Considerations for effective prompts engineering and prompt frameworks
In this article, we will discuss importance of prompt frameworks for an effective prompts design, mental model to simplify prompts writing, and pros and cons of each of the frameworks.
Published Oct 9, 2024
In the realm of Artificial Intelligence (AI) and Natural Language Processing (NLP), prompts serve as the catalysts to ignite meaningful interactions between humans and machines. These short, instructive sentences or phrases guide AI Large Language Model (LLM) in generating responses that are contextually relevant, accurate, and insightful. Whether you're asking a AI bot to set a reminder or querying a complex AI model for data analysis, it all starts with a well-crafted prompt. This article talks about prompt writing strategies and considerations for a scalable and effective prompt designs.
In the fast-paced world of AI and NLP filled with large number of models and where models are becoming increasingly sophisticated, the need for a structured approach to prompt writing is more crucial than ever. Frameworks provide a systematic way to approach the formulation of prompts, incorporating various elements like context, specificity, and clarity, which are essential for generating meaningful responses.
Each LLM is unique in terms of its maturity level, syntax, capabilities, model quality and performance, and hence approach to writing effective prompts differ based on the model you’re interacting with. There are several resources and article available (some from the model providers) on the prompt engineering techniques and approaches specific to the LLM. However, its important to have a mental model and structured method that provide users logical framework in approaching the prompts irrespective of the LLMs in question. This conceptualization does not only help simplify approach to write effective prompts for anyone, and in-turn eases adoption of any LLM, but also helps Customers improve the management and scaling of prompts design or migration to newer version of the LLM or different LLM.
On the other hand, an unstructured approach often results in prompts leading to challenges that can undermine the effectiveness of an AI model. Below are some of the common issues:
Ambiguity - Unstructured prompts can be vague, leading to ambiguous responses from the AI model. For example, asking a financial AI model, "Tell me about stocks," could yield a wide range of answers, from stock definitions to current market trends. The lack of specificity makes it difficult for the model to provide a focused and useful answer.
Inconsistency - Without a structured framework, prompts can vary greatly in their construction and intent, leading to inconsistent results. This inconsistency can be particularly problematic in professional settings where uniformity and reliability are crucial.
Inefficiency - Unstructured prompts often require multiple iterations to get the desired output, wasting both time and computational resources. This inefficiency becomes increasingly problematic as AI models grow more complex and resource-intensive.
Ethical and Compliance Risks - In sectors like healthcare and finance, where compliance with regulations is mandatory, unstructured prompts can lead to outputs that are not only incorrect but also legally risky. A structured approach ensures that all necessary guidelines and parameters are considered when formulating a prompt.
Ambiguity - Unstructured prompts can be vague, leading to ambiguous responses from the AI model. For example, asking a financial AI model, "Tell me about stocks," could yield a wide range of answers, from stock definitions to current market trends. The lack of specificity makes it difficult for the model to provide a focused and useful answer.
Inconsistency - Without a structured framework, prompts can vary greatly in their construction and intent, leading to inconsistent results. This inconsistency can be particularly problematic in professional settings where uniformity and reliability are crucial.
Inefficiency - Unstructured prompts often require multiple iterations to get the desired output, wasting both time and computational resources. This inefficiency becomes increasingly problematic as AI models grow more complex and resource-intensive.
Ethical and Compliance Risks - In sectors like healthcare and finance, where compliance with regulations is mandatory, unstructured prompts can lead to outputs that are not only incorrect but also legally risky. A structured approach ensures that all necessary guidelines and parameters are considered when formulating a prompt.
In brief, Prompt frameworks offer a solution to these challenges by providing a structured methodology for crafting effective prompts. They act as a blueprint, guiding the user in asking the right questions in the right way. By adhering to a framework, users can ensure that their prompts are clear, focused, and aligned with their objectives, thereby maximizing the utility and efficiency of AI models.
There are several frameworks out there to help with the mental model and conceptualization. These frameworks offer a systematic way to construct prompts, ensuring that they are clear, focused, and effective in eliciting the desired response from LLMs.
In this article, we will discuss various frameworks such as RTF Request, Task, Format), SMART (Specific, Measurable, Achievable, Relevant, Time-bound), RISEN (Role, Instructions, Steps, End goal, and Narrowing), RODES (Role, Objective, Details, Examples, and Sense Check), Chain of Thought (CoT), and Chain of Destiny (CoD) frameworks that are designed to optimize prompt writing and their strengths, applications, and best practices.
