Introduction

When interacting with advanced Artificial Intelligence (AI) models like Claude, how you structure your prompts can significantly impact the quality of responses. Prompt engineering involves more than just asking questions—it is about strategically designing prompts to get clear, accurate, and useful answers. With large language models (LLMs) and other forms of generative AI increasingly being incorporated into applications, prompt engineering is key to saving costs and time while improving the overall model performance. 

In this article, we will break down essential best practices for prompt engineering with Claude, covering techniques like clarity, iteration, and context that help optimize responses for tasks ranging from code generation to creative writing.

What is Prompt Engineering?

Prompt engineering involves the careful design of instructions that guide AI models, such as Claude, to generate desired outcomes. These prompts are the natural language inputs that provide context and set clear expectations, ensuring that the model delivers accurate and relevant results. Properly structured prompts lead to high-quality outputs, while poorly designed prompts can result in inaccuracies, hallucinations, or irrelevant responses. Through precise prompt engineering, the capabilities of these models can be harnessed to produce consistently reliable and useful answers.

Best Prompt Engineering Practices for Claude

These are the recommended guidelines for prompt engineering with Claude. Try to follow these in the same order and build upon each one if they do not get you your desired results.

0. Basic Structure

When interacting with Anthropic’s Claude model via the Messages API, it is important to follow a structured format to ensure optimal responses. Key parameters include:

• <model>: Specifies the version of Claude to use (e.g., <claude-v1>).

• <max_tokens>: Defines the maximum number of tokens Claude can generate before stopping. It ensures responses are concise.

• <messages>: An array of alternating input messages between user and assistant roles. The conversation always begins with a user message.

Optional parameters like <system> provide additional context and instructions to guide Claude’s behavior. For example, the <system> prompt can dictate response style, such as encouraging critical thinking or limiting response variability with the temperature parameter.

Correct formatting is crucial, as messages must strictly alternate between <user> and <assistant> roles, with the conversation always starting from the <user>. If this structure is not followed, the API will return an error.

System prompts are particularly valuable for setting guidelines on how Claude should frame its responses. For example, a system prompt might instruct Claude to avoid directly answering questions but instead prompt further inquiry, providing a more dynamic and interactive response experience. By building a good base, developers can ensure that iteration time to get to the ideal prompt will be minimal and avoid incurring unnecessary charges from the API.

1. Be Clear and Specific

When working with Claude, clarity is key. Imagine explaining a task to a new employee who has no prior knowledge of your processes. The more specific you are, the better the result will be. Keeping this in mind, once you have finished writing your prompt, ask a colleague or friend to follow your prompt, ideally someone who has minimal context on the problem. If they are confused by your instructions, it is likely that Claude will be too. This is the “Golden Rule of Clear Prompting”. 

To avoid vague instructions, keep these tips in mind:

Provide Contextual Information: Offer relevant details that help Claude understand the task better. These can include the purpose of the task, the target audience, where the task fits within a broader workflow, and what a successful completion would look like.

Be Specific About the Desired Output: Clearly state what you expect from Claude. For instance, if you only need code, explicitly mention that.

Provide Instructions as Sequential Steps: Use numbered lists or bullet points to outline the task clearly. This approach helps ensure that Claude follows the instructions precisely.

Providing this level of detail prevents confusion and ensures Claude delivers the most accurate and relevant responses.

2. Provide Examples

Using examples in your prompts is an effective strategy to guide the model toward generating accurate and structured outputs. This technique, known as multishot or few-shot prompting, involves providing several well-crafted examples to demonstrate the desired format or behavior.

Including 3-5 examples in a prompt can be a shortcut to achieving high-quality results. The more examples provided, the better Claude will understand expectations, especially for tasks that require structured outputs or consistent formats. These are some things to keep in mind when crafting the prompts:

Relevance: Ensure the examples are directly applicable to your use case. Claude will perform best when the examples align with the task at hand.

Diversity: Cover edge cases, variations, and challenges within your examples. A diverse set of examples prevents the model from picking up unintended patterns, ensuring it can handle a wide range of input scenarios.

