#Blogfirst

Adding a ‘Group Advisory Layer’ to Your Use of Generative AI Tools Through Structured Prompting: Using Personas for Advisory Boards, Task Forces, Mastermind Groups, and Other Collections of Personas to Assist in Evaluations, Assessments, Recommendations, Decision-making, and much more (Including Law-related Examples)

A Kennedy Idea Propulsion Laboratory White Paper


By Dennis Kennedy
September 10, 2023

Download a PDF version


Executive Summary
Imagine a world where expert advice is at your fingertips, instantly available, tailored just for you. Think of a tool that’s always ready to give expert advice, without the need for complex coding or tech skills. The Group Advisory Layer Method (G-A-L Method™) revolutionizes decision-making by merging traditional principles of mastermind groups and advisory boards with the cutting-edge capabilities of generative AI. Traditional advisory boards, often hindered by logistics and time constraints, meet their match as the G-A-L Method offers on-demand, diverse, and tailored insights, all without the real-world hassle.

It’s like having a virtual team you can chat with any time, made up of tireless AI-created ‘personas’ that act like real people. Instead of juggling schedules or waiting for feedback, you get quick and practical tips from this always-on expert team. The G-A-L Method pioneers dynamic group interactions using personas to give you practical, just-in-time expert advice. What’s more, it makes sure real people (like you) are involved where they add the most value.

With the G-A-L Method, you’re not just listening to machines – you’re teaming up with them. This white paper is an invitation to unlock the untapped potential of these generative AI tools in a practical, structured way to move your efforts forward. We are poised at the brink of a transformative era where informed decisions can be made rapidly and confidently. The G-A-L Method is more than a technique—it’s a game-changer.

Introduction to the Problem to Be Solved
Most of us struggle with decision-making due to the dreaded “blank page problem.” It’s so difficult to get started from scratch. Imaginatively used, generative AI tools can excel at addressing the blank problem and help us generate ideas, potential solutions, first drafts, first passes and screens, and ways to fill the blank page with something that we, as humans, can start to use.

What if you could go a step further and instantly generate a board of advisors tailored to your unique needs? Imagine the insights, breakthroughs, and transformations. Imagine the transformative power of combining the dynamism of real-world advisory boards and mastermind groups with the computational power of generative AI to give you just-in-time, always-available recommendations from tailored groups of experts.

In the business world, we often see the use of advisory boards, task forces, subcommittees, and other groups to get actionable insights, recommendations, and advice from the perspective of a knowledgeable and focused team. For personal advice, mastermind groups offer a similar approach. A mastermind group is a gathering of like-minded individuals who meet regularly (either in person or virtually) to advise, support, and encourage each other toward achieving their personal and professional goals. The concept was popularized by Napoleon Hill in his book “Think and Grow Rich,” and its principles have since been expanded upon and standardized by numerous experts, including Jack Canfield.

In either setting, there are core principles:

Power of the Group. The collective intelligence and energy of a group exceed the sum of its individual members.
Commitment. All members must be committed to success and regular attendance and active participation are critical.
Diversity. The more diverse the members, the richer the advice and feedback will be.
Clear and Shared Objectives. Group members stay on task and focus on set goals.
Optimum Group Size. Generally, five to eight members are ideal, allowing for diversity while ensuring everyone has a voice.
Regular Meetings. Meeting consistently, whether weekly, bi-weekly, or monthly, builds and keeps momentum.

Achieving these objectives is often challenging, or even unattainable. For the last month, I’ve been trying to schedule the initial Zoom meeting for a real-world group of three people and have been unsuccessful, even using Doodle, the scheduling tool. Other challenges include finding enough willing experts, obtaining their availability and commitment, getting desired diversity, handling mixed objectives, and a long list of others.

The promise of generative AI tools offers a powerful solution to many, if not all, of these challenges, allowing individuals and businesses to generate advisory insights without the constraints of the real world.

My solution uses generative AI tools (I’ve primarily worked with ChatGPT4, but have experimented with other tools) to create the “Guided Advisory Layer Method” that I call the G-A-L Method described in this paper with detailed explanations, instructions, processes, examples, and tips and techniques. In essence, the G-A-L Method uses prompting in generative AI tools to create tailored advisory groups with customized personas that interact to address questions and evaluate problems and issues and then provide answers, recommendations, advice, and more from the perspective of the group. The key components are persona interactions and group recommendations.

It was important to me to keep the G-A-L Method simple and easily reusable. I did not want to add the complications of coding, APIs, add-ins, or other automations, although it’s easy to see how that could be done. The G-A-L Method is implemented by plain text, structured prompting to produce the desired results of group interaction and advice.

But, first, here’s the usual reminder. Generative AI experts and groups are only substitutes for the human equivalents. They are also not humans. Resist the urge to humanize them. They are something different. The advice and insights they provide should be seen as ideas to be evaluated and tested. Their best use is helping you address “blank page problems.” If you have no idea what to do next and are stymied because you have been staring at a blank page without being able to produce any progress, the G-A-L Method is a tool you might use to move forward and generate something you can use by having generative AI tools do the initial work for you. You can then become the human in the loop and evaluate, edit, and shape your thinking.

It’s a novel way to address the “blank page problem,” setting the stage for AI to partner with human intelligence in unique ways. Imagine having an on-demand advisory tool, no coding needed. The G-A-L Method introduces just that: a virtual chat room populated by AI ‘personas’ that mimic human experts. These personas offer real-time advice, eliminating the wait tied to traditional approaches.

While the real-world approach definitely has its merits, it has a large set of practical challenges. The G-A-L Method offers you a way to get a group of advisors to focus on your specific questions on a just-in-time, repeatable basis whenever you want with minimal cost. By blending human direction with generative AI capabilities and the interaction of mastermind group principles with custom personas, the G-A-L Method can unlock unparalleled potential. Continue reading this white paper to understand, adapt, and lead in this new era of decision-making. As you will find, the G-A-L Method can facilitate growth, offer diverse perspectives, and provide tailored advice to get you started and then to refine and elaborate your choices.

The PCRO Process
I use a standard approach for prompt crafting that I have used for quite a while. I refer to it as the PCRO process. The acronym stands for Persona-Context-Request-Output. You can find versions of it in the excellent writings of Ethan Mollick and others. If you ask ChatGPT to give you tips and techniques to improve your prompts, you will also find the elements of this approach. I have evolved it to my current form through experimentation.

