examples could cause the model to overfit its responses to those examples, while too few may not provide enough context. Adjust and test until you find the optimal number for your use case.

Focus on Patterns, Not Antipatterns

When providing examples, prioritize showing the model what to do rather than what to avoid. Examples that demonstrate correct patterns are far more effective than those that highlight incorrect ones.

⛔ Negative Example:

Don’t end haikus with a question:

Haikus are fun
A short and simple poem
Don’t you enjoy them?

✅ Positive Example:

Always end haikus with an assertion:

Haikus are fun
A short and simple poem
A joy to write

Maintain Consistency in Formatting

Ensure that all examples follow a consistent structure and formatting to avoid undesired variations in the model’s output. Few-shot examples aim to guide the model on the desired response format, so consistency is critical. Pay close attention to elements such as XML tags, whitespace, line breaks, and how examples are separated. This consistency helps the model generate responses that align closely with your expectations.

Summary

Providing example prompts and responses helps the model learn and replicate desired patterns effectively. Instead of showing examples of what to avoid, focus on illustrating the behavior you want the model to follow.

Experiment with how many examples to include in your prompt. A small number of examples may not significantly influence the model's behavior, while too many can lead to overfitting. Strive to find the right balance based on the specific model you’re working with.

Ensure examples are consistently formatted to make patterns clear. Additionally, include contextual details and instructions in the prompt rather than assuming the model has all the necessary information. This extra context helps the model better understand the constraints and specifics of your request, leading to more accurate responses.

Example:

Here’s a sample prompt for troubleshooting a router issue:

Prompt:

What should I do to fix my disconnected Wi-Fi? The light on my Google Wifi router is yellow and blinking slowly.

Response:

A yellow blinking light on your Google Wifi router usually indicates an issue with the internet connection. Follow these steps to troubleshoot and resolve the problem:

1. Restart your router and modem:
   • Unplug your modem from the power outlet and wait for 30 seconds.
   • Plug the modem back in and allow it to fully boot up.
   • Unplug your Google Wifi router and wait for 30 seconds.
   • Plug your Google Wifi router back in and wait for it to fully boot up.
2. ... (include additional troubleshooting steps as required).

gemini-1.5-flash

The response currently seems like generic troubleshooting information and lacks specificity to the router or the LED indicator lights. To improve accuracy, you can customize the response by providing the router's troubleshooting guide as contextual data in the prompt. This ensures the model has detailed, relevant information to work with.

How to Customize the Prompt:

Include the following contextual information in the prompt:

• LED Indicator Colors and Their Meanings: Provide the guide detailing what each light color and blinking pattern signifies.
• Troubleshooting Instructions: Add specific steps for resolving issues based on the LED indicators and related scenarios.

Suggested Prompt Example:

Use the text below to answer the question. Only respond with the relevant instructions provided in the text.

Question: What should I do to fix my disconnected Wi-Fi? The light on my Google Wifi router is blinking slowly in yellow.

Contextual Information:

• Color: Slowly Pulsing Yellow
  Meaning: There is a network error.
  What to do: Check that the Ethernet cable is securely connected to both your router and modem, and ensure both devices are powered on. If the issue persists, unplug and reconnect each device.
• Color: Fast Blinking Yellow
  Meaning: The router is being factory reset (reset button is held down).
  What to do: Continue holding the reset button for about 12 seconds until the light turns solid yellow. Then release the button to complete the reset process.
• Color: Solid Yellow
  Meaning: The router is in the process of factory resetting.
  What to do: Wait up to 10 minutes. Once complete, the device will begin pulsing white, indicating it’s ready for setup.
• Color: Solid Red
  Meaning: Critical failure.
  What to do: Perform a factory reset. If the light remains red, contact customer support.

Response:

Check that the Ethernet cable is securely connected to both the router and modem, and ensure both devices are powered on. You may also need to unplug and reconnect each device.

Summary of Best Practices for Prompt Design:

• Provide Context: Include detailed troubleshooting information, such as LED color meanings and corresponding steps, in your prompt.
• Set Clear Instructions: Specify exactly how the model should respond using the provided context (e.g., "Only respond with relevant steps").
• Use Prefixes: Add input and output prefixes to ensure the model understands the structure of the prompt. For example:
  • Input Prefix: Label different input segments (e.g., "Color:", "Meaning:", "What to do:") to make it easier for the model to parse the data.
  • Output Prefix: Use a prefix like "Response:" or "JSON:" to signal the expected format of the response.

In the following example, "Text:" is the input prefix and "The answer is:" is the output prefix.