In the following sections, we will delve into various prompt frameworks, their unique features and applications. Whether you are a novice looking to understand the basics or a seasoned professional aiming to refine your skills, these frameworks offer invaluable insights into the art and science of prompt writing.
The RTF Framework stands for Request, Task, and Format. By breaking down the prompt into three distinct components RTF Framework ensures that the AI model receives a well-defined, actionable request that is easy to understand and execute. This Framework is particularly useful in scenarios where a complex or multi-step interaction is required. It is ideal for professional settings, such as customer service, healthcare, and data analysis, where clarity and precision are paramount. The framework is also beneficial in educational contexts, helping students interact more effectively with educational AI tools.
Request - The "Request" component is the initial part of the prompt where the user specifies what they want the AI model to do. This is usually a direct question or command aimed at eliciting a specific type of response. The Request sets the stage for the interaction, providing the AI model with the context it needs to generate a meaningful output.
Task - The "Task" component elaborates on the Request, providing additional details or specifications. This could include parameters, conditions, or any other information that helps the AI model understand the scope and requirements of the task at hand.
Format - The "Format" component is optional and is used to specify the desired format of the AI model's response. This could be a particular data structure, a specific layout, or even a preferred language style. Format helps tailor the output to meet specific needs or preferences, making the interaction more user-friendly and effective.
Task - The "Task" component elaborates on the Request, providing additional details or specifications. This could include parameters, conditions, or any other information that helps the AI model understand the scope and requirements of the task at hand.
Format - The "Format" component is optional and is used to specify the desired format of the AI model's response. This could be a particular data structure, a specific layout, or even a preferred language style. Format helps tailor the output to meet specific needs or preferences, making the interaction more user-friendly and effective.
Prompt Template:
Request: \[insert the question or command you want AI to act on.\]
Task: \[Additional details or specifics to elaborate on the Request.\]
Format: \[Desired output format.\].
Request: \[insert the question or command you want AI to act on.\]
Task: \[Additional details or specifics to elaborate on the Request.\]
Format: \[Desired output format.\].
- Customer Service Use CaseRequest: "Tell me about your refund policy."
Task: "Specifically, I want to know the conditions under which I can return an electronic item."
Format: "Please provide the information in bullet points." - Healthcare Use CaseRequest: "List the symptoms of Type 2 diabetes."
Task: "Include both common and rare symptoms."
Format: "Organize the symptoms in order of severity." - Data Analysis Use CaseRequest: "Generate a sales report for the last quarter."
Task: "Include revenue, expenses, and net profit, broken down by department."
Format: "Present the data in a bar chart."
By applying RTF Framework, users can craft prompts that are not only clear and actionable but also tailored to their specific needs and contexts. This structured approach significantly enhances the quality of the interaction, making it a valuable tool for anyone looking to leverage the power of AI and NLP effectively.
The CoT Framework is a specialized prompt engineering technique designed to enhance the reasoning capabilities of language models like Claude, LIama, Mistral, GPT-4. By instructing the AI to approach a problem "step-by-step," this framework guides the model through a logical sequence of thoughts, making it particularly effective for complex analytical tasks and problem-solving. This framework excels in scenarios that require detailed analysis or problem-solving. Whether you're looking to dissect a complex issue or find a solution to a challenging problem, CoT helps by encouraging the AI to think logically and sequentially.
There are two most common and effective ways to invoke CoT with models a) Adding a phrase such as “Think step-by-step” at the end of your prompt. b) breaking down the complex task into series of intermediate reasoning steps or questions in the prompts to help guide model’s thinking. This is similar to humans' approach to complex problem by decomposing it into simpler and small steps. These strategies help AI model to break down the problem into smaller, more manageable parts and think through each one logically and make it easier for us to debug and inspect the model behavior.
- Use Case : Analyze Market Trends Prompt Template using Step-by-Step directive: "\[insert your prompt instructions\]. Think step-by-step."Prompt: "What factors are contributing to the declining market share of our product? Think step-by-step."In above example with Claude 3 family model, AI would systematically evaluate various factors such as competition, consumer behavior, and marketing strategies, providing a comprehensive analysis that could lead to actionable insights.