Clarity: Clearly wrap your examples in <example> tags, and use <examples> tags to structure multiple examples. This helps Claude differentiate between different examples and understand the scope of the task.

3. Let Claude Think (Chain of Thought Prompting)

Chain of Thought (CoT) prompting is an effective technique to boost Claude’s performance when tackling complex tasks, such as analysis, problem-solving, or research. By encouraging the model to think step-by-step, CoT allows Claude to break down problems, leading to more accurate, coherent, and nuanced responses, helping reduce errors in logic, analysis, and mathematical computations. 

This structured approach also leads to more organized and cohesive outputs, improving overall response quality. Additionally, allowing Claude to demonstrate its thought process provides a valuable debugging tool, enabling users to identify areas of unclear prompts and refine them for better results. 

CoT is especially useful for tasks such as solving complex mathematical problems, conducting multi-step analyses, drafting detailed documents, or making decisions involving numerous factors. However, not all tasks require this depth of thinking, and using CoT in situations where latency is a concern may be counterproductive. Therefore, CoT should be applied selectively to balance performance with response time. Here are some techniques to keep in mind when implementing CoT:

Basic CoT Prompting: The simplest method involves asking Claude to "Think step-by-step." While effective in some cases, this lacks detailed guidance on how to structure the thought process and may not work as well for highly specific tasks. For example,

User: Generate personalized product recommendations for a customer. 
Customer Information: <customer>{{CUSTOMER_DETAILS}}</customer> 
Product Information

Guided CoT Prompting: In this approach, users outline specific steps for Claude to follow in its thinking process. This ensures more thorough reasoning but can still make it difficult to separate the final answer from the underlying thought process. This is how it differs from basic CoT prompting:

User: Generate personalized product recommendations for a customer. 
Customer Information: <customer>{{CUSTOMER_DETAILS}}</customer> 
Product Information

Structured CoT Prompting: The most detailed method, using tags like <thinking> and <answer> to clearly separate Claude’s reasoning from the final output. This provides both structure and transparency, allowing users to easily strip out the thinking process if only the answer is needed. This is what a structured prompt would look like:

User: Generate personalized product recommendations for a customer. 
Customer Information: <customer>{{CUSTOMER_DETAILS}}</customer> 
Product Information

4. Use XML Tags

Using XML tags in prompts can significantly improve the clarity, accuracy, and flexibility of interactions with Claude. By incorporating tags such as <instructions>, <example>, and <formatting>, users can help Claude distinguish between different components of the prompt, minimizing confusion and enhancing the quality of responses. This method ensures that Claude interprets instructions accurately and follows the structure laid out in the prompt, leading to more reliable and organized outputs.

Additionally, XML tags provide flexibility by allowing easy adjustments to specific sections of a prompt without the need to rewrite the entire input. They also improve post-processing by enabling easy extraction of specific parts of Claude’s responses, especially when working with structured outputs.

Although there are no predefined XML tags that Claude is explicitly trained to recognize, it is important to use logical and intuitive tag names that align with the content they enclose. Some best practices to keep in mind are:

Be consistent: Use the same tag names across your prompts and refer to them when necessary.

Nest tags: Use hierarchical tags for content that has layers of detail.

Combine techniques: For better results, mix XML tags with multishot prompting or chain-of-thought techniques to create well-structured, high-performance prompts.

5. Use System Prompts Effectively to Give Claude a Role

Using role prompting in Claude can greatly enhance its performance by giving it a specific role through the system parameter. This technique transforms Claude from a general assistant into a virtual expert tailored to the task at hand. By defining a role, such as a legal analyst or financial advisor, you can see the following benefits:

Enhanced accuracy: Helps Claude handle complex tasks like legal analysis or financial modeling more effectively.

Tailored tone: Adjusts Claude’s communication style to suit specific roles, whether it is a formal CFO or a creative copywriter.

Improved focus: Keeps Claude on track by setting a clear role context, ensuring it stays within the task’s boundaries.

In practice, role prompting is implemented by specifying the role within the system parameter of the Messages API. For example, setting Claude as a seasoned data scientist could help focus on a specific task, such as analyzing a dataset for anomalies.

