Claude is a family of large language models developed by Anthropic. Claude models have three key capabilities:

• Advanced reasoning
• Claude can perform complex cognitive tasks that go beyond simple pattern recognition or text generation
• Code generation
• Start creating websites in HTML and CSS, or debugging complex code bases
• Multilingual processing
• Translate between various languages in real-time, practice grammar, or create multi-lingual content

We are often familiar with creating prompts for the GPT series and tend to use the same style when working with other models. However, when using Claude, we need to make some adjustments because Claude’s models are trained using different methods or techniques.

It’s important to understand that prompts that work well with GPT may not be as effective with Claude. According to Anthropic’s documentation, Claude performs best with prompts that are clear, direct, and detailed. To achieve better results with Claude, it’s essential to tailor your prompts accordingly. The complexity of your prompts should match the complexity of your task- the more complex the task, the more detailed your instructions should be. Here are some prompt engineering techniques for Claude Al:

1.   Use XML Tags

We are used to creating prompts in various formats for GPT. However, Claude models are more familiar with XML tags because these models have been finetuned with XML tags. Therefore, it’s important to include tags like<> and> to 9. Limiting Extraneous Tokens References Best Practices for Prompt Engineering 3 differentiate between instructions, examples, questions, context, output format, and input data.

2. Be Direct, Concise, Specific, and Allow Claude to Say “I Don’t Know”

Instead of telling the models what to avoid, it’s better to use affirmative instructions to specify what the model should do rather than saying what it shouldn’t do.

We should also allow Claude to say “I don’t know” to prevent hallucination and avoid generating inaccurate or misleading information.

3. Specify the Output Format

Claude’s models generally tend to be chatty. This can be a problem if we need the model to follow a specific output format. To address this, we can use an assistant message to ensure that Claude models start their responses consistently.

4. Assign Claude a Role (System Prompts)

We can assign Claude a role to mimic the style or character of an expert, such as an elementary teacher, content writer, or any other relevant persona. This defined role can also help improve the model’s performance.

5. Use Examples or Few-Shot Learning

Some articles suggest that providing Claude with examples is a powerful way to guide it in generating the best response. We should identify a general example relevant to our use case and observe how this leads to more accurate and consistent results. While using more examples can improve outcomes, it may also increase cost and latency.

6. Allow Claude to Think (Chain of Thought)

We can instruct the Claude model to think before answering a question. By adding a new output format, we can allow Claude to share its thought process or provide reasoning before giving a conclusion or final answer. This technique can reduce errors, especially in math, logic, analysis, or other complex tasks.

7. Chain Complex Prompts

We can break down complex tasks into steps to help Claude perform better on such tasks, as shown here:

8. Tips for Handling Long Contexts

If we’re dealing with longer documents, we should place our key question or instruction at the end of the prompt.

In general, there are two versions of the Mistral model, each available in various parameter sizes:

1. Mistral

Mistral offers different model variations in various sizes:

• Mistral 78: Ideal for tasks such as answering questions, generating outlines, or interpreting text. It is a strong performer in multilingual capabilities, reasoning, math, and code generation.
• Mistral Large: It reaches top-tier reasoning capabilities. It can be used for complex multilingual reasoning tasks, including text understanding, Best Practices for Prompt Engineering 8 transformation, and code generation.

2. Mixtral

Mixtral is an upgraded and larger model compared to Mistral:

• Mixtral 8×78: Suitable for real-time applications, demonstrating strong capabilities in mathematical reasoning, code generation, and multilingual tasks. It supports languages such as English, French, Italian, German, and Spanish.
• Mixtral 8x22B: With its massive parameter size, this model excels in understanding subtle nuances in natural language. It provides more intelligible and logically relevant responses, making it ideal for tasks like experimental writing, complex question answering, and writing synopses.

1.   Understand Mixtral’s Capabilities

• Review Model Strengths: Mixtral is particularly strong in generating creative text, understanding complex instructions, and maintaining conversational context over longer interactions. Familiarize yourself with these strengths to leverage them effectively.
• Task Specialization: Mixtral models excel in tasks such as summarization, translation, and content creation. Tailor prompts to these specific tasks for better results. However Mixtral models can be used to do classification tasks with step­ bystep instructions and few-shot examples.

2.   Use Clear and Specific Instructions

• Direct Commands: Use straightforward and unambiguous language. Avoid vague terms. For example, instead of asking, “Can you summarize this?” you could say, “Summarize the key points of the following article in three bullet points.”
• Contextual Prompts: Provide necessary context within the prompt to reduce ambiguity. If you’re asking the model to generate a story, include details like the genre, characters, and setting.

Example:

3.   Experiment with Iterative Refinement

• Test and Refine: Start with a simple prompt and gradually refine it based on the output. Adjust the prompt length, wording, or structure if the initial results are not satisfactory.
• Iterative Feedback: Use the model’s responses as feedback for prompt adjustment. If the output is too generic, add more details or constraints to the prompt.

Example: 

Initial Prompt:

Refined Prompt:

4.   Leverage Mixtral’s Advanced Features

• Chain-of-thought: For complex instructions, break down the instructions into step-by-step smaller instructions.
• Role Assignment: Assign specific roles or perspectives to the model to guide its response style. For example, ask the model to respond as an expert in a particular field.

