Building personalized GPTs - Igor Pogany

Introduction

OK, welcome.

So I'm going to be the practical demo, but I have 20 minutes, I suppose, right?

I want to use half of that to tell you a story about my Sunday and how I solved the issue that has been plaguing me for years now, accounting, the nightmare of every entrepreneur.

So I'm going to tell you about that.

And then we're going to go into the practical demo part.

1Matter of fact, it's going to be quite simple today, because I'm going to be showing you two GPTs that kind of do all the work for you.

The Story of My Sunday: Accounting Woes

But I hope, my hope is that in the first 10 minutes, you will understand why

using these two gpts in the second half actually makes sense right you need a little bit of context so let's begin so let's talk about how i spent my sunday the reason all of you came here right hearing about my sunday but trust me on this it's going to be interesting because

This has been my Sunday.

Look at this.

Isn't this glorious?

Look at that.

Let me scroll through.

Oh, my god.

We could spend like minutes scrolling here.

The point is this.

Consumer-Focused ChatGPT Use Cases

I want you to start thinking about ChatGPT interactions.

And we're going to be doing consumer products, OK?

This is not like enterprise, Azure, a land chain, function calling, all of that.

This is pretty basic stuff how to use ChatGPT as a consumer, OK?

I want you to think about it in two buckets.

First bucket, you just get going.

You solve problems.

You don't give it much context.

You don't use custom instructions.

You just get going.

You start prompting.

You start writing.

You start solving problems.

Second bucket, you want to solve a problem for somebody else and make it a little easier on them.

Maybe that somebody else is an employee.

Maybe it's a friend.

Maybe it's a family member.

Maybe it's somebody who's not as good with this technology.

Who knows, okay?

Very different approaches in these two buckets.

In the first one, you kind of just get going and you can start having a conversation with the tool.

As you might know, one of the best things you can do is talk to it as you would to a human, right?

If there's an issue, just tell it and let it correct its mistakes.

This is what I did here, okay?

Personal Problem Solving with ChatGPT

We're not going to assume any Python knowledge or anything, but basically,

I had this problem in my life, and let me tell you, it has to be top three problems, because it has been plaguing me since eight years now.

I had a company back in Austria, now I have a Portuguese one, I have a corporation set up in Slovakia, and I myself still do most of the accounting for that, meaning I need to find all the invoices, I need to submit them to my accountant, but I also have to go through the entire Excel sheet and make sure that every transaction on my bank account for every month

company bank account has the corresponding invoice.

A nightmare if you're subscribed to all of these AI tools, right?

By the way, if you don't know what I do for a living, I'm an AI educator and YouTuber and all of this good stuff.

So basically, it's kind of my job to subscribe to all these new tools that come out.

And I want to expense them, right?

But then if you have a list like this of different AI tools that you're paying every single month, and then at the end of the month, you've got to find every invoice, you're going to spend

Half a day, maybe a full day, just logging into different websites, downloading invoices.

It is a nightmare.

And I've had enough of it.

I was like, okay, this might take up my entire day, but I'm going to automate this away, at least a part of it.

So this is what I did here, okay?

I just went in and I started talking to it.

By the way, I'll skip forward to the result.

1I spent nine hours automating something that would have took me one hour, but then, hey, no more one hour every single month from here on out, okay?

And what I just did here, I just started describing my problem.

I said I have a basic script, and there's all these basic invoices that I have to retrieve every single month.

And we're not going to go into this conversation, obviously.

But what it did there, it gave me suggestions on how I can tackle this.

Then I went ahead and I told it, OK, I'm going to have to do this every single month, and I have sort of an idea of what we could do here.

We could let you manipulate the Excel sheet.

I'm not going to do that anymore.

You're going to go ahead.

I want you to do the color coding.

I want you to do the filtration.

I want you to do all the things that I've been doing for eight years every single month.

And it went ahead and wrote code for me that I could run.

First version, second version, 50th version, whatever.

At the end of the day, I got to a point where it does this for me automatically.

Can you see this?

This is just a screenshot.

It's just a part of it to demonstrate it.

But it basically color codes it for income.

These are irrelevant.

These are only invoices for.

And these are automatically retrieved.

And then all that is left for me is these white columns every single month.

