Introduction

So my name is Yanis. I work for Yodek. My background is in business, data, and marketing.

I don't know if you remember this movie. It's a bit old one, but the main character actually fell in love with a bot.

From Sci‑Fi Bots to Practical Personas

I couldn't imagine myself speaking with a bot with the same passion when I created Chloé. Chloé is one of Yodek personas.

Are you familiar with personas? Are you using personas? Yes, a little bit.

So personas actually are fictional, let's say, characters that are really close to the real customers of a company, right?

Building Chloe: A Custom GPT Persona

So I built Chloe, which is a custom GPT. It's actually a bot. It has all data we know about our personas.

Grounding the Persona in Real Data

It's also based on reviews from, let's say, Captera, G2, other platforms, and also demos from sales demos with real customers. All of them filtered if they, let's say, have the same job, let's say, title with Chloe, which is internal co-manager. working for HR department.

What Chloe Is Good At—and What She Isn’t

So Chloe, as a GPT, it's really good to tell us her opinion, what she likes, what she dislikes about us, and how she feels about, for example, a new feature or maybe how she feels about a landing page or an ad that she sees.

Although she's really bad to implement her suggestion, right? It's trained to work only as a person, not as an expert of design, for example.

How to Set It Up

How to set it up? It's really simple.

Learning Resources and the Prompting Approach

I learned it while doing the course in MindStone. It's one of the basic, let's say, lessons. I really recommend to do it.

There is a small code here.

It's not a code. It's actually natural language.

Structuring the System Prompt

saying, for example, you are this persona. Your goal is to behave like this persona. Here is more context around your persona, et cetera.

And I have this a bit more, how to set the base is a bit larger.

Rediscovering an Old Persona and Bringing It to Life

And it's interesting that I find exactly this persona six years ago in a hidden document. somewhere on our documentation, let's say up, and the company CEO actually wrote it.

So it was always there, but I think nobody really used it a lot. So recreating as a virtual character, it was much easier to use this persona on our everyday tasks in marketing.

Using Emotional Framing in Prompts

Another interesting part of the code is that there is a bit of emotional pressure. So it's a best practice, according to the course, to say to the bot, take a deep breath before answering. Based on the studies, this helps a lot to be more accurate.

Or I will tip you $300 if you reply me correct, et cetera. This kind of pressure that you think works only in humans eventually works also with bots.

In the beginning, it did not sound funny. I mean, did you believe it would have impact? No, I was laughing as well.

But it works. Yeah, it works. It works.

And it's interesting that you cannot use this part every day, every time you have a prompt.

Why Custom GPTs Beat One‑off Prompts

But if you make a custom GPT, like Chloe, for example, then it's there. It exists. You don't have to repeat it every time.

So you don't need the perfect prompt. You just need, let's say, to speak more like a human language from the moment you build it.

Importing Real‑World Data

As a last step, we import all data we had about this type of customer. We have a lot of reviews, like 5,000 reviews in total from these directories. We also have some AI notes from demos we do with customers. So we took everything, we collected everything, and we gave it to AI.

Live Demo: Creating and Using the Persona GPT

I can show you how this looks like. The interface looks like GTP, but I can demonstrate it live. Also, I can show you how to build it.

If you are in chat GTP, you can go here to GPTs, and then here, create. Do you see here? Here you add the code that I just mentioned, right? And you give a name, of course, et cetera.

And here you can add the files. It's really easy, just three steps, and you have a persona to fall in love with. And then I can show you some really interesting discussions I had with this persona.

What We Learned from Chloe

First of all, I asked her, how are you using screen scheduling feature, which is one very famous feature we have in Yodeck. And if you do this prompt in the chat GPT, I think you will not get any relevant answer. But with Chloe, the answer was really relevant, saying, oh, look, I have some safety reminders for my employees. So it starts really behaving like the real customer.

Landing Page Critiques that Spark Experiments

Another interesting example, I asked her to evaluate a landing page. It was a landing page that was built for this persona initially. The Office Digital Signage landing page. Here is the URL.

And then Chloe responded, look, I don't really like it. Because, for example, the testimonials are not from HR managers like me. They are from CEOs or from IT guys. Why do you have these testimonials? I know why, because we had the same testimonials in all industry pages, for example, which was to save some time.

She found it. She also found that some of our call to actions were very generic, like start for free, download this, nothing specific for her needs, like I want to see sign-ups for HR teams, for example. In fact, Google is penalizing if you have really generic call to actions, like learn more. In every button you click, you will learn something more. It gave some really nice suggestions, and after this, we planned some experiments to do in the future.

Ad Feedback and Multi‑GPT Workflow

Also, I asked her to evaluate Google Ads. She gave me some suggestions, although if you know a bit of Google Ads, you will think, okay, these suggestions might be too long. It's out of specs to implement.

Then you can use the multiple GPTs. So build one to start with, provide context with a personal GPT. Then ask her about your opinion about what she might want to see.

Then ask from a copywriting GPT to build the content based on the specs that this bot already knows. And then if you want to add also a design, build another GTP to build a design. And this time, this GPT also knows the specs.

Orchestrating Bots to Ship Faster

So we have a workflow. And you can call all of these, let's say, AI bots to collaborate to have something that's according to the specs. so they minimize the time from idea to have a ready ad or a ready blog or a landing page.

Scaling to Multiple Personas

In the end, we built the bot for every of our persona. It's a fun fact.