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-3-5-sonnet-20240620",
    max_tokens=2048,
    system="You are a seasoned data scientist at a Fortune 500 company.", # <-- role prompt
    messages=[
        {"role": "user", "content": "Analyze this dataset for anomalies: <dataset>{{DATASET}}</dataset>"}
    ]
)

print(response.content)

Role prompting can be customized depending on the task’s needs, with different roles yielding unique insights. A data scientist specializing in customer insights may interpret the same dataset differently than a marketing strategist, offering diverse perspectives based on the given role. Even specifying that the data scientist in question works at a Fortune 500 company can yield different results.

6. Prefilling

​​Prefilling Claude’s response is a powerful technique that allows you to guide the assistant’s behavior and control its output more precisely. By pre-populating part of the Assistant message, you can instruct Claude to skip generic preambles, enforce specific output formats like JSON or XML, and even maintain character consistency in role-playing scenarios. This method can significantly improve Claude’s performance, especially in situations where it might otherwise deviate from the desired response structure.

You can implement prefilled responses by including the desired initial text in the Assistant message, allowing Claude to complete the rest of the response. Here is an example using the Anthropics API:

import anthropic

client = anthropic.Anthropic()
response = client.messages.create(
    model="claude-3-5-sonnet-20240620",
    max_tokens=1024,
    messages=[
        {"role": "user", "content": "What is your favorite color?"},
        {"role": "assistant", "content": "As an AI assistant, I don't have a favorite color. But if I had to pick, it would be green because"}  # Prefill here
    ]
)

In this example, Claude’s response is prefixed with "As an AI assistant, I don’t have a favorite color...," which guides the rest of its answer. Prefilling like this provides more control over the structure and tone of the response.

Some examples of prefilling are:

Controlling Output Formatting and Skipping Preambles: When working with specific formats like JSON or XML, you can prefill the opening characters to enforce proper formatting. For instance:

{"role": "assistant", "content": "{"}  # Prefill here for JSON output

This forces Claude to output a clean JSON object by skipping any introductory text, making the output more concise and programmatically useful without needing additional processing.

Maintaining Character in Role-Play Scenarios: In scenarios where Claude is required to play a role, such as a specific character in a dialogue or role-play, prefilling with a bracketed role name can keep it consistent over time:

{"role": "assistant", "content": "[CFO] I believe the best financial strategy would be..."

This technique helps ensure Claude stays in character, even in long or complex conversations, and is especially effective when combined with role prompting through the system parameter.

7. Prompt Chaining

Prompt chaining is an effective method to break down complex tasks for Claude, ensuring that each step is managed with precision. When a single prompt tries to handle too many elements, Claude can overlook or mishandle parts of the task. By dividing tasks into smaller, manageable subtasks, you can boost the accuracy and clarity of Claude’s outputs, making it easier to trace and fix any issues that arise during the process.

When tasks involve complex transformations or require multiple instructions, chaining prompts ensures that Claude does not drop steps or muddle the output. If a specific step doesn't perform well, isolating and fine-tuning that step becomes straightforward without having to rework the entire task. It is particularly useful for multi-step processes like research synthesis, document analysis, or iterative content creation, where each step builds on the previous one.

To chain prompts, follow these best practices:

Identify subtasks: Break the task into smaller, sequential steps. Each step should have a single, clear objective.

Use XML for handoffs: XML tags can be used to pass outputs from one prompt to the next in a clear and structured manner.

Iterate and refine: Based on Claude's performance, refine individual subtasks as needed to improve the overall workflow.

In an advanced use case, one could also chain prompts to have Claude review its own work. This involves generating an initial output, then creating a second prompt where Claude evaluates its accuracy, clarity, and completeness, and finally using this feedback to refine the original content.

For example, in a research summary task, you can chain prompts to first summarize a medical research paper, then ask Claude to review the summary, and finally, use feedback from the review to improve the original summary. This step-by-step approach allows each part of the task to be handled with greater care.

By chaining prompts in this way, Claude can handle complex tasks with multiple steps while maintaining focus and delivering higher-quality results.