So look, this is bucket number one.

Customized Solutions for Others

This is solving a very personal problem, something that is unique to you.

This is not going to be applicable to you because you're not going to be subscribed to the same services.

The filtration is going to be different, right?

And here you have like, I don't know, 11 labs, Canva, ChatGPT.

Here you have like...

Doesn't matter, it's gonna be personal to you.

This is where you need to have an interaction with it, okay?

This is the first mode of operating with ChatGPT.

Talk to it.

I always say talk to it as if it was an assistant, a world-class assistant that you just hired and massively overpaid.

Gonna have high expectations for an assistant like that, right?

So if she doesn't deliver something you're happy with,

Tell it.

Have a conversation.

Make the thread as long as mine.

Don't hesitate.

Don't like the result?

Go ahead, tell it.

I don't like the result.

Can we change this and this?

It's the first mode of operation.

This is how most people use it, and I would encourage you to be a bit fearless with it.

It understands so much context.

You don't need to be some God-tier prompt engineer, whatever that means.

Just interact with it.

Talk to it like you would to a human.

That's the first mode of operation.

Now, let's slowly but surely transition to the second mode.

Oh yeah, true.

One more point.

1Having basic knowledge of no-code automations like Zapier make really helps.

This just like really improved my workflow too, because one part is like sorting the invoices, but the other one is getting the invoices.

This little three-step automation that you don't need any coding knowledge to set up,

Pah, blessing.

I hook up my email that all the invoices go to.

It only continues if it has an attachment.

And if it has the attachment, save it to this Dropbox file that my account has access to.

So simple.

You can learn this in 20 minutes on YouTube.

It's really simple and you can get things done.

Okay, so this is kind of the first part, right?

This is like solving things for yourself.

And also, I kind of just wanted to throw this in here.

Have you seen this?

US Software Developer Employment Index?

I mean, heck, if I would have to do all of this, the coding, I would need more than nine hours.

I would need to learn the different packages.

I would need to do so much debugging, it would not be worth it.

I would rather spend one hour every month for the rest of my life than spend three days learning all of this.

But I managed to do this.

So I don't know.

This is just an interesting kind of data point.

around development, but here's the second use case.

The second use case is you're not doing it for yourself.

You're trying to solve a problem like this for your employee.

Maybe your marketing department, maybe your software developers, maybe your salespeople, they don't have a clue about

you know, GPT-4 or even existing.

They don't know that there's other models out there.

They don't know you can use vector databases.

They don't know that you can do function calling.

They don't know that it can write code, it can execute code to do data analysis.

They just want the results.

What do you do there?

And arguably, this is...

Maybe even the more critical question.

Because you yourself can make yourself more effective.

Right now, probably not 10 times more effective, but probably like, I don't know, 20%, 40%, 60% more effective.

Depends on the use case.

But if you manage to do that for a team of 10 people under you, and maybe just five of them you make 20% more effective, that compounds so hard.

Let me tell you, running a little team of people, equipping everybody with little GPTs here and there for some repeating tasks,

makes a real difference.

And that 20% improvement across five people frees them up, reduces emotional stress.

This accounting thing is not about the one hour.

It's about the fact that it's the 15th, and I already know that by the 22nd I have to submit the invoices, and I'm already freaking out about it because I don't want to spend the time of it.

That is so much more valuable.

That peace of mind and calmness is so much more valuable than having an extra hour.

So if you're solving this for other people, what do you do?

My answer to that, on a consumer level, sure, on a higher level, we just saw fantastic talk there, what you would build.

You would build them out even further.

But how do you even begin?

How do you start building these solutions?

How do you start building these various agents that have access to your knowledge and function calling and that you will be upgrading over time?

You start in here.

They've kind of killed it, in my opinion, in really lowering the bar

so low that everybody, and I'm talking everybody, even my, okay, maybe not my grandma, because she doesn't speak English, but literally almost everybody can start building these simple solutions that make people's life easier, a bit easier.

Just think about it as in like 20% more effective.

Practical Demo: Building GPT Solutions for Others

Okay, here we get to the practical part, okay?

My proposed solution to building these solutions for others is start with what you have here.

Start with building a simple GPT because it is transferable.