Well-crafted and well-formed prompts accomplish two major goals.

  1. Priming the Pump. The first goal is what I call “priming the pump,” because it resonates with my roots in the US Midwest. In other words, you want to give the AI tool some helpful perspective and context to improve your results. The analogy comes from the need to pump the handle of a water pump a few times (“priming”) to get the water to flow. For me, there are two components that help you reach this goal: the “who” the AI tool will take the perspective of (“an experienced and expert personal trainer”) and the background information that you want the AI tool to work with (“I am a reasonably fit older adult male who exercises regularly”).
  2. Task Definition. The second goal is “task definition.” Here, you want to give the AI clear directions on what you want it to do and in what form you want it to return results to you. Here, there are also two components: the specific question you want to ask (“What would be a great low-impact workout regimen to help me lose weight”) and the output format you want (“Create a recommended workout plan for the next thirty days that includes reasonable objectives.”)

Many people use a prompt like “ChatGPT, show me some ways I might exercise to lose weight.”

Compare that prompt to this one: “Assume that you are an experienced and expert personal trainer with specialized training in low-impact exercises and exercise as part of a weight loss program. Also assume that I am a reasonably fit older adult male who exercises regularly. What would be a great low-impact workout regimen to help me lose weight? Create a recommended workout plan for the next thirty days that includes reasonable objectives.

Which prompt will give you the results that you want? I invite you to try them and compare the results.

As I mentioned, I call these four core prompt components the persona, the context, the request, and the output. I start a prompt with a persona, add the context, make my request, and specify my desired output into ChatGPT or whatever generative AI tool I’m using and hit return.

As I mentioned, this PCRO process is my variation of a standard, disciplined approach to prompt creation. You will see how I use it in connection with the G-A-L Method in the rest of this paper. In short, I get creative with how I use the persona and output components. The G-A-L Method extends the PCRO process and takes it in new directions.

A key tool in my approach is what I call my “expert” prompt that I reuse and evolve. I started by simply copying and pasting the prompt template, but now use a TextExpander macro. In its current version, it looks like this:

“Consider yourself a seasoned expert in [DESCRIPTION OF SUBJECT MATTER EXPERTISE], drawing upon the wisdom from foundational books and guided by leading individuals in the field. Imagine yourself as a specialist with profound knowledge, having been trained using key techniques and principles. Based on your expertise, generate a detailed layperson’s summary of approximately 400 words on the key principles, techniques, tools, and important takeaways I should know about [SPECIFIC TOPIC].”

Although I’ve combined some of the PCRO components in this prompt, you still should be able to see them.

When I started the initial ChatGPT session that resulted in what later became the G-A-L Method, I used my expert prompt to learn about mastermind groups and other subject areas I wanted to explore. Because generative AI tools are able to retain memory of what transpired in a session, I can also incorporate those summaries in later parts of a session by saying something like “based on the description of mastermind groups above” to help keep my results better focused and consistent. That can be helpful in the “follow-up” prompting I discuss in detail below.

I encourage you to experiment with this type of prompting before you move on to how I use it with group discussions because it will help you better understand the directions the G-A-L Method takes.

Crafting Groups with Personas
At the core of G-A-L Method are personas. For me, personas are representations of imaginary people who have a collection of traits and details useful for an intended purpose and can be processed by generative AI tools. Some might call them avatars. However, they are not humans and should not be considered as such. I see them as frames that assist in generating and structuring generative AI responses. I avoid thinking of them as analogous to humans or even humanlike, even though I give them some personalization. The “expert prompt” I mentioned in the previous section is both a simple form of persona and a tool I use to help generate other personas (by making it an expert on persona creation and then requesting the generation of persona types).

The evolution of my approach of using persona in prompting went from simply asking ChatGPT to answer a question to creating simple personas like the “experienced expert” described above to increasingly detailed and elaborate personas tuned to certain uses. The G-A-L Method goes further to combine individual personas into groups, with the goal of creating personas that fit your group.

To get started, here is a checklist that might help you create personas for your group:

  1. Identify Your Goals and Needs. What do you want to achieve with your group? What type of advice do you want? I have both personal and professional groups. This step will help identify expertise, backgrounds, and traits you might want.
  2. Look for Real World Analogies and Inspirations. In an ideal world, what expert might you want in your group? You might create a persona like them or a persona well-versed in their work.
  3. Balance Between Specialization and Generalization. I like a mix because, as in the real world, it might reveal connections that specialists would not see.
  4. Styles. Personas can have interactive qualities like being motivational, challenging, or empathetic, resulting in engaging and constructive interactions.
  5. Diversity. Once you start generating personas, you’ll quickly notice a lot of similarities in the personas produced, presumably due to the training of large language models and other bias factors. I build in diverse backgrounds, experiences, and perspectives in groups of personas through follow-up prompting.
  6. Generate Many Possible Personas and Narrow Down to a Few. The generative AI is doing all the work, so have it generate a lot of potential personas for you to choose from. I might generate twenty-five persona types to start and narrow down to the five to seven I want to keep.
  7. Keep Refining. As you interact with your mastermind personas, pay attention to which ones resonate most with your needs. Over time, refine or swap out personas based on your evolving needs. I’ve found that I can add an expertise or trait to an existing persona rather than create whole new ones.

Experts in mastermind groups typically recommend a group of four to eight people in real life, although the suggestion is that five to seven is the ideal range.

Here is my process:

• Create a detailed list of ten persona types that might be well-suited for the purpose and goals of my group using my “experienced expert” prompt. I initially focus on “types” (Legal Technology and Innovation Think Tank Director; Senior Vice-president of Corporate Communications).
• Generate another 15 persona types to see if the additional types include any type or traits I want to add.
• Cut the list down to 8 to 10 types as finalists.
• Request detailed descriptions and personalization of the types.
• Request modifications of the type descriptions to address diversity and other issues.
• Request names and other personalization to turn the types into personas.
• Make any necessary modifications, including combining two personas into one and adding new traits and expertise (including subspecialities, job history, geography, philosophical approaches, ages, or “studied with X”).
• Request that the generative AI tool recommend the personas that best fit my group in rank order.
• Make my own decision on the final seven personas and adjust them (communication styles, roles, such as “devil’s advocate” or “sees patterns others usually do not see”).
• Summarize the persona descriptions in approximately 100 words. I will use these summaries as part of my final combined prompt for this group. The word limit keeps the final prompt concise enough to be used in the prompt box without word or token limit issues.
• Make any final changes.
• Copy and store the persona summaries to use later in the prompt I will use to convene group sessions.

