You’re in the R&D lab at GoolyBib, gearing up to help Gustav Gieger, the CEO, figure out how to simulate feedback from virtual customers using a clever technique called “Personas of Thought.” Instead of relying on just one expert opinion, you’ll be tapping into a whole crowd of diverse personas to generate richer, more balanced insights. It’s all about capturing different viewpoints before combining them into one smart, nuanced response — a real game-changer for innovation brainstorming. Let’s get started!

I keep hitting a wall when trying to predict how different parents will react to our new smart bib features. One perspective just isn’t enough—there are so many types of users out there.  
If only I could gather insights from a whole crowd of virtual customers, each with their own unique viewpoint. That kind of feedback would save us months of guesswork.  
But how do I simulate that wisdom of the crowd without actually assembling a panel of hundreds of testers?  
I need a method that lets me generate diverse opinions and then combine them into a clear direction—something smarter than just asking one expert.  
Maybe this “Personas of Thought” technique is exactly what we need to finally crack this problem.

**Personas of Thought: Harnessing Diverse Perspectives for Smarter AI Responses**

In the evolving landscape of AI and natural language processing, one of the challenges is getting responses that go beyond generic, surface-level answers. Typically, when we prompt a large language model (LLM) naively, the result tends to be bland and predictable—often sycophantic, overly positive, and lacking in depth. To overcome this, a promising technique called *Percentages of Thought* has emerged, an adaptation of the popular chain-of-thought prompting, that leverages multiple personas to enrich AI responses with diversity and nuance.

At its core, Percentages of Thought invites the model to imagine a variety of different perspectives—either world-class experts or everyday personas—each contributing their unique viewpoint on a given question or task. Instead of a single, monolithic answer, the AI generates multiple diverse opinions, which are then aggregated into a thoughtful, well-rounded final response. This method has shown to produce far more insightful, non-obvious, and human-like answers.

### Experts Prompt: Wisdom from the Top

One version of this technique is the *Experts Prompt*. Here, the AI is asked to first name several world-class experts relevant to the question—past or present—and then answer from each of their perspectives. For example, when considering the best way to learn a new skill, the prompt might include Malcolm Gladwell, Angela Duckworth, Tim Ferriss, and Richard Feynman. Each expert offers a distinct viewpoint: Gladwell emphasizes 10,000 hours of practice, Duckworth focuses on grit, and Ferriss shares practical hacks.

By aggregating these expert opinions, the final response becomes richer and more comprehensive than one generated naively. The experts prompt encourages critical thinking and diversity of thought, exposing insights that would otherwise be overlooked. Additionally, it fosters a balanced critique—highlighting strengths and potential weaknesses—rather than simply endorsing an idea.

### Personas Prompt: The Power of the Crowd

Complementing the experts prompt is the *Personas Prompt*, which asks the AI to consider viewpoints from everyday demographic personas relevant to the task. For instance, when evaluating a product name, the AI might draw on the perspectives of a busy parent, an athlete, a fashion-conscious individual, and an industry expert. Each persona highlights different priorities—convenience, style, clarity of purpose—which collectively surface a more nuanced understanding of the problem.

This approach taps into the "wisdom of the crowds

Personas of Thought

  All right. I'm going to walk you through a technique that I call percentages of thought. It's just a play on the chain of thought idea where you would get an LLN to spend some tokens, to think about the task first, before doing it. And this adaption that I I've been using more recently is that you ask it to come up with a bunch of different personalities or personas to demographic percentages or experts. Who will have a different opinion on on the task and then you can aggregate their opinions together to get the fund a result. 

And I find that it's really good for. The rest of your thought and uniqueness of insights, rather than just the generic, bland stuff that you get from Anela. So here's some code that I'm running here. Just to have a question that I want to ask, like I'm trying to come up with an idea for the, I should use this product name for a new product I invented and I just call open AI. 

This is two fifty four, a mini.  And the prompt is very simple. It's just, I want a paragraph response to the following question.  And it's the typical sycophantic stuff. It's like a strong choice, and it's really good. And it's. He catchy and easy to remember. And I actually think this is a good name, but yeah, this is a typical response. 