Prompt:

Classify the following text into one of these categories:

• Large
• Small

Examples:

Text: Rhino

Answer: Large

Text: Mouse

Answer: Small

Text: Snail

Answer: Small

Text: Elephant

Answer:

Response:

Answer: Large

(Generated by gemini-1.5-flash)

Using Models to Complete Partial Input

Generative language models function like advanced autocompletion tools. When provided with partial content, they predict and generate the continuation of the content based on the context given. By including specific examples or contextual details, you can guide the model to produce more precise and relevant responses.

Example of Structured Prompting:

Prompt:

For the given order, return a JSON object containing fields such as cheeseburger, hamburger, fries, or drink, where the values represent the ordered quantities.

Order: A burger and a drink.

Response:

{ "hamburger": 1, "drink": 1 }

(Generated by gemini-1.5-flash)

Guidance for Better Results:

While the model followed the instructions, natural language prompts can sometimes leave room for interpretation. For instance, if a restaurant’s menu contains many items, it’s often better to exclude items not mentioned in the order to simplify the JSON response. In such cases, providing a clear example and a response prefix can help the model generate more concise and accurate outputs:

Prompt:

The valid fields for orders are: cheeseburger, hamburger, fries, and drink.

Example Order: Give me a cheeseburger and fries.

Output:

{ "cheeseburger": 1, "fries": 1 }

Example Order: I want two hamburgers, a drink, and fries.

Output:

{ "hamburger": 2, "drink": 1, "fries": 1 }

(Generated by gemini-1.5-flash)

Note: In the first example, "cheeseburger" was included in the output because it was part of the order. In the second example, "cheeseburger" was excluded since it was not requested. The model accurately formats the output to reflect only the items mentioned in the order.

Guiding the Model to Format Responses

You can use the model's completion strategy to ensure responses are formatted as needed. For example, if you'd like the model to create an essay outline, you can prompt it with specific instructions or provide a pattern to follow. Here's how:

Prompt: Create an outline for an essay about hummingbirds.

Response:

Hummingbirds: A Symphony of Speed and Color I. Introduction A. Hummingbirds are fascinating creatures known for their dazzling colors, lightning-fast flight, and unique feeding habits. B. Brief overview of the diversity of hummingbird species and their geographic distribution. C. Thesis statement: Hummingbirds are an embodiment of resilience, adaptation, and the wonders of the natural world. ...

(Generated by gemini-1.5-flash)

In this case, the prompt did not specify the format for the outline, so the model chose its own structure. However, if you need a specific format, provide an example or starting text to establish a clear pattern for the model to follow. This way, the model will complete the response in the desired format.

Prompt:

Create an outline for an essay about hummingbirds.

Response:

Hummingbirds: Jewels of the Air

I. Introduction

• Hook: Start with a vivid description or captivating anecdote about hummingbirds, highlighting their dazzling colors, lightning-fast movements, and unique feeding habits.
• Background: Introduce hummingbirds as the world's smallest birds, known for their extraordinary adaptations that allow them to thrive on a nectar-based diet.
• Thesis Statement: Present the essay's central focus. Some possibilities include:
  • The remarkable adaptations of hummingbirds equip them to flourish in diverse habitats.
  • Hummingbirds play an essential role in pollination and maintaining biodiversity.
  • The elegance and charm of hummingbirds inspire both scientific exploration and artistic admiration.

Tips for Creating Effective Prompts:

• For partial inputs, provide a clear structure or example to guide the model’s completion. For instance, including a partial answer can orient the model toward your desired format and style.
• Simplify complex tasks by breaking the prompt into smaller, digestible components, ensuring clarity and consistency.
• When dealing with multiple instructions, separate them into individual prompts. This approach allows for better organization and ensures that the model addresses each instruction effectively.

Chain Prompts

For tasks involving multiple sequential steps, break down each step into individual prompts and link them in a logical sequence. In this approach, the output of one prompt serves as the input for the next, and the process continues until the final prompt delivers the desired result.

Aggregate Responses

When handling parallel tasks on different parts of the data, use aggregation to combine the results into a final output. For instance, you can instruct the model to perform one operation on a specific section of the data and a different operation on another section, then merge the outputs for a comprehensive result.

Summary

• Break down complex instructions into individual prompts and determine which prompt to use based on the user’s input.
• For sequential tasks, structure prompts in a chain where the output of one prompt feeds into the next.
• For parallel tasks, divide the work and aggregate the responses for the final output.