If you're going towards Azure and if you're going to be using

you can take the instructions from your GPT and just copy paste them.

We had a suggestion, was it over here somewhere with like the AI adjusting to the tone of voice and kind of the emotional state of the person talking to it, right?

You can build that into your GPT here already.

You can test all of these things here already.

You can adjust the tone of voice.

You can adjust the interaction sequence, how it will work with a person.

But that all seems like a lot, right?

You kind of need to have all of these different building blocks in your mind.

You need to know exactly what it can do.

It's a bit much.

We actually took this, we've done it so many times that we took it and we built like this sort of like, I don't know, 15,500 character gigaprompt that kind of does it all for you.

And what we managed to do, actually just a few weeks ago, we managed to put it all into GPT.

So now,

And it's completely free.

It's completely free.

Now it's accessible to everybody since two to three weeks.

And we have this custom GPT builder.

If you want to just kind of rip it off, you can just say, download the attached files.

We've got the full prompt, and you can build your own.

But it's here, freely accessible.

You can get started here.

You can just find it on a GPT store completely for free.

Custom GPT builder, what does it do?

It builds GPTs for you.

And GPTs are these little specialized chatbots that help other people use ChatGPT, although they have no idea what they're doing.

How does it work?

Well, let's say maybe we want to build an onboarding GPT for our new marketing employees, OK?

Random example.

We could take others.

Just turns out these GPTs are really good at guiding people and teaching them new skills.

So I'll hit the only button that's here, OK?

So this is my live demo.

This is going to show you how to use this and how to build a GPT for your employees.

OK, start by clearly defining your chatbot's purpose.

A detailed description of the goal is key.

OK, so what is our goal here?

Educate my on

So you could make it as simple as this, but obviously, more detail helps.

So if we say, one, analyze existing accounts and our niche.

Two, brainstorm similar ideas.

Three, craft copy and

and generates a mockup, let's say.

Cool.

If this is my little process, what's up with that?

This is like an opening, I think.

OK.

If I just type in this, I could keep it as simple as just the first message, right?

But you just have to tell it what am I looking, what kind of process am I looking to simplify, right?

It can be as simple as this.

It can be as simple as the first few words.

I'm just looking to educate my new marketing intern.

And then what it does, what I built into this thing is it comes up with follow-up questions to kind of clarify what exactly you mean.

This is great because you don't have to know what to tell it.

You don't have to know prompt engineering.

You don't have to know the building blocks that go into it.

You don't have to know the different capabilities that are hidden with it.

For example, I built in one thing where it gives you interactive quizzes sometimes if it's appropriate for the GPT.

So for an educational one like this, it might be great to kind of

First teach your intern, but then tell it like, hey, would you like an interactive quiz to test your abilities?

This builder will do this and many more little hidden tricks like this for you.

It just builds it for you.

All you need to do is respond to these five clarifying questions.

In the interest of time, your time, my time, and efficiency here, there's a little feature here.

It's hidden behind this little arrow.

But for all of the lazy users, just like me, I included the auto feature.

So you can just type auto, and you don't have to answer any one of these.

It's just going to come up with answers for all five of them.

If you build it for your employees, colleagues, whatever, you might want to fill them out.

It's going to make it more relevant, right?

But in this case, I'm just going to say auto.

And it's going to answer all these five questions for me.

And there you go.

That's it.

We did it.

We built a personalized assistant for a specific use case.

It starts with a communication sequence.

How is it going to engage the user?

How is the interaction going to go?

Where is it going to start?

Where is it going to continue?

And then inside of the prompt that you can look at, we have various building blocks that define its behavior.

In here, you can see these are all the instructions.

We don't need to go into the details, but there's things like, look, offer interactive exercises and quizzes to reinforce learning and help the user practice content creation skills.

We reinforce the use of the browser tool, because it seems appropriate if we're doing social media, right?

It does it by default, but if you reinforce it with a building block like this, it just uses it more often.

What else?

We want a structured response with clear, numbered steps, right?

Because it's an intern and you want to make it as easy as possible.

That makes a lot of sense.

So it makes all of these assumptions for you, right?

You can give it some answers here and it will consider them.

But here, as you can see, it makes all of that.