- Use Case : Essay Reviewer and Commentator Prompt:" You are a commentator. Your task is to write a report on an essay.
When presented with the essay, come up with interesting questions to ask, and answer each question.
Afterward, combine all the information and write a report in the markdown format.# Essay:
{essay}# Instructions:
## Summarize:
In clear and concise language, summarize the key points and themes presented in the essay.## Interesting Questions:
Generate three distinct and thought-provoking questions that can be asked about the content of the essay. For each question:
- After "Q: ", describe the problem
- After "A: ", provide a detailed explanation of the problem addressed in the question.
- Enclose the ultimate answer in <>.## Write a report
Using the essay summary and the answers to the interesting questions, create a comprehensive report in Markdown format."In above prompt with Mistral model, we break down the task into three steps: summarize, generate interesting questions, and write a report. This helps the language to think in each step and generate a more comprehensive final report. We also asked LLMs to generate three questions and provide detailed explanations for each question to automatically guide the reasoning and understanding process of the model.
The Chain of Thought Framework simplifies the process of tackling complex issues by encouraging a step-by-step analytical approach, making it a valuable tool for anyone looking to harness the problem-solving capabilities of AI. This technique can be useful for generating more thoughtful and well-reasoned responses from language models.
The RISEN Framework is a prompt engineering technique designed to break down complex or constrained tasks into actionable components for better execution. The acronym RISEN stands for Role, Instructions, Steps, End goal, and Narrowing (constraints), and it provides a structured approach to guide AI in executing tasks with multiple layers, such as blog posts, research projects, or business plans.
This framework is particularly useful for tasks that require a multi-layered approach, such as creating content, planning projects, or developing business strategies. It is effective in situations where you need the AI to consider multiple variables and constraints while still focusing on a specific end goal.
The RISEN Framework consists of the following key components:
Role (R): Define the role you want the AI to take. This sets the tone and expertise level for the output.
Instructions (I): Clearly state the main task you want the AI to complete.
Steps (S): Provide a numbered list of steps for the AI to follow in completing the task.
End Goal (E): Specify the goal of the output, what you aim to achieve with it.
Narrowing (N): List any constraints that the AI should consider, such as word count limits or specific focus areas.
Role (R): Define the role you want the AI to take. This sets the tone and expertise level for the output.
Instructions (I): Clearly state the main task you want the AI to complete.
Steps (S): Provide a numbered list of steps for the AI to follow in completing the task.
End Goal (E): Specify the goal of the output, what you aim to achieve with it.
Narrowing (N): List any constraints that the AI should consider, such as word count limits or specific focus areas.
Prompt Template:
Role: \[insert the role you want AI to take.\]
Main Task: \[Insert the task you want AI to complete.\]
Steps to complete task: \[Insert numbered list of steps to follow.\]
Goal: \[Insert goal of the output\]
Constraints: \[Enter constraints\].
Role: \[insert the role you want AI to take.\]
Main Task: \[Insert the task you want AI to complete.\]
Steps to complete task: \[Insert numbered list of steps to follow.\]
Goal: \[Insert goal of the output\]
Constraints: \[Enter constraints\].
The flow is sequential, starting with defining the role and ending with setting constraints, to ensure that the AI understands the task in its entirety and can execute it effectively.
- Use Case : Plan a Marketing Campaign for a New Product Prompt:
Role: You are a seasoned marketing strategist with a decade of experience in launching successful products.
Main Task: Develop a comprehensive marketing plan for the launch of our new eco-friendly water bottle.
Steps to complete the task:
1. Begin by outlining the target audience and market research findings.
2. Discuss the marketing channels to be used and why they are effective for this product.
3. Provide a timeline for the campaign, including key milestones and deadlines.
4. End with a budget allocation and expected ROI for each marketing channel.
Goal: The goal is to create a well-rounded marketing plan that maximizes reach and ROI, while aligning with our brand values.
Constraints: Maximum of 1000 words. - Use layman's terms. - Include both online and offline strategies. - Make it actionable.By using RISEN Framework prompt, you can expect to receive a comprehensive marketing plan that starts with identifying the target audience, moves through channel selection and timeline planning, and ends with budget and ROI considerations. All of this will be done within the constraints of a 1000-word limit, easy-to-understand language, and a mix of online and offline strategies.
