Hey,

Most of the times, when I write here, I draw inspiration from conversations happening with clients or David.

Today's mail is not different, I had a really insightful call with a client of our Automation Engineering Bootcamp program, and I think it could be helpful to you too.

Why use multiple AI personas?

One of the key insights from the call was that why one needs to use multiple personas and switch them during ChatGPT conversations, or when building AI agents.

And this really shows a difference and a similarity between how the human brain works, at the same time.

We humans are good at doing many things, even in quick succession. I don't believe in multitasking being productive, but if I introspect now in my brain as I'm writing this email I can notice a lot of different processes working.

I'm thinking about the topic itself. I'm thinking about how to teach it to you, and to do that, I'm also thinking about teaching styles and the knowledge I assume you have about the topic. I'm also thinking about how to write, and use my previous experience and education I got about copywriting and marketing. I'm constantly thinking about what to write and what not to write. I'm thinking about how this fits into your daily lives and our marketing goals.

This is a lot of thinking that's happenning. It's also called metacognition: the thinking about thinking.

I can do this, WHILE maintaining a high quality of work (I believe :D ) but LLMs can't do this.

If I were to give an LLM everything I learned about copywriting and marketing, and our company, etc. it would not be able to perform the taks with high precision.

But if I know and can articulate how I think when I do something, I can create these mini-personas, and make them focus on only one thing at a time.

And LLMs are really good at doing one task at a time. And as the manager of the AIs, it's my job to guide these together.

So instead of having one LLM that does all the tasks of what one human brain would do, we map certain parts of the human brain responsible for each task and have an LLM do that instead.

The art of conversation design

In our Prompt Master AI course, I teach the methods of conversation design, which is the process of writing prompts in a strategic way to guide ChatGPT to make things that would otherwise be impossible with one prompt.

On the call today, our client not named John, has asked me about how do I pre-prompt ChatGPT and feed it information before I actually run the task.

I said "give me any task"

He says, "ok, let's do a financial plan"

I was like, sure, and opened up ChatGPT.

First prompt I sent was:

Ask me a list of questions that you'd need to know the answers to in order to give me good financial advice for my personal finances

This made ChatGPT throw me like 15-20 questions about finances. Like what's my monthly salary, how many savings I got, etc.

Here, if it was a real scenario, I would answer these questions based on the actual information I know.

For demo purposes, I made ChatGPT come up with the answers:

Act as a 27 year old single man and come up with answers to these quesitons, just guess if you don't know.

This made ChatGPT respond with answers to the questions it made. They were surprisingly accurate... :(

Anyway, then I asked it to:

Okay, based on this, now act as a financial advisor, and give me advice on how could I buy a $300k apartment in Budapest Hungary.

And it gave me a really good financial plan to execute on for the next 5 years. Even though the data it was going off on was not even my own.

So that's how I made ChatGPT load its own context window with information. Using the keyword "browse" you could also make it find actual information about companies, etc.

How this transfers to AI agents

So the next logical step was to build an agent that has conversation design built in. This is fundamental to David's Proxy architecture, where, in essence, multiple agents talk to each other to do complex tasks at human quality level or even beyond.

So we pulled up an example about building a StoryBrand (from Donald Miller). It's a seven-step process, if you've read Donald's book, you know what it is.

I told our man not named John to "Let's draw up the architecture flowchart of a hypothetical bot, that would build a StoryBrand like brand messaging from any company's information. How could we build that AI agent?"

And then we went into the fundamental building blocks and fundamental personas required to carry out this task.

This is where First Principles Thinking is key.

Can you break down a task into it's Fundamental steps? Basically the atoms that we can't break down even more?

In this case, the first step was the "Character" of the StoryBrand.

So our input was a lot of information about a certain company. A person using this bot would provide this information, and then press "Run" and the end result would be a brand messaging.

So here we introduced a few different personas.

Persona1: An interviewer doing customer interviews

Persona2: An ideal client of this company

Persona3: A summarizer agent

Persona4: The StoryBrand writer agent

Here's how it would work in theory:

1. Persona1 takes all [input about company] and writes questions for customer interview. This is [output1: questions]
2. Persona2 acts as the ideal client of company based on [input about company] and answers [output1: questions]. This creates [output2: character answers]
3. Persona3 acts as a summarizer, and takes all [input about company] [output 1 and 2] and writes a detailed summary of all the key insights, quotes, everything that's important. This is [output3: ideal client insights]
4. Persona4 then takes on the role of writing the brand messaging using [output3: ideal client insights], and adds that to the final document.
5. Output 1-3 are stored in an archive document.

This agent can now take ANY information about ANY company, and write the first part of the messaging.

It is an AI agent of multiple personas, that conducts a pre-determined conversation with persona switching, each response feeding the next prompt.

How well this would perform will come down to how well you can write the prompts, set restrictions, etc. the 6 key elements.

And I can guarantee you that it will save a lot of time without much compromise on quality.

Maybe for time savings it's not the best example because you only do this messaging once per company, so if you do it for yourself, just do it with ChatGPT, don't build an agent for it.

If you want to offer it as a chatbot service, then it makes sense to automate the steps in the flow.

And the similar idea can be used for many other things. Repetitive marketing tasks like writing the social media post captions that everyone hates to write and nobody reads.

Time savings are everywhere, and if you want to save time with AI, our Prompt Master AI course is you.

By the end of it, you will become a Prompt Master who can automate tasks with no-code tools and AI, without writing a single line of code. Click here to learn how to become one!​

Talk soon,

Dave

P.S: We offer a bonus consultation call and 90-day money back guarantee for people who buy now, so don't wait on it, take action now. And if you have questions about the training or want to just chat with us, hit reply and let us know via email.

‍