It's in a very rarely as useful. Information, if you just prompt it, naively. So the next one we're going to try is the experts prompt in to run this. But this is very different. I want a paragraph response to the phone question, but first I want you to name a number of world-class experts past or present who would be great at answering this. Then for each expert, I want you to answer critically from their perspective, given their background experience, and then finally combine all of that together into a single response as if they'd collaborative. Collaborated and writing a joint. Anonymous sensor. And I've given it some structure in terms of the format. 

So we'll see what we get. I came up with in this case at Nike designer. From an FDA commissioner and event to Harvard business school, professor Tim Ferris.  Tony Shay. These are all fairly relevant, which is great. And there's good at coming up with good ideas here. For inspiration. And then as given individual feedback for each tinker Hatfield thought it was good. But this one, this guy, David Kessler. I was worried about that. Whether it resonates with consumers, health needs. And that makes sense. Ah, because he's an expert in consumer health. So that's the sort of thing that just wouldn't come up and didn't come up in the previous feedback that we got from the LLM. 

But because it's been pushed to get a diverse crowd of opinions then it ends up being quite useful in my opinion. So then you can go through, you look and see which feedback you agree with. And I find it that's really great for brainstorming in general. But then also we get this final response at the end. And it has some of that diversity of thought in here. 

It says, however, it's essentially, the name is distinctive enough. And then Yeah, it doesn't specifically mention the health benefits in this case. But it does. Does it, it's not very sick of antic. Like it does. Bring a bunch of these ideas together, which is really helpful. I find this technique works really well now. This this is another type of prompt, which is exact same thing as the experts prompt, but with just regular people. 

So personas prompt is what I call it. Name a number of demographic personas. Who'd be relevant, Francine this. What we're trying to do is get like a crowd of people. And instead of them being experts, we just want regular people. And w what are the relevant demographics? So in this case we do have a, Footwear industry expert, but we also have a busy parent and an athlete because they're going to have different perspectives. And then it's a fashion conscious person. 

So we're going to tackle this from multiple angles. You can say, it does sound in a direct and Trendly, but doesn't specifically mention footwear. Only fit might be too vague in this case. Suggest convenience for busy parents have a doesn't clarify the product's purpose. So I'm getting a much better. In a broad sense of the problems with this name, which I think you didn't really get in the previous ones. 

And then this case says, we do think it's a promising name. However it needs to clearly convey the product's unique selling position proposition. I think that's a much more helpful, final response in my opinion. Cool. So bringing all that together, like here are the three responses side by side.  And in general, whenever I run this city, even for any type of question, I'm finding I'm getting much better results with the persona one and sometimes the expert one as well. 

All right. We can judge this, like we can use an Ella and judge in this case. This is just another prompt where I'm asking it to compare the two responses. And I want to make sure that the responses are unique, that they're non-obvious and that they sound human, not like an AI. And then it's going to choose which one is better. 

This is like a common format for. Judging. Responses, because I want to run this a bunch of across a bunch of questions and not just one question. And I don't want to just like manually check each one. So I'm automating the checking in this case. This code basically just runs. It just runs the judge prompt and we can see for just.  Run this now.  It's gonna. And the naive versus experts prompts a chose experts. 

Number two. And now you versus personas that chose number two and then experts versus percentage that chose number one. So what that means is they both an improvement over the naive prompt. The first one we tried but that the experts one was better and it comes with some reasoning here. You can see like why it chose that. But it's it says responsible and provide more comprehensive analysis, which is useful. Cool. 

You'll get something different when you run it and especially if you change the question, but I find that the judge also, It's pretty, pretty good at. Agreeing with what my assessment would be which in this case, I did think that the percentage one was a little bit better. 

But yeah, in, in, in this case, it's pretty clear that they're both better than the naive prompt.  All right. So bring all this together. I'm just going to run this all this does is take all the things we just did. And put it into a function called run test. That takes in a question. And then we're just going to run through every question here. 