This example demonstrates how the RISEN Framework can be applied to complex tasks that require a structured approach for effective execution. It ensures that all critical aspects of the task are covered, making it a valuable tool for project planning and strategy development.
The RODES Framework is a structured approach to prompt writing that is particularly useful when you have good examples of your desired output. The acronym stands for Role, Objective, Details, Examples, and Sense Check. Each component serves a specific purpose. This framework is most effective when you have a clear idea of what you want but need the output to adhere to specific styles or examples. It is excellent for creative tasks, marketing copy, or any situation where the style and tone are as important as the content itself.
- Role (R): Specifies the role you want the AI to take on. This sets the tone and expertise level for the output.
- Objective (O): Clearly states what you want the AI to accomplish.
- Details (D): Provides any context or constraints that the AI needs to consider for generating a good output.
- Examples (E): Offers examples that the AI can use as a model for its answer. These examples serve as a guide for the style, tone, or structure.
- Sense Check (S): Asks the AI to confirm its understanding of the objective and guidelines.
- Objective (O): Clearly states what you want the AI to accomplish.
- Details (D): Provides any context or constraints that the AI needs to consider for generating a good output.
- Examples (E): Offers examples that the AI can use as a model for its answer. These examples serve as a guide for the style, tone, or structure.
- Sense Check (S): Asks the AI to confirm its understanding of the objective and guidelines.
- Use Case : Create a LinkedIn Headline for a Digital Marketing ExpertPrompt:
Role:**** You are an experienced copywriter specializing in LinkedIn profiles.
Objective:**** Craft a LinkedIn headline that will attract recruiters in the digital marketing field.
Details:
- The headline should be no longer than 120 characters.
- Use language that highlights expertise and experience.
- Avoid using buzzwords or clichés.
Examples:
Here are some examples to model your answer after (note - these are not on my desired topic, but they illustrate the kind of impactful language and structure that works):
1. "Transforming businesses through data-driven strategies. ROI is my middle name."
2. "Empowering teams to reach their full potential. Leadership through innovation."
3. "Cutting through the noise to deliver measurable PR results. Your story, well told."
Sense Check: Do you understand the objective and the specific guidelines for this task?By employing the RODES Framework in this manner, you can expect a LinkedIn headline that not only stands out but also closely aligns with the style and tone set by the examples. This ensures that the output will meet your specific needs, making the RODES Framework an invaluable asset for tasks that require a blend of accuracy and creativity.
The Chain of Destiny Framework is an iterative approach designed to refine and improve content through multiple cycles of feedback and revision. It is particularly useful for tasks like summarizing articles, enhancing long-form content, and even refining your prompts for better AI output. The CoD Framework is most effective when you have a piece of content that requires multiple iterations for improvement. It's excellent for refining marketing materials, academic papers, or any content that can benefit from a recursive, iterative process
The framework consists of the following key components:
Instructions: Define the content you want to improve.
Recursion: A set of steps that are repeated multiple times to refine the output.
Benchmark: Additional information to guide what constitutes a good output.
Additional Guidelines: Specific rules or constraints to follow during the process.
Instructions: Define the content you want to improve.
Recursion: A set of steps that are repeated multiple times to refine the output.
Benchmark: Additional information to guide what constitutes a good output.
Additional Guidelines: Specific rules or constraints to follow during the process.
Instructions: Here is a draft of my LinkedIn summary: "Experienced in marketing with a focus on digital strategies. Looking for new opportunities."
You will generate increasingly improved versions of this LinkedIn summary.
Recursion: Repeat the following 2 steps 5 times.
Step 1. Identify 1-3 points from the initial output that are missing or could be improved.
Step 2. Write a new, improved output of identical length which includes the missing or improved points.
Benchmark: Here is more information on what makes a good LinkedIn summary:
- Be Clear: The summary should succinctly describe your experience and goals.
- Be Specific: Include specific skills or achievements to stand out.
- Call to Action: Encourage the reader to connect or reach out to you.
Additional Guidelines: Keep the summary under 300 characters. Use professional language. Avoid buzzwords.
You will generate increasingly improved versions of this LinkedIn summary.
Recursion: Repeat the following 2 steps 5 times.
Step 1. Identify 1-3 points from the initial output that are missing or could be improved.
Step 2. Write a new, improved output of identical length which includes the missing or improved points.