8. Long Context Prompting

Long context prompting in Claude 3 models takes advantage of an extended context window of 200K tokens, allowing the model to process large amounts of information effectively. This is ideal for complex, data-rich tasks, such as analyzing lengthy documents or synthesizing information from multiple sources. To fully leverage this power, a few key strategies can enhance Claude’s performance, especially when dealing with extensive inputs.

Place Longform Data at the Top: When working with large documents (20K+ tokens), positioning them at the beginning of the prompt—above the query, instructions, and examples—can significantly improve the output. This gives Claude immediate access to the most crucial information.

Position Queries at the End: When queries are placed after the documents, especially in multi-document prompts, Claude’s ability to provide accurate responses improves by up to 30%, based on internal tests. This structure allows Claude to process the relevant information before tackling the query.

Use XML Tags to Organize Content: When handling multiple documents, structure the content using XML tags like <document>, <document_content>, and <source>. This organization helps Claude navigate and process the information more efficiently, reducing the chance of confusion or missed details. For example:

<document>
  <document_content>Text of the first document...</document_content>
  <source>Document 1: Author, Title, Date</source>
</document>
<document>
  <document_content>Text of the second document...</document_content>
  <source>Document 2

Quote Extraction for Focus: For long document tasks, it is useful to instruct Claude to extract and ground responses in relevant quotes from the documents before proceeding with analysis or summarization. This helps Claude focus on the most pertinent parts of the input, minimizing distractions from less relevant sections. For instance, you could ask Claude to quote key sections before providing an overall summary, ensuring that the response remains aligned with the original documents.

These techniques optimize Claude’s extended context capabilities, allowing for better accuracy and clarity in responses, particularly when dealing with complex, multi-document inputs.

Best Image Prompting Practices With Claude

The Claude 3 family’s advanced vision capabilities allow it to analyze and interpret a wide range of visual content, from charts and diagrams to complex images. This ability warrants a few additional tips to keep in mind, in addition to all the general text-based prompt techniques discussed already.

1. Image Placement and Size

For optimal performance with Claude 3’s vision capabilities, place images at the start of the prompt. Resize images to balance clarity and size, as recommended by Anthropic’s guidance on image sizing.

2. Applying Traditional Techniques

To reiterate, traditional prompting techniques, such as providing clear instructions and allowing time for Claude to think, are also effective with images. This approach ensures detailed and accurate analysis of visual content.

3. Visual Prompts Directly in Images

Images can serve as prompts themselves. For example, you can embed questions directly within the image or use the image content to inform responses, as shown in examples involving counting objects or solving geometric problems.

4. Handling Multiple Images

When using multiple images in a prompt, clearly label each image with tags to distinguish them. This helps Claude analyze and provide insights based on the specific visual data from each image.

5. Detailed Descriptions for Charts and Graphics

For complex charts or graphics, guide Claude by asking it to describe data points or identify color HEX codes. This helps in accurately interpreting visual information, especially when dealing with similar colors or numerous data points.

6. Narration for Slide Decks

Convert slide decks into individual images for each slide and ask Claude to narrate each slide’s content. Ensure Claude includes all numbers and details from the slides without adding extraneous information. This method provides a comprehensive overview and aids in answering questions about the entire slide deck.

Trouble Getting Started?

This guide only applies once you have an initial prompt in hand. Anthropic has developed a prompt generation tool to assist users in creating effective AI prompts, particularly when they encounter difficulty starting from a blank page. This tool automatically generates high-quality prompt templates, tailored to specific tasks, and incorporates established prompt engineering best practices. By providing an initial framework, it offers users a valuable tool allowing for easier testing and refinement of the prompts. Check out the various ways to get started with prompting for Claude on their official website.

Conclusion

In conclusion, effective prompt engineering with Claude involves clear, structured instructions, relevant examples, and utilizing advanced techniques like Chain of Thought prompting and prompt chaining. By incorporating XML tags for organization, system prompts for tailored responses, and prefilled content for control, prompt engineers can significantly enhance the model's performance. These practices empower users to generate accurate, contextually aware, and high-quality responses.