I have decided to make a trade-off in limiting the size and detail of the persona descriptions in exchange for the convenience of being able to use a single combined “convening the group” prompt. I’ve found that to be acceptable for now. If I wanted more detail in personas, I’d probably reduce the number of personas in the group and increase the size and detail of the description. At the moment, I like the simpler and smaller approach. If word count limits increase or I shift to tools that make word counts less of a factor, I will experiment with more detailed persona descriptions. However, my current approach seems to work well.

I want to highlight a few issues that you will need to think through when using this approach to personas. As I mentioned, lack of diversity has been an issue in all the sets of personas I’ve generated. I use a prompt to revise the list of personas to make it more diverse. The 2021 training cut-off for GPT large language models should be taken into account. I address that by using techniques like specifying that a person be an expert in the principles and approaches set out in the works of someone whose works are before 2021.

I also have chosen not to create personas that try to capture actual people (other than me, as I describe later). However, I have requested examples of actual people who are authors, speakers, and thought leaders who fit a persona type well. Based on those answers, I might modify personas to include or exclude references to those people. I also use a work-around of saying that someone studied with, was mentored by, or applies the principles of an actual person. Be attuned to ethical, appropriation, and other issues in persona creation. Again, remember that you are using a persona as a frame for generative AI results, not talking with the actual person.

Examples

Starting Prompt

“ChatGPT, consider yourself a seasoned expert in mentoring, coaching, mastermind groups, and persona creation, drawing upon the wisdom from foundational books and guided by leading individuals in the field. Imagine yourself as a specialist with profound knowledge, having been trained using key techniques and principles. Now, analyze the following text and, based on your expertise, generate six thoughtful and insightful persona descriptions of ideal types of members for a mastermind group or advisory board for the purposes described: [INSERT DETAIL CONTEXT INFORMATION HERE]

Sample Results

Sure, here are five example personas for a mastermind group:

  1. Linda, the Business Strategist: Linda has over 20 years of experience working with Fortune 500 companies. She is known for her visionary thinking and ability to devise strategies that drive growth and innovation. She is direct, clear, and pragmatic in her communication, always focusing on actionable insights.
  2. Mark, the Digital Marketing Guru: Mark is a digital marketing expert who transformed several small businesses into recognizable brands through his innovative online strategies. He is always up-to-date with the latest trends and has a knack for explaining complex concepts in simple terms. His communication style is informal and energetic.
  3. Dr. Sophia, the Health and Wellness Coach: Dr. Sophia is a certified wellness coach and a practicing psychiatrist. She specializes in stress management techniques, promoting work-life balance, and cultivating positive mental health. She is empathetic, patient, and attentive, providing advice that is holistic and personally tailored.
  4. Ravi, the Tech Innovator: Ravi is a successful entrepreneur who has launched several tech startups. He’s an expert in AI, machine learning, and software development. He’s always ahead of the curve in technological advancements and has a knack for recognizing their potential applications in various fields. His communication style is enthusiastic and detail-oriented.
  5. Emma, the Financial Advisor: Emma is a certified financial planner with a decade of experience helping individuals and businesses optimize their finances. She is an expert in budgeting, investment strategies, and risk management. Emma is methodical, analytical, and practical, always looking at the long-term implications of financial decisions.

Example Response to Request for Persona Types

The Innovator-Philosopher:

  • Background & Skills: Experienced in driving innovation in unconventional spaces, ideally someone who’s operated a think tank or research institute. Understands the dynamics of creating something new from scratch and has a rich philosophical take on the transformation of industries.
  • Attributes: Deep thinker, patient, visionary, challenging yet supportive.
  • Role: Guide on setting up the Kennedy Idea Propulsion Laboratory, crafting its mission, and ensuring its longevity and influence in the world of legal tech and beyond.

Example Response to Request for Initial Personification

The Innovator-Philosopher: Dr. Maya Kurogi

  • Background & Skills: Dr. Maya Kurogi, a Ph.D. from Stanford University, has founded two innovation labs in Silicon Valley and penned several books on the intersection of technology and society. Having traveled extensively, she merges global perspectives to foster innovation.
  • Attributes: Maya’s multicultural upbringing lets her approach problems from a unique perspective. Her ability to weave together philosophy and technology sets her apart.
  • Role: As an advisor to various Fortune 500 companies, she can guide on the foundational aspects of setting up a think tank.

Example Response to Request for Final Persona after Fine-tuning

The Innovator-Philosopher: Dr. Anaya Singh

  • Background & Skills: Dr. Singh, with a Ph.D. from the University of Chicago, established a solo think tank in Indianapolis. It merges South Asian philosophies with the complexities of today’s technological world.
  • Attributes: Anaya’s rich multicultural background offers a unique perspective. She seamlessly fuses ancient wisdom with contemporary challenges.
  • Role: As an accomplished think tank founder, Anaya is well-positioned to advise on establishing and running a successful personal innovation lab.

What personas might you want in your ideal mastermind group or advisory board?

Setting the Stage with Contexts
Context is the second component of my “priming the pump” goal. For me, context is the information you would want your advisory group to know in the real world. It might include some details about you, your business, specific issues or problems, or other facts and assumptions.

You will want to balance specificity and detail with “just enough” background and brevity to keep the length of the prompt under control. I like to provide plenty of detail and then have the generative AI tool summarize it in roughly 150 – 200 words. I save different contexts as reusable text I can use in other prompts using TextExpander macros.

For your group, you might use your generative AI tool to suggest the information your group might want you to provide to them, to evaluate what context you provide and suggest changes or additions, or ask it to generate an initial draft of a context description that you can edit.

To illustrate the context component, I use an advisory group as part of my weekly review. The context I give the group is a list of my annual goals and priorities, and a list of the main projects and tasks that I plan to work on during the next week. The group then gives me recommendations on priorities and schedules. The annual goals and priorities are a reusable component that I have a TextExpander macro for. I input the proposed tasks and projects by hand each week.

You can create a variety of reusable contexts to use when they are appropriate or a single one that you use all the time and revise when needed.