Benchmark: Here is more information on what makes a good LinkedIn summary:
- Be Clear: The summary should succinctly describe your experience and goals.
- Be Specific: Include specific skills or achievements to stand out.
- Call to Action: Encourage the reader to connect or reach out to you.
Additional Guidelines: Keep the summary under 300 characters. Use professional language. Avoid buzzwords.
By following this framework, you can expect a LinkedIn summary that not only stands out but also closely aligns with professional standards. This ensures that the output will meet your specific needs, making the Chain of Destiny Framework an invaluable asset for tasks that require a blend of accuracy and creativity.
Prompt frameworks are essential tools for generating precise and effective outputs from AI and Natural Language Processing systems. Each framework has its unique features, making it more or less suitable for certain types of tasks. In this section, we will compare the frameworks discussed above, in terms of ease of use, effectiveness, and application areas.
RTF: This framework is straightforward and easy to use, as requires only three main components. Plus RTF prompts typically tend to be short and hence may help with latency and cost objectives associated with the AI models.
Chain of Thought: Also fairly easy to use, whether simply adding directive or steps to break down the task in your prompt. Prefer using directive over the task decomposition if it meets the output quality goals, as former is not only simplifies the prompt design but may also benefit on cost and latency dimensions.
RISEN: Requires a bit more planning as it involves multiple steps and constraints, making it moderately easy to use. In general, this results in higher latency and cost in comparison to RTF and CoT.
RODES: Similar to RISEN but includes examples, adding a layer of complexity and potentially latency / cost. However, can provide improved quality responses over output attained through RISEN approach.
Chain of Destiny: This is the most complex and typically tend to have higher latency and cost when compared with RTF and CoT. However, depending on the use case and prompt design, it could yield better latency and cost compare to RISEN or RODES.
Chain of Thought: Also fairly easy to use, whether simply adding directive or steps to break down the task in your prompt. Prefer using directive over the task decomposition if it meets the output quality goals, as former is not only simplifies the prompt design but may also benefit on cost and latency dimensions.
RISEN: Requires a bit more planning as it involves multiple steps and constraints, making it moderately easy to use. In general, this results in higher latency and cost in comparison to RTF and CoT.
RODES: Similar to RISEN but includes examples, adding a layer of complexity and potentially latency / cost. However, can provide improved quality responses over output attained through RISEN approach.
Chain of Destiny: This is the most complex and typically tend to have higher latency and cost when compared with RTF and CoT. However, depending on the use case and prompt design, it could yield better latency and cost compare to RISEN or RODES.
RTF: Highly effective for generating specific outputs but may lack depth for more complex tasks.
Chain of Thought: Exceptional for problem-solving and analytical tasks.
RISEN: Highly effective for complex tasks that require a structured approach.
RODES: Effective when you have good examples to guide the output.
Chain of Destiny: Extremely effective for refining and improving content through iterations.
Chain of Thought: Exceptional for problem-solving and analytical tasks.
RISEN: Highly effective for complex tasks that require a structured approach.
RODES: Effective when you have good examples to guide the output.
Chain of Destiny: Extremely effective for refining and improving content through iterations.
RTF: Best for quick queries, data extraction, and simple content generation.
Chain of Thought: Ideal for problem-solving, decision-making, and analytical tasks.
RISEN: Suited for research projects, blog posts, and other complex tasks requiring a structured approach.
RODES: Useful when you have examples to guide the output, such as copy-writing or content creation.
Chain of Destiny: Perfect for iterative content refinement, such as academic papers, marketing materials, or any content that benefits from multiple revisions.
Chain of Thought: Ideal for problem-solving, decision-making, and analytical tasks.
RISEN: Suited for research projects, blog posts, and other complex tasks requiring a structured approach.
RODES: Useful when you have examples to guide the output, such as copy-writing or content creation.
Chain of Destiny: Perfect for iterative content refinement, such as academic papers, marketing materials, or any content that benefits from multiple revisions.
While each framework has its pros and cons, the key is to choose the one that aligns best with your specific needs. For quick and specific outputs, RTF is your go-to framework. If you're tackling complex problems, Chain of Thought and RISEN are more suitable. RODES is excellent when you have good examples to guide the AI, and Chain of Destiny is the ultimate choice for tasks requiring iterative refinement. Understanding the strengths and weaknesses of each can help you make an informed decision for your specific use-case.