Evolving Insights: When the Top Persona Changes

When I finished this, I took some data, because this persona is something like six years old. I took some data, and Chloe was not any more relevant, was not one of the top personas.

In fact, the marketing manager persona was much more interesting. It had much more clients from this persona.

But anyways, it was something to start with.

Wrap‑up and Contact

More or less, that's the presentation.

Here is my email. If you need anything, feel free to send me. I'll be really glad to assist.

Q&A Session

Any questions? Yeah?

Nice work.

My question is, have you or are you thinking to implement a server? And if yes, how do you plan to do it? Yeah, that would be the holy grail if we could do it.

But I think self-learning GPTs are not yet available. Last time I checked, it was not yet available. So I know if this changed. But back then, it was not an option when we built it.

This is like five months old, let's say.

Here is Joshua. You can ask him if it's available. Yep.

Validity and Overfitting Concerns

Have you seen any study or have you tested the persona with the real thing to see if what the persona says is similar to what the real person would say? That would be a nice idea for a secondary study to check if they are close.

But this bot is based on thousands of comments we have online from this type of customers, let's say. Personally, I found it really close to what I was reading from these comments. That's why I promote it internally and said, look, I see value here. It's really simple and it's a pity if we don't use it.

All right, but that could be overfitting the comments you gave. Have you tested it? Indeed. It could hallucinate a lot. That's part of the game, right? But we could extra check it.

We haven't done any, let's say, reality check if this persona really reacts 100%. It would be really hard to test it, actually, because each person replies differently. But if we have thousands of data, then probably we can build something that represents this average user, let's say. But there is no average user in reality.

for a nice study to see if those personas actually correspond to what a real person would say in words. Yeah, it would be nice. It would be nice. I agree with you.

Yes, you've been pressed. Sorry. Just a follow up to this question.

How Much Is New vs. Summarized?

So like Chloe. gives you feedback based on the feedback it reads online, or it actually creates some new feedback based on what it feels?

It actually builds a character based on feedback online, but he's trying to simulate this. So if we go through all the comments, more or less see the same comments that we did, like a summary.

With this comment, she is building her character, right? But if I ask something new that doesn't exist in this comment, for example, I ask her, give me your feedback about this landing page, right?

The comments ask feedback about Yodek in general, not about specific ads, landing pages, features, whatever you can build from now on, right? Yeah? Yeah, interesting question.

Sharing and Permissions

I think here they are. So here is Chloe. I can go to edit.

So you can just decide to share it. And in the settings, you can share it with more members in the company. And if somebody shares something with you, you will see it here.

So somebody shared this Sailor Moon bot with me. But no, I think there are GPTs that are publicly available, so you can decide to make it public. You have the option, yeah. No, this is confidential, but yeah, you can decide to have it public.

For example, you can build a GPT for three suggestions, you know, and make it public. I don't know if there is, let's say, if there are paid GPTs yet. It would be nice to see in the future.

And then it has to have an OpenAI account, a pro? Ah, yeah, I forgot to mention, you need to have a pro account for this. Plus email setup for the company. Yeah.

It's something, I think, really simple. Yeah, yeah, yeah. Yes?

Cross‑Model Use and Internal Adoption

Have you tried the prompt with the persona from the other models? Not yet. Not yet.

I'm trying to convince my team to use this first in order to make, you know. Let's say the CR specialist, yes, they get ideas from personas, but now I'm trying to convince content team and other teams to use it.

Yeah. It's not so easy. Yeah.

Yeah? Yeah, good question.

Role of Real Customer Research

Did you use any advice from your CXR, from your customer experience research team to build this? How many? I'm not sure you're doing customer experience research with publishing feature, but if you do, how many cycles research with real customers,

To be honest, I don't think it can yet replace real customer feedback. It can just give you an early feedback about something you want.

To educate it, we get all the demos from the sales team. And this was, let's say, an input.

But I don't think in the future. In the future, I think it's useful to still speak to customers, to real customers. Maybe to less if you have a bot like this. But don't completely skip this part yet.

Data Ingestion Mechanics

So I just uploaded in, let's say, an Excel format. So I just upload here these reviews. I just try to upload it here.

So you can, again, if I edit it. Edit.

But of course, it might be another way to do it better.

OK. Yeah? Hi.

Can you just continue? OK.

Last question. Yes.

Creating the Marketing Manager Persona

Can you tell us the process that you followed to create the fourth persona, the marketing manager? Sorry, please repeat. Can you tell us the process that you followed to create the fourth persona, the marketing manager?

I just checked which persona is giving us more money, and I found out that we get more money from this persona than Chloe. That's it. Just like MRR per persona, let's say, per job role.

So it's same data, different personas? I mean, what data did you use to build the Maria? Ah, the same, the same, yes, the same.

I just realized when I finished. Was it also exactly the same? Exactly the same prompt, just changing the context inside the prompt. So here when you specify, for example, when you specify here who you are, you just change name and this paragraph, which is the description of the persona and how you should behave.

But you feed the same data. The same data from different reviewers, let's say. Testimonials from different reviewers that are marketing managers in job role, not HR managers.

Data Limits and Correct Labeling

So there is a limit on the data that we can fit in? There is a limit, but we want to simulate the correct persona with the correct data, correct reviews, let's say.

So I wouldn't give to Maria, let's say, HR manager reviews, because it would be a wrong input, right? I didn't get all reviews. I get only the reviews with the correct, let's say, job title. Yeah.

Conclusion

That's it. Thank you very much for your attention.