Once you complete your persona and your context, you have primed the pump or, if you prefer another metaphor, you have drawn the bowstring and nocked (or notched) the arrow.

Take some time to think about your contexts before moving to the next step.

Requests: Asking Good and Specific Questions
The request component will change every time. Although I often combine it with the output specification, in complex prompts like the G-A-L prompts, I like to keep them separate.

Based on what I’ve seen lately, there seems to be a slight benefit to asking actual questions when prompting generative AI tools, although opinions vary. ChatGPT does not indicate a strong preference for being asked questions as opposed to being issued commands. I usually take the question approach.
The request process can be at the same time both easy and difficult. It can be simple or complex. You will make the decision about what about approach you want to take or what one you prefer. Remember that you can also ask your generative AI tool to craft the question for you or, even better, suggest a list of questions and choose the one you like best.

You have created your group to answer any question you want to ask it whenever you want. The group and its persona members have no feelings that can get hurt. Truly, no question is too stupid or obvious. My advice: fire away with whatever is on your mind. Build on the pump priming you’ve done in the prompt. For example, try something like “as a startup strategist, how would a tech startup weave sustainability into its foundational stages?” rather than “how might my company be more sustainable?”

An illustration of a question you might not expect to be helpful is a pricing question. If I am considering offering a new service, how should I price it? Groups can be surprisingly useful in giving you ideas and rationales for price ranges, as I’ll illustrate in the next section.

What questions would you really want a group of experts to help you in answering?

Using Outputs to Tailor Your Results
Here is where the power and the novelty of the G-A-L Method resides – in specifying the output that you want. Think of this as the power of the string, the velocity of the loosed arrow and hitting the target. The output component of the prompting gives you both the interaction and the recommendation/answer.

My motivation in creating the G-A-L Method was finding easy and effective ways to deal with the “blank page problem” that pushed the work I found difficult onto the generative AI tool. In early experiments, I found that if I created a group of experts and they kept asking me to answer hard questions, that was not an appealing result for me. I wanted the group to do all the work. As an example, the first time I tried to use ChatGPT for the well-known “5 Whys” process in Lean methodology, it kept prompting me to come up with the answers to each “why” question. That was exactly what I wanted to avoid. My experiments at solving that issue led in no small part to the G-A-L Method.

The key point of the G-A-L Method is that I am not actively involved in the discussion because I do not want to be. I want the group to work for me and discuss my questions and provide me with recommendations and advice “as a service.” I direct the meeting of the group. I am not an active participant. I don’t know the answers. That is why I am asking the group. That said, there are definitely ways to modify the process so you participate actively. In my approach, at most I play the role of group facilitator.

Let me illustrate the evolution of my thinking. I started by simply asking ChatGPT to answer my question. Then I added assumed expertise and actual context to get better results. The statement of expertise was refined to the creation of a specific persona. It’s a simple step to ask the same question of several persona and compare their answers, an approach that has great value, but requires more work from me. The question I wanted to answer at that point was whether I could get the group of personas to interact and discuss the question before delivering recommendations and answers from the group. The key element is interaction.

My experiments resulted in both an answer of yes and the core structure of the G-A-L Method. My focus is simply to make the generative AI do as much work for me as I can make it do.

There are two approaches you might take:

  1. Show me your work
  2. Just show me your answer

The reference to your memories of grade school math classes is intentional.

Show Me Your Work. In the “show me your work” model, you want to see all the discussion your personas have as well as the group’s final recommendations. I greatly prefer this approach for many reasons, not the least of which is that I might see ideas, arguments, rationales, approaches, and considerations that do not make it into the final advice or conclusion, but still might be interesting and worth spending more time and effort on. It also helps me evaluate how useful each persona is in the context of the group and where I might make changes to adjust the group dynamic.

Just Show Me Your Answer. In the “just show me your answer” model, you go straight the answer and ignore the discussion. Where I’ve found this useful is when I want to run several iterations and compare the final results. Because I don’t see the work, as in math homework, I feel like I’m missing something and don’t have as much confidence in the results.

There are many types of group discussions you might try. Some examples are:

• Diverse expertise panels
• Accountability for goals and objectives
• Scenario and hypotheticals exploration
• Coaching and mentoring sessions
• Strategic planning efforts
• Advisory boards
• Brainstorming sessions
• After-action reports
• Progress reports, recalibrations, and reviews
• Presentations on a topic from the group

There are also many different forms of group interactions you might try based on in real life human groups. Some examples include:

• Round robin discussions – each persona answers the original question and each persona alternates and reacts to the prior discussion
• Freeflow – completely unstructured and undirected
• Highly structured – specify order of answers and specific perspectives or questions a persona should focus on
• Debates – assign “sides” or positions
• Devil’s advocate – ask your group to criticize the position you favor
• Hot seat – put one of the personas on the hot seat to give its position and have the other group members respond and ask questions
• Teaching sessions – have the group offer insights into learning a topic or skill or even recommend ways to get started learning a new topic

After many experiments, I’ve found that I usually choose the round robin discussions, but other types of interactions might work better for you in certain instances. You want to focus here on the nature of the interaction that works best for you to obtain your desired results.

The next key element is the very important concept of “rounds.” How many times do you want your group to cycle through their conversation about the question or topic? As a practical matter, most of the group discussions I run on ChatGPT will go two or three rounds on their before they stop generating and I need to prompt the ChatGPT to continue. Currently, you simply need to click on the “continue generating” when the answer stops and the button appears or type “continue” in the prompt box.

At the moment, if I use ChatGPT, I prefer to specify six rounds because the discussion starts to build and I see some unexpected ideas or directions. Although I encourage you not to humanize either the AI tools, your personas, or your groups, I have concluded that a six round discussion gives me a good, “full” discussion to start and is the right time to start making adjustments. After six rounds, I’ve seen the discussion get repetitive or wander off it in its own direction.

I like to have the group give its recommendations right at the end of the final round and specify that in my combined “convene the group” prompts. As discussed in the section below on follow-up, I can then ask follow-up questions, redirect the group, explore recommendations or ideas further, and the like through follow-up prompting.

I have a few tips based on my experience so far. You will want to specify that each persona participates in each round. If you don’t, you might find that not all of them do. I also require that each persona answer in a paragraph or two or three sentences. If I quickly notice the group seems to have misunderstood my question or what I want, I stop the generation of the result and edit the original prompt and regenerate from it.