Example:

 5.   Provide Facts, Examples, and Formats

• Provide facts: Include facts within the prompt to improve the context accuracy of model’s result.
• Input Examples: Include examples within the prompt to show the model the type of response you expect.
• Structured Output Requests: If you need the output in a specific format (e.g.,
• bullet points, JSON), clearly specify this in your prompt.

Example: 

Prompt example:

Llama 3.1 is the latest iteration of Meta’s Large Language Model (LLM) series, representing a significant advancement in Al technology. This open-source model is designed to be highly versatile and powerful, catering to a wide range of applications from natural language processing to specialized domain tasks. Here are the key features and variants of Llama 3.1:

Key Features of Llama 3.1

1. Multilingual Support: Llama 3.1 supports eight languages, including English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai, making it accessible and useful for a global audience.
2. Extended Context Length: The model boasts an extended context length of 128k tokens, allowing it to process and understand much longer pieces of text for more complex tasks and analyses.
3. Tool Calling Capabilities: The instruct-tuned models in Llama 3.1 are finetuned for tool calling, making them suitable for agentic use cases. They come with two built-in tools (search and mathematical reasoning with Wolfram Alpha) and support custom JSON functions for further extensibility.
4. Improved Instruction and Safety Measures: The instruct models have been optimized to follow user instructions more effectively. With the introduction of Llama Guard 3 and Prompt Guard, Meta is offering robust tools to improve the safety and security of Al applications built with Llama 3.1.

Variants of Llama 3.1

Llama 3.1 is available in three sizes: 88, 708, and 4058 parameters, each offered in both base and instruct-tuned versions. This variety allows users to choose the model that best fits their specific needs, whether it’s for efficient deployment and development, large-scale Al applications, or synthetic data generation

1. llama -3.1-8b – base pretrained 8 billion parameter model
2. llama-3 .1-70b – base pretrained 70 billion parameter model
3. llama-3.1-405b – base pretrained 405 billion parameter model
4. llama-3.1-8b-instruct – instruction fine-tuned 8 billion parameter model
5. llama-3.1-70b-instruct – instruction fine-tuned 70 billion parameter model
6. llama-3.1-405b-instruct- instruction fine-tuned 405 billion parameter model  (flagship)

Here are the best practices for crafting effective prompts from Llama Prompting guides:

1. Be clear and concise: Your prompt should be easy to understand and provide enough information for the model to generate relevant output. Avoid using jargon or technical terms that may confuse the model.
2. Use specific examples: Providing specific examples in your prompt can help the model better understand what kind of output is expected. For example, if you want the model to generate a story about a particular topic, include a few sentences about the setting, characters, and plot.
3. Vary the prompts: Using different prompts can help the model learn more about the task at hand and produce more diverse and creative output. Try using different styles, tones, and formats to see how the model responds.
4. Test and refine: Once you have created a set of prompts, test them out on the model to see how it performs. If the results are not as expected, try refining the prompts by adding more detail or adjusting the tone and style.
5. Use feedback: Finally, use feedback from users or other sources to continually improve your prompts. This can help you identify areas where the model needs more guidance and make adjustments accordingly.

1.   Explicit Instructions

Detailed, explicit instructions produce better results than open-ended prompts:

1. Stylization

2. Formatting

3. Restrictions

2. Zero-Shot Prompting

Large language models like Llama 3 are unique because they are capable of following instructions and producing responses without having previously seen an example of a task. Prompting without examples is called “zero-shot prompting”.

3. Few-Shot Prompting

Adding specific examples of your desired output generally results in more accurate, consistent output. This technique is called “few-shot prompting”.

In this example, the generated response follows our desired format that offers a more nuanced sentiment classifer that gives a positive, neutral, and negative response confidence percentage.

4. Role Based Prompts

Llama will often give more consistent responses when given a role. Roles give context to the LLM on what type of answers are desired.

5. Chain of Thought Technique

Simply adding a phrase encouraging step-by-step thinking “significantly improves the ability of large language models to perform complex reasoning”. This technique is called “CoT” or “Chain-of-Thought” prompting.

6. Self-Consistency

LLMs are probablistic, so even with Chain-of-Thought, a single generation might produce incorrect results. Self-Consistency introduces enhanced accuracy by selecting the most frequent answer from multiple generations (at the cost of higher compute}:

7. Retrieval-AugmentedGeneration

Retrieval-Augmented Generation, or RAG, describes the practice of including information in the prompt that has been retrieved from an external database. It’s an effective way to incorporate facts into your LLM application and is more affordable than fine-tuning which might also negatively impact the foundational model’s capabilities.

8. Program-Aided Language Models

LLMs, by nature, aren’t great at performing calculations. While LLMs are bad at arithmetic, they’re great for code generation. Program-Aided Language leverages this fact by instructing the LLM to write code to solve calculation tasks.

9. Limiting Extraneous Tokens

A common challenge is generating a response without extraneous tokens (e.g. “Sure! Here’s more information on…”).

By combining a role, rules and restrictions, explicit instructions, and an example, the model can be prompted to generate the desired response.

Article originally posted on Datasaur.ai as a PDF download. Reposted with permission.