So there's five different questions and we just want to see generally how often does the persona prompt when versus the. Expert. Prompt say. Right now. It's just setting up the naive response. And then sending them to the experts response and then sending it with punch sentence response or for that same question. 

And then it judges them.  And here, this is where it's judging. And it's printing stuff out for us, and then it returns to find our results. So you can see that.  And this case, you have the question, what's the best way to learn a new skill. And the native response is useful. 

But the expert response it's pulled in all these heavyweights, right? Pull them Malcolm Gladwell, Angela Duckworth. Erickson K and as Tim Ferris, it's versus pretty common. Actually in all of these questions but also people like Richard Feynman. And they all have a different opinion. 

Malcolm Gladwell talks is sharp. It talks about the 10,000 hours. Angela Duckworth talks about grit. Hope book was. Called grit. So as you would expect, but because you're forcing it to be a little bit broader, it does tend to work a lot better.  

You can see at the end, hit the.  

I shall send this response is quite useful.  

Actually, we can pull this open in a text editor. We can see.  It's a federal response.  

 

and yeah, I think this is really useful. Like it talks about structure lending, hands on practice, community engagement. This is the sort of thing I would recommend people do as well. Practical skills, dive into real-world projects, seeking mentorship. This is exactly what I do to learn a new skill. And you can see as well that in this case, the percent is prom. Beat everything. Experts be naive and then present those beat experts. What's going through the rest of the questions.  And we'll see how it gets to the end.  

What we're going to do at the end is just add up. How many wins each one got, and then they'll say how many times they're a tested. When we run naive versus experts. Then we'll count one for naive in terms of trials and one for experts in terms of trials. And then we're also going to count.  In this case here, if naive one. Then we'll get. Incorrect the winds for naive or effects, but it's one that we can make around the wet and sick.  What will happen is we'll get an accurate percentage of. How many times did it when it was being tested against one of the other prompts.  And we're just going to see what that looks like at the end.  

Okay, everything's finished running now, surgical to run this. And here we go.  So on average, across the different trials. The nave prompt only one, one and 10 times. So sometimes it is the better one, but it's pretty rad. The experts prompt was much better. 60% of the times at one and the percentage from 180% of the times in the trials. 

Just this. Just like pretty common actually I'm funding across different questions. I try this on when I'm measuring it. The personas prompt actually tends to be. The best and it's like wisdom of the crowds, so I'm going to show, I think with the experts prompt, it does tend to be better than not what you naive the get. But it sometimes goes a bit over the top with the The details of what Tim Ferris, what Daniel Conoman, whatever. All these famous people will specifically push like the 10,000 hours. 

And I think that harms the diversity of thought sometimes. Whereas when it's looking for regular people and it's having to be a bit more creative with how. Those people would approach answering that question.  Cool say. This is just meant to be a way to demonstrate how it works. I use this technique in lots of different ways, just like chain of thought. It's not, it doesn't necessarily have to be applied exactly like this. 

It's just more of a concept that you can try. And you could do this and chatty putty as well, but in general what I tend to do is I'll create the percentage first. As like a separate step and then I'll run each one of those in parallel. And then I'll aggregate at the end. So that tends to be how I approach it. I just want to just show a simple version here where it just does it all at one prompt. Cool. 

Yeah, hopefully that's a useful for you guys. 

What is the main idea behind the "percentages of thought" technique described in the tutorial?  

How does the "experts prompt" differ from the "naive prompt" in the example given?  

According to the tutorial, what is one advantage of using the "personas prompt" over the "experts prompt"?  

According to the tutorial’s experiments, how do the naive, experts, and personas prompts compare in terms of winning percentages, and what can we infer from this?  

ChatGPT is designed to be a helpful assistant, but can it help you with customer research? By getting AI to roleplay as your potential customers, you can get more diverse and realistic feedback without paying for a focus group or survey.

Vexpower Learning Certificate of Completion

Congratulations Steve Kozel 💪