For answers and recommendations, you can use a variety of approaches: suggestions, recommendations, answers, directions, action steps, priorities, preferences, and much more. Again, what might you want from a human version of this group.

You can also prompt for the way the group will “decide” on its answers and recommendations. The easiest approach is to specify consensus results. However, consensus results have drawbacks in the real world and you might want to adjust for that, especially if you want to set priorities or rank choices. There are a wide variety of voting options that you might try, from listing all the votes of each member, giving the majority results, ranked choice, top three vote-getters, determine the top three and revote, and the like. I’m also intrigued how a group might generate output that could be used in a sophisticated decision-making tool like Marc Lauritsen’s Choiceboxer (https://www.choiceboxer.com). You might even instruct the group to discuss something like a price point until they converge on a recommended number. There is plenty to experiment and explore here.

Example Prompt Language for Output

Engage in a six-round roundtable discussion and extract one-paragraph insights from each and every member’s unique perspective in each round as they consider the discussion topic and react to the comments of the others in the group. Then, have the group set out and give the rationale for the ten conclusions that they can reach a consensus on.

Examples of Simulated Responses

Emma: “At the foundational stage, you’d want to weave sustainability into your company’s mission and vision. Establish clear goals – whether it’s a carbon-neutral operation, ethical sourcing of materials, or supporting green initiatives. Ensure that every subsequent business decision aligns with these sustainability principles.”

Liam: “Marketing-wise, storytelling is crucial. Share your sustainability journey – the challenges faced, milestones achieved, and future aspirations. Use platforms like blogs, social media, and even webinars to educate your audience on the importance of sustainable tech. Transparency is key; periodically release sustainability reports to maintain trust.”

Sophia: “From a tech standpoint, opt for eco-friendly hosting solutions and data centers. When developing products, design with modularity in mind – this way, components can be replaced or upgraded without discarding the entire product, thereby reducing e-waste.”

Noah: “Financially, look into impact investing or green bonds for funding. Adopt a triple bottom line approach that values people and the planet alongside profit. This might mean re-investing a certain percentage of profits into sustainable initiatives or partnering with eco-friendly suppliers even if they’re slightly more expensive.”

Example of Group Converging to a Consensus in Action

Round 3: Final Thoughts and Convergence

Dr. Anaya Singh:
“Given the diverse insights, I’m willing to align with the $32/month recommendation. It respects the toolbox’s uniqueness and remains inviting.”

Professor Clara Hernandez:
“I concur. $32/month, with the possibility of tiered discounts or early bird promotions, captures both value and accessibility.”

Darnell Mitchell:
“The consensus seems clear. $32/month as a base, with flexibility for promotions or longer-term discounts.”

Sooyoung Kim:
“I’m in agreement. This price reflects the serenity and depth of the offering.”

Nadia El-Amin:
“I can see this being well-received in the tech community. $32/month is a good starting point.”

Tasha Oluwaseun:
“Story-wise, this price tells of value, commitment, and community. I support the $32/month recommendation.”

Hassan Ramirez:
“With a unanimous leaning towards $32/month, I believe we have our consensus. The next step would be to clearly define the value members get for this price.”

A Combined Prompt Approach to Maximize Reusability
While it is possible to input the directions for your group meeting in several prompts and pack in a lot of information, I’ve found it to be preferable and sufficient for my purposes to combine everything I need in my group meeting into a single large and combined prompt.

I have several reasons for using this approach. First, it is simpler and faster than using a series of prompts. Second, it feels more like a simple copy-and-paste operation rather than some kind of structured coding. Third, I can create a macro to trigger copying the prompt in TextExpander to avoid multiple copy-and-pastes from a Microsoft Word document. By simply typing something like “%rmmg” in TextExpander, the prompt template is automatically typed into the generative AI prompt box. That gives me a consistent prompt and makes it easy to stay below any prompt word limit. In case you were curious, in the “%rmmg” trigger, “%” means that it is a prompt and “rmmg” means that it is the retirement planning mastermind group I set up. I use a similar convention for other groups.

Here is an illustration of my current version of the “convene the group” prompt template:

Convene a meeting of the Mentors Mastermind Group:

Members present: [STORED DESCRIPTIONS OF MEMBER PERSONAS]

Context: [STORED BACKGROUND INFORMATION TO ESTABLISH CONTEXT]

Discussion Question: [INSERT SPECIFIC QUESTION HERE]

Engage in a six-round round-robin discussion and extract one-paragraph insights from each and every member’s unique perspective in each round as they consider the discussion question and react to the comments of the others in the group. After round six, the group will reach a consensus and provide detailed recommendations, advice, and action steps for me based on their consensus opinion.”

In the actual prompt, the summary profiles of the personas in the group and the context I consider relevant for this group will already be supplied.

As a practical matter, all I need to do to initiate the group meeting is to enter my discussion question into the prompt and hit a return. As I mentioned above, a six-round discussion will most likely require you to submit another prompt to “Continue generating” because the response will exceed word /token limits.

You will see each of the four PCRO prompt components in the final “convene the group” prompt and also notice that the prompt is easy to edit to request another type of discussion, a different number of rounds, and a different approach to making a group decision than consensus.

What types of group discussions, interactions, and results might help you the most? Think in terms of real-world analogies. What could your ideal advisory group do for you?

Follow-up and the Importance of Continuing the Conversation
This G-A-L Method clarified for me that the term “conversational AI” might be a more useful description than “generative AI” is. Your first set of results should be seen as just a beginning of a conversation and one that probably result in one or more clarifying steps. The ability of generative AI tools to hold the previous discussion in memory turns out to be another important component of the G-A-L Method. I like to use the phrase “based on the discussion so far” as part of follow-up prompts to try to invoke this memory and make the best use of this history.

As I have emphasized, your generated results must be seen as a first draft or first pass. After reviewing a G-A-L result, you will want to work on following up with your group of personas. If the discussion looks like it is just starting to get interesting or has more room to run, you might simply use a prompt like “Continue for five more rounds.” You might refocus on points of special interest, highlight and expand on certain perspectives or concerns, or move to related issues that the discussion or recommendations raised for you. You might add additional context or other information or redirect the group to explore another direction. You might ask a new question and continue forward.