Experiment with Parameter Values

Each model call includes parameters that influence how the model generates responses. Adjusting these parameters can help fine-tune the results for your specific needs. Note that the available parameters may vary depending on the model. Below are some commonly used parameters:

• Max Output Tokens: Defines the maximum number of tokens the model can generate in a response. One token is approximately four characters, so 100 tokens roughly equal 20 words. Use lower values for concise responses and higher values for longer outputs.
• Temperature: Controls the randomness of token selection during response generation. Lower values create more deterministic and focused responses, while higher values encourage creativity and diversity. A temperature of 0 ensures the most predictable results. For general tasks, start with a temperature of 0.2. If the responses feel overly generic or too short, consider increasing the temperature.
• Top-K: Limits the model’s token selection to the top K tokens with the highest probabilities. Adjust this parameter to control response variability.
• Top-P: Also known as nucleus sampling, it restricts token selection to the smallest set of tokens whose cumulative probability exceeds a specified threshold (P). Use this to fine-tune response creativity.

Experiment with these parameters to discover the optimal settings for your specific task and model. Small adjustments can significantly impact the quality and relevance of the results.

Top-K

Top-K determines how the model selects the next token in a sequence. A top-K value of 1 means the model will always pick the most probable token (known as greedy decoding), while a top-K of 3 allows the model to choose from the three most probable tokens based on their likelihood, influenced by the temperature setting.

During each token selection step, the model identifies the top-K tokens with the highest probabilities. These tokens are then further filtered through the top-P setting, with the final choice made using temperature sampling.

Set a lower top-K value for more deterministic responses or a higher value for more varied and creative outputs. The default top-K value is 40.

Top-P

Top-P, also known as nucleus sampling, adjusts how the model selects tokens by considering their cumulative probabilities. The model selects tokens starting from the most to the least probable until the sum of their probabilities matches the top-P value. For instance, if tokens A, B, and C have probabilities of 0.3, 0.2, and 0.1 respectively, and the top-P value is 0.5, the model will only choose between tokens A and B while excluding C.

Lower top-P values result in more focused and less random responses, while higher values encourage greater diversity in outputs. The default top-P value is 0.95.

Prompt Iteration Strategies

Crafting effective prompts is an iterative process, often requiring several adjustments to achieve consistent and desired results. Below are some strategies to refine your prompt design:

Experiment with Different Phrasing

Varying the wording or structure of your prompts can lead to different responses from the model, even if the intended meaning stays the same. If the response isn’t as expected, try rephrasing the prompt.

Examples:

• Version 1: How do I bake a pie?
• Version 2: Suggest a recipe for a pie.
• Version 3: What's a good pie recipe?

Switch to a Similar Task

If the model struggles to follow instructions for a specific task, try reframing it as an analogous task that can accomplish the same goal.

For example, to categorize a book using predefined categories:

Prompt:

Which category does The Odyssey belong to?

• Thriller
• Sci-fi
• Mythology
• Biography

Response:

The Odyssey belongs to the category of mythology.

Here's why:

• Mythology: The Odyssey narrates the epic journey of Odysseus, a legendary hero from Greek mythology, as he struggles to return home after the Trojan War. Along the way, the story weaves in gods, mythical creatures, and supernatural events that are hallmarks of ancient Greek lore.

(gemini-1.5-flash): While the response is accurate, the model provided an answer outside the set boundaries of the options. To ensure concise answers, you can reframe your instructions as a multiple-choice question, prompting the model to pick one of the given options directly.

Example Prompt:

Multiple-choice question: Which of the following best describes the book The Odyssey?

Options:

• Thriller
• Sci-fi
• Mythology
• Biography

Response:

The correct answer is Mythology.

(gemini-1.5-flash): If the model's response is not as expected, try adjusting the order of the content within the prompt. The sequence of examples, inputs, and context can influence the model’s behavior. Experiment with reordering to optimize results.

Suggestions for Prompt Structure:

• Version 1: [Examples] → [Context] → [Input]
• Version 2: [Input] → [Examples] → [Context]
• Version 3: [Examples] → [Input] → [Context]

Fallback Responses:

A fallback response occurs when the model triggers a safety filter or encounters unclear instructions. For instance, the fallback might be: "I'm not able to assist with that as I'm just a language model."

How to Address Fallbacks:

If you encounter a fallback response, consider increasing the temperature setting to encourage more varied and creative responses. This small adjustment can often help achieve the desired outcome.

Things to Avoid

Refrain from depending solely on models to produce factual information, as their outputs may not always be accurate.

Exercise caution when using models for solving math or logic-related problems.

Next Steps

With a stronger grasp of prompt design, take the initiative to create your own prompts using Google AI Studio.

Interested in expanding your skills? Explore Prompting with Media Files to learn more about multimodal prompting.