Summary

The "Flipped Interaction Pattern" is a method for leveraging AI to ask users targeted questions, guiding them towards achieving a specific goal without the user needing to research or process extensive information.

Abstract

The article discusses the "Flipped Interaction Pattern," a prompt engineering technique where an AI model leads the interaction by asking questions rather than the user. This approach is beneficial when users have some but not all the necessary information to complete a task. It allows users to be guided through a process, such as creating an itinerary, diet plan, or investment portfolio, by simply answering questions posed by the AI. The pattern is particularly useful when there is an overabundance of information, making the task seem daunting. By using key phrases, users can instruct the AI to ask questions that will lead to the desired outcome, such as creating a CV or making investment decisions. The article suggests that providing the AI with as much known information upfront can streamline the process and avoid unnecessary questions. Additionally, users can set boundaries on the number of questions or specify their involvement level in key decisions. The "Flipped Interaction Pattern" is presented as a convenient and efficient alternative to traditional research and decision-making methods.

Opinions

The author believes that the "Flipped Interaction Pattern" is a helpful solution for tasks that require gathering and synthesizing information, offering a more convenient alternative to DIY research.
The article suggests that this pattern is not only useful for situations with little information but also for scenarios with too much information, which can be overwhelming for the user.
It is implied that users can save time and effort by letting the AI model ask the necessary questions to achieve their goals, which is seen as a more efficient approach.
The author emphasizes the importance of providing context and known information to the AI to avoid irrelevant questions and to make the interaction more productive.
The article conveys that users can maintain control over the interaction by setting limits on the number of questions and by specifying their desired level of involvement in decision-making processes.
The author concludes that the "Flipped Interaction Pattern" is a powerful tool in prompt engineering, particularly when used with an understanding of its capabilities and limitations.

Prompt Engineering via Prompt Patterns — Flipped Interaction Pattern

The article is part of series: Prompt Engineering via Prompt Patterns

Many a times you are working on a task where you have some information to complete the task, but not all the information. And completing the task requires you to research the missing information and then piece all the information together. You are not always in a mood to go in so much detail, and come up with a great output. You would rather have someone ask you all the necessary questions and then create the output.

You can switch to video version of this article

Anyone who has travelled as a tourist know this well like finding out about all landmarks in destination and come up with an itinerary, creating a diet plan, creating a low risk stock portfolio, or picking the right insurance product from many options available in market. Rather than researching, comparing and identifying one that matches your requirement, you just want to be quizzed about your requirements and suggested about the product that best suits your goals.

In comes the helpful solution. The ‘flipped interaction pattern’.

The pattern is exactly what is being marketed. Your usual session with a large language model involves you asking questions or instructing the model to do something, and the model conforming. With this pattern, you let the model ask questions from you, and guide the entire process towards an end goal. This is extremely helpful if you don’t know or don’t want to know the missing information, so you delegate the task of filling in missing pieces to the expert.

Taking cue from our earlier example, rather than researching about 100 stocks and evaluating their risk profile to choose those you want to pick for your low risk portfolio with say an investment of 10000 dollars, you let the model ask you relevant questions to come up with the portfolio. Come to think of that, answering questions like what is your budget, what is your intended ROI and how much risk are you willing to take, and then boom a ready made portfolio of 20 stocks ready for your consumption is magnitudes better than identifying and shortlisting those 20 yourself.

Yes you could gather all the information by researching yourself or questioning the model, but isn’t flipped interaction amazingly more convenient than DIY? In this example, the situation was not like you did not have access to all the information to make the decision. Here there was too much information to process which essentially equates to no information since you don’t have bandwidth or drive to accomplish the task. Bottom-line, it is a really helpful pattern when you have little or no information about a task, or processing available information is not feasible for whatever reason, even if that reason is pure laziness and lethargy, or just convenience since you have read this article now so why even bother 😊

So the way to go about it is to instruct the model using key phrases like

I would like you to ask me questions to achieve my task

You should ask questions until this condition is met or to achieve this goal or better

Ask me questions about my goals regarding my task until you have enough information to suggest a solution for me.

Then optionally, you can scope the session with statements like

When you have enough information, show me the solution or response or whatever. You can begin by asking the first question. Or ask me questions one at a time, or 2 at a time etc.

Examples are worth a thousand contextual statements so here goes

I would like you to ask me questions to help me create a CV. You should ask questions until you have sufficient information about my current job, experience and skills . Ask me the first question.

Its best to tell the end goal as to why is the model asking you questions. Questions are means to an end. It can be a general quiz where no end goal is necessary, but if you have end goal in mind, convey it. Better context lead to better and to the point questions. Also you can tell how many questions to ask or how many questions to ask at a time. Ask the first question automatically tells the model it is to ask one question at a time. Also its best to scope the interaction like you can ask 15 questions max, or you can even instruct it to limit questions mid session once it has started. Use the magic sentence ‘until you have enough information’ to limit questions. You can exercise fair degree of control on the number, frequency and formatting/ordering of questions to keep the session fun, tolerable and productive.

Another way to limit questioning is to inject as much known information in advance as possible. Say if you are asking it to make a diet plan, and you are a vegan, let it know in advance rather than waiting for it to ask you if you prefer or avoid certain type of food. Same goes for dietary restrictions and any health issues you may have. While the intent of pattern is to avoid being in the driving seat, avoiding to be asked ‘obvious’ questions can lead to better utilization of the pattern. In short, you want to avoid being in driving seat for questions related to information you don’t have about the topic. In worse case, if trying to limit questions to your set max, ChatGPT might not even ask some basic questions and make guesses and assumptions.

Another related aspect is to make yourself known to the model i.e. your knowledge and education level etc. Say if you are starting a session about house prices in a region and you are a realtor yourself, making it known to ChatGPT would make lot of questions meant for ordinary folks go away. Any flipped interaction session would involve key decisions and you can let the model know in advance whether you would like to be involved via questions in every key decision, if you would like to do so than let ChatGPT know explicitly otherwise it might want to make educated guesses to keep session limited.

ChatGPT can add questions related to a particular answer that you specified in middle of interaction as it has more information which may lead to more questions to come up with a proper answer.

This concludes our discussion on flipped interaction pattern. I hope this article would help you apply this pattern to some problem you need to solve. If you liked the article, please clap/share and subscribe on the YouTube channel. Thank you and good bye.

Next article: Prompt Engineering via Prompt Patterns — Role/Persona Pattern