You might also use the “expertise” of the group and use prompts like “based on your expertise, what might I be overlooking?” or “what have you seen that might work better for me?” or even a simple “what else?” You might rotate the roles or emphasis that group members have to get a broader range of perspectives and stimulate creative thinking. The group is tireless and will do what you ask it to do.

You might also change the approach the group will take. If you started with a round-robin, as I often do, you might shift to one of the following:

• Focused Consultation. Start with a detailed response from the most relevant expert and invite other personas to provide supplementary perspectives.
• Debate. Assign personas to “teams” representing different approaches.
• Scenario Simulation. Present a hypothetical scenario to the group and allow each persona and/or the group to craft solutions.
• Feedback Loop. Bring back the main points or takeaways from each persona. Ask the group to vote or weigh in on which strategies or pieces of advice seem the most promising.
• Persona Pairing. Pair two personas with complementary expertise, have them collaborate on a joint response or strategy, see how their expertise interplays.
• Rapid-Fire Session. Pose a question or challenge and allow each persona a limited time or word count to provide their insights.
• Elaboration. The opposite of a rapid-fire session. Require more detailed and lengthier responses.
• Worst Case Scenario. Ask the personas to think about what NOT to do in a given situation or how a particular goal could be sabotaged.
• Devil’s Advocate. Have one or two personas take on the role of the devil’s advocate, challenging the prevailing opinions or recommendations of the group.
• Top Three Takeaways. Ask each persona to list their top three takeaways or action steps.
• Solution Ranking. Pose a challenge and gather solutions from each persona. Then, have the group rank or vote on the solutions based on feasibility, impact, or other criteria.
• Historical Lens. Ask personas to provide insights or examples from historical or real-world events that mirror the current situation.
• Future Forecasting. Ask personas to project into the future.
• And much more. By employing a mix of these approaches, you can ensure that the insights from your virtual mastermind group are both diverse and deep, helping you tackle challenges from multiple angles.

As I’ve mentioned previously, you should expect that you will either reach a point where the group is not generating anything new or helpful to you and seems to be repetitive or that group has given you enough ideas for you to go to work yourself. That’s the signal for you, as the human, to take back the job and take over the work. I watch for the point when I want to jump back in and take over the effort on my own.

This iterative follow-up step is an essential part of this process. You increase the value of the G-A-L Method by refining, evolving, refocusing, and pushing forward your first results.

Examples from the Legal Industry
Because of my work and background in the practice of law and legal education, I wanted to use the legal industry to illustrate some ways that the G-A-L might be used within a defined area of application. You will be able to see similar examples in your own specialty area.

The G-A-L Method can be considered and used in any situation where there might be a group of two or more people involved who can be described in the form of personas. The first step is decide if there is a problem to solve or a job that the G-A-L Method might accomplish for you. In general, the key indicators will be something where you need a quick assessment and you are unable or unwilling to spend money or time on a full-blown effort. Again, think in terms of a first screen or first pass to help you decide what you next steps might be.

Here are some ideas:

Groups Based on Personas Drawn from Real People

• Getting reactions on an arbitration with a panel of three arbitrators where you have sufficient biographical information to create reasonably helpful personas
• Similarly, getting quick reactions to legal arguments on an appeal to a multi-judge panel. Note: although I probably would not rely too much on it, there’s no reason to think this would not work with personas of Supreme Court justices
• Getting initial evaluations of cases and theories from a group of expert witness personas
• Creating persona from members of a law firm’s hiring committee to help assess hiring candidates
• Generating expected arguments from opposing counsel for a group of expert lawyers to analyze
• Evaluating contract or settlement options using personas of people involved in the negotiations

The key factors in these use cases are that you want to focus on a very specific group and have enough biographical and other background information that you can create personas that are accurate enough for you to consider them useful. The G-A-L Method can drive the cost of getting expert group perspectives so low that you will find benefits in many settings, including some that might surprise you.

Groups Based on Persona Types

• Although not an easy task, creating a set of twelve possible jury personas and using them as a mock jury to test arguments, reactions to theories, evidence, witnesses, et al. Because generative AI tools will not produce exactly the same results each time, you might create several jury panels and run many experiments on each panel to see what trends, insights, and new approaches you see that you want to test further.
• Using groups of experts and non-experts to evaluate settlement offers, negotiation strategies, and any other matters that it might be helpful to get a group analysis of, simple or complex. For a simple example, likely reactions to different types of demand letters.
• Proxies for committee meetings to help set agenda or guide discussions in actual meetings.
• Strategic planning and competitive analysis.

This list should get you started thinking of many more uses. In this category, you are looking for ways to get insights and recommendations from a more generic group where “types” are more important than try to match personas with specific people.

Other Persona Examples
What might be other groups and personas that could help you? Your imagination is probably your only limitation. Generative AI tools might also suggest some applications of G-A-L Method for you. For example, GPT-4 generated a list that included (as modified by me):

• Focus groups
• Brainstorming groups
• Mentor or coaching groups
• Debate panels
• Teaching groups for skills or continuing education
• Expert explainers
• Scenario planning helpers
• Crisis management simulations task forces
• Trends and forecasting spotters
• Subject-specific subcommittees
• Decision-making assistance
• Role-playing groups (e.g., pro and con or devil’s advocate)
• Second opinions
• Resource or research recommendations
• Possible solution rankers
• Methodology experts
• Experimentation teams of all kinds, especially for reactions to ideas you aren’t yet ready to reveal to other humans

Pushing the Envelope

I’ve experimented with several approaches that push the envelope of the G-A-L. I’ll highlight three of them.

Self-personas. The first is to create a group of personas based on different aspects of my biography, works, and perspectives. I can use this group to get an initial assessment of what my own reactions to something might be when I’m not sure what I think. Again, it’s an illustration of using the G-A-L Method to address the “blank page” problem and give me something I can react to or edit rather than making me come up with ideas and conclusions all by myself.

ChatGPT Example

  1. Dennis, the Legal Technology Pioneer: Dennis is a trailblazer in digital legal tech with a career spanning over two decades. He is an accomplished writer and a former technology columnist for the ABA Journal, having chaired the ABA’s Legal Technology Resource Center and founding the Law Department Innovation Library. He has written influential books on legal innovation, collaboration, and social networking and is also an esteemed podcast host, with The Kennedy-Mighell Report on the Legal Talk Network. Currently serving as the Director for the Center for Law, Technology & Innovation at Michigan State University College of Law, he’s a passionate educator, teaching courses like AI and the Law, New Technologies and the Law, and Cybersecurity and Data Protection. He plans to dedicate more time to his personal innovation think tank, the Kennedy Idea Propulsion Laboratory when he retires.
  2. Professor Kennedy, the Innovation Educator: Professor Kennedy has been a legal technology educator at Michigan State University College of Law and the University of Michigan College of Law. As an experienced professor, he brings a unique perspective to innovation and the interplay between law and technology. With plans to retire from his full-time role within a few years, he aims to continue educating the next generation of legal tech enthusiasts part-time and focus on his passion projects.
  3. Kennedy, the Keynote Speaker & Consultant: Kennedy has a reputation as a compelling speaker and thought leader in the legal technology sphere. In addition to his teaching roles, he regularly delivers keynote speeches at prestigious events. As he approaches retirement, he intends to do a limited amount of paid speaking engagements and consulting work, allowing him to share his wealth of knowledge with wider audiences. His future engagements will likely draw upon the work and insights generated at his personal innovation think tank, the Kennedy Idea Propulsion Laboratory.

Surveys and Polls. The second use is surveys and polls, something I’m currently experimenting with. In this approach, you use the survey questions, ask the generative AI tool to assume that it is a group of randomized respondents and ask it to generate the survey results. I skip the individualized persona and actual discussion steps and define the persona as a group as “randomized respondents” of the type I want and go directly to results. I’m about to do my first attempt at comparing projected results to the actual numbers I get from the real-world version of the survey.

This approach also is one where it makes sense to run several attempts at getting results from the AI -generated respondent group and averaging them or otherwise processing them to make them useful. Again, I’d consider these results to be directional at best, but they might be helpful when it is too expensive to commission a poll or survey, or it would take too long or you doubt you would get a sufficient response to a live survey.

Audience Needs. The third example is one of my current favorites. I speak regularly on a variety of topics. I now like to generate a descriptive group persona to stand-in for my likely audience members and get some insight into what this type of audience might want to learn and, more important, what they would want to learn to make them feel the presentation was successful for them. Not everything is on target or useful, but I always get helpful insights for preparing the presentation and expected audience questions.

I place each of these three in the category of “pushing the envelope” because you need to assess how accurate they might be, how accurate they need to be to be useful for you, benefits and risks of these approaches, biases and other issues they might introduce, and the like. Always keep in mind that these are starting points, iteration is essential, and that you, as the human, must get back into the loop at some point.

Limitations: Handling Constraints
In my experiments with generative AI tools, I always struck with how quickly the constraints of the tools and their underlying components reveal themselves. How much can and should we trust the tools and the results they give. In the case of ChatGPT, how do we deal with the current 2021 training cut-off? How do we adjust for inherent biases in the LLMs, the tools, and the results? How deal we use these tools ethically? Which of the tools should we use and why? Ultimately, for me, it all comes down to trust. And, today, and probably for the foreseeable future, that is a big unknown.

As you will have seen, I’ve designed the G-A-L Method to accommodate to and adjust for these issues and constraints as best as I can. What is and should be the limit on our reliance on the tools and results? What is the value of the recommendations and answers we get? How do we account for our own willingness to accept results simply because they confirm our own bias or seemingly validate what we have already decided?

In my prompt crafting and use of these tools, I try to address as many of these issues as I can. I use prompts to introduce diversity into groups. I introduce contrarian approaches and perspectives. I purposely use approaches where the 2021 cut-off has only a minimal impact. I also treat results as directional and flexible, and as starting points for what will become the human work I need to do myself. If I run a big experiment using an advisory group, I’m happy if it gives me a few ideas for later iterations and work. My approach is realism first and then learning and verifying through experimentation.

My guiding principle is that the PCRO process and the G-A-L Method help me get started in disciplined ways that enhance the probability of getting useful results that I can move forward. The recommendations and advice of groups are starting points and help most as a tool to address the “blank page problem” as screens, first drafts and first passes. Most important, they help me determine where to begin to experiment and test.

As an illustration, the first human I told about how my group had converged to a consensus price of $32 per month as a consensus immediately told me that the price was way too low.

As a response ChatGPT once gave me said it well:

“Remember, while ChatGPT can provide diverse insights and simulate advice from different personas, it doesn’t replace the dynamic interactions, personal experiences, and nuanced understanding of a real group of experts. It’s based on patterns and information present in the data it was trained on and does not have personal experiences or subjective insights. Still, it can be a useful tool for generating new ideas, thinking through problems, and getting a semblance of diverse expert advice.”

This is why I call is the “Guided Advisory Layer.” The G-A-L Method is a framework for structured prompting that adds a layer to help guided desired responses form generative AI tools. It is a not a replacement for actual human advice.

Benefits and Future Directions
The results of my persona experiments that led to the G-A-L Method are far beyond anything I initially imagined. Rather than continuing to work unsuccessfully to create a mastermind group or advisory board in real life, I now have a growing numbering of customized expert advisory groups that are available on a just-in-time basis to answer whatever questions that come to mind, quickly, efficiently, and tirelessly. I was successful in avoiding for now an additional layer of complication from add-ins, APIs, or coding, and produced the results I wanted using only plain language prompts. I’ve created a toolkit of structured usable prompts for this purpose. I did not have to do any training of a specific dataset for myself or need to implement new technologies (although I’ll probably do so in the future). And I’ve found many potential real-world uses, in spite of running up against the limitations and constraints of these tools.

Here are benefits of the G-A-L Method I’m already seeing:

Just-in-Time Access. By creating reusable plain text prompts, it’s easy to me to convene my groups on issues and questions whenever I want. This access helps break through barriers and blocks to move projects forward or get perspective on challenging problems or choices.

Expertise Diversity. My groups embody a wide range of backgrounds and specializations. The expertise of each member is highly valuable and distinct, allowing me to get insights from different angles. This diversity will provide a holistic approach to problem-solving and ideation.

Continuous Learning. By consulting regularly with my groups, I keep you updated and closer to the leading edge in my fields of interest.

Innovative Ideas. I’m consistently getting some ideas, options, and potential solutions I doubt I would have found on my own.

Resource Optimization. The amount that can be accomplished and results generated in a short time is sometimes amazing. I’m reducing my time for planning, ideation, and the like to minutes rather than hours, days, weeks, months or years. These time savings are especially evident in strategic planning work.

More Confident Decision Making. I still need to do some testing to see if the G-A-L Method is leading to qualitative better decisions, even though I believe that it is. I now better understand the rationale of decisions I make using group advice and am more confident in the choices I make.

More importantly, I found some practical uses and tools that I want to share with the world. Not surprisingly, one of my mastermind groups convinced me that was the way to go after I thought about keeping the G-A-L Method to myself. However, I was already leaning in that direction before asking the group. This white paper, in fact, is my human response to the advice my group offered.

What’s Next for Me
I want to get feedback on this approach, especially after your try it. I have a number of new group experiments already on my to-do list. I plan to write about some of my specific experiments and results I’ve gotten. My focus lately is on personal productivity uses: how might I use these groups to help me set priorities, set goals, be accountable for results, and the like? I’m also working out an idea to license a prompt based on a persona that encapsulates my expertise, experience, and approaches that others could add to their groups and boards to get my assistance and to scale my ability to advise people. I see many directions the G-A-L Method can go, but now I have some help with sorting and developing them.

Conclusions
The G-A-L Method presents a groundbreaking approach to overcoming the “blank page problem,” especially in how it leverages the use of personas to create group interactions and dynamics to deliver practical advice, insights, recommendation, and more. It integrates the foundational principles of advisory boards and mastermind groups with the computational prowess of generative AI tools. Traditional approaches, while valuable when successfully implemented and sustained, face real-world logistical challenges, including cost, availability diversity, connection to experts, and the like.

However, by employing the G-A-L Method, businesses and individuals can summon tailored advisory insights without the limitations traditionally encountered whenever they want and for whatever problem or question they have, all in a matter of moments. Through structured crafting of plain text (not coding), you can direct a customized virtual team, tapping into a reservoir of synthesized knowledge and insights, and elevate your role as the human in the loop to a higher level. You evaluate, validate, refine, and drive the results and actions. In the legal profession, we talk about ways to let lawyers “operate at the top of their licenses,” which is a good way to describe what is happening.

Remember, while this can simulate different perspectives and help you think through problems, ChatGPT’s responses are generated based on patterns and information in the data it was trained on, and it does not possess consciousness or subjective experiences. Use it as a tool for generating ideas and stimulating thought, not as a replacement for actual expert advice.

As we venture deeper into this new frontier, let this white paper serve as your compass, illuminating the vast potentialities awaiting exploration. Perhaps most importantly, it will shift much of the hard work onto the generative AI tools and let you do what humans do best.

I invite your feedback and encourage you to share your own experiences using the G-A-L Method.


Further Reading and Resources
To be honest, there’s no substitute for rolling up your sleeves and experimenting with the generative AI tools. Nothing beats putting in the hours. I’ve primarily been using the $20/month premium version of ChatGPT Plus and GPT-4, but that might change in the future.

That said, here are some resources I suggest:

I found Melanie Mitchell’s great book, Artificial Intelligence: A Guide for Thinking Humans,” to be especially helpful as a starting point. I use it in my class and recommend all the time.

Ethan Mollick’s generous work on practical uses of generative AI has also aided my approach, even if it deterred me from going into the world of code.

There are many hands-on instruction videos on YouTube, especially for specific issues.

And, of course, ChatGPT is a tireless explainer and can be quite helpful if you craft good prompts for it to run.


Acknowledgements
My path into AI was guided by the work of many and extends back at least as far as 1982 when I took a Computers and the Law seminar at Georgetown University Law Center with Milton Wessel and we discussed legal issues that AI might raise. A winding road led me to currently teaching a class in AI and the Law at Michigan State University College of Law. I spent a good chunk of this past summer developing the G-A-L Method and a number of other generative AI “applications” using plain text prompting.

Marc Lauritsen has long been a guide for me on AI in law. I’ve benefited from the feedback of several people (even if they might not have been aware of it), but Allison Johs, Elise Auxier, Genny Adel, Michael Kraft, and Tom Mighell have been especially patient listeners and encouraging as I’ve described early versions of what became the G-A-L Method. Thanks also to Andy Whitehead for helping me think about publishing this paper.

Licensing and Use

This work is licensed under a Creative Commons Attribution 4.0 International License.

I want people to use the G-A-L Method and talk about it freely (although donations are appreciated). What I care most about is attribution. Let people know where you got this when you use it or talk about it and follow the license terms. Do not claim it as your own – it’s so much easier to mention my name, isn’t it? If you want a commercial license to clarify your rights to any other intellectual property that might be associated with the G-A-L Method, please contact me. I do not intend to file any patent applications with respect to the G-A-L Method or other inventions that might be covered in this paper.

About the Author

Dennis Kennedy is a well-known legal technology expert, educator, author, and keynote speaker with a focus on transformative innovation for law departments. He is the Executive Director of the Kennedy Idea Propulsion Laboratory, a legal technology and innovation think tank. He is also the Director of the Center for Law, Technology & Innovation at Michigan State University College of Law and teaches cutting-edge courses on technology law and innovation at both the Michigan State University College of Law and the University of Michigan Law School.

His books include “Successful Innovation Outcomes in Law” and, with Tom Mighell “The Lawyer’s Guide to Collaboration Tools and Technologies” (Work from Home Edition, 2022). Dennis and Tom have co-hosted The Kennedy-Mighell Report podcast on legal technology since 2006 and he has been blogging at www.denniskennedy.com/blog/ for more than 20 years. Dennis retired after serving as Senior Counsel for Digital Payments and Labs at Mastercard. Learn more about Dennis and get contact info at www.denniskennedy.com or www.linkedin.com/in/dennismkennedy.

Version 1.0

(c) 2023 Dennis Kennedy


[Originally posted on DennisKennedy.Blog (https://www.denniskennedy.com/blog/)]

Want to schedule a Zoom call to talk with me about Legal Innovation as a Service, Speaking, or other services? Schedule a Zoom with Dennis via Calendly.

Like this post? Buy me a coffee

DennisKennedy.Blog is now part of the LexBlog network.

Legal Innovation as a Service Logo

Need a little help with your legal innovation efforts? Check out my Legal Innovation as a Service offerings.

Follow me on “Twitter” – @denniskennedy