And what is it?

It's an expert in social media marketing with a friendly and encouraging demeanor.

Clear and actionable guidance.

Love that.

And then last thing, like up here, the communication sequence.

How does it engage our user?

Welcome to the Social Media Mentor, your guide to mastering Instagram content creation.

Good, so this gives you a fantastic base version, right?

So what do you do?

You just copy this and you have your very first set of instructions.

If I wouldn't be talking here and explaining everything step by step, this would literally take you under a minute.

You just hit, let's build a GPT, you tell it what the goal is, you let it auto-generate the answers, boom, you have your first set of instructions.

And you can start building towards more advanced solutions like this.

with this very first step.

All you need to do is go to create a GPT.

For this, you do need a Plus account, right?

But then you can bring it over to any other LLM.

You can use it with Lama free.

Just send it as the first message also works.

But then you can put it here in the GPT.

Social media mentor.

And here you have version 1.0.

You're good to go.

This is the first thing.

You can start using it.

You can give it to your internal, try it out, see how it goes.

And then you can start enhancing it.

Here, you can give it a bit of context.

I'm not the biggest fan of the implementation of the knowledge base, I gotta say.

It's not as good as a lot of the rack solutions that uses vector DBs.

It's OK.

It works.

I always say up to 15 PDF pages, whether that's 15 documents with one page each or one document with 15 pages.

I won't go over that.

It starts getting very mushy about that.

But up until then, you can provide it with extra context.

You can also put it into the instructions.

Into the conversation starter, I always just throw in, let's start.

And then what you can also do is if you have other prompts, which I'm sure you're at an AI meetup.

It's 2024.

I'm sure your email is full of newsletters with various prompts.

Your notepad, I'm sure, is full of various prompts that you've saved away here and there.

Or at the very least, you know where to go to check them out, YouTube, Twitter, newsletters.

A lot of prompts out there these days.

They're kind of commoditized at this point.

But basically, if you have some useful prompts for yourself, and I provide them too.

If you check out my YouTube bunch of videos on prompting, just copy them from the description or from wherever you have them.

But if you have other useful prompts for your intern that you might want them to use, you can put them in here.

And this way, we really just set up our very first version of kind of an assistant for one of our employees in, I don't know, I would dare to say a minute, maybe two, right?

Now you can start customizing it, you can start like watching educational videos like mine or others and improving

what goes on in the instructions, you can kind of evaluate the results and watch the intern use it and see like, okay, maybe we should add this, or maybe we should add a second GPT that trains them on a specific skill, not just like how to handle Instagram content creation, but like overall.

But here you have the very first version.

You can copy the link, you can send it to them, and this is really, my opinion, this is the best way to start.

If you're at like,

Starting out, this is such a great way to set up a very first version of a personalized assistant.

Cool.

One more thing.

How much more time do we have?

OK, a couple of minutes is great.

Good.

So this is one part.

This is kind of giving it the persona, giving it the interaction sequence, a base framework for how to build these.

But there's a second part to it.

All the power in these LLMs hides within the prompts.

You need to speak results into existence, so to say.

You need to use words to evoke certain behavior.

And there's so much of it hidden in it.

You can go all the way from,

I don't know.

You can go all the way from writing email to, let's pick something really fancy.

What have I been looking at today?

To citing specific quotes from within academic papers while avoiding any hallucination whatsoever.

That prompt looks like this.

Another prompt is like three words, right?

But you can like do these things and everything in between.

You can summarize, you can, I don't know.

You guys know what prompts can do, right?

They're powerful.

There's a lot of capabilities here, a lot of specific use cases.

But there's a big problem I ran into when working with a lot of people.

Often they're inconsistent.

At the end of the day, this thing, where is their logo?

ChatGPT.

This little application, it's not the most consistent thing in the world, right?

At the end of the day, it's a very fancy random number generator.

Very, very fancy.

But at the end of the day, there's still this randomness aspect to it, right?

Token generation is...

always a bit unpredictable.

Hallucination is a feature, not a bug.

The creativity that you encounter within it, it's a good thing.

You want that.

You want it to think outside of the box a little bit.

But just like with humans, with all the creativity, there's also a lot of problems.

It's kind of a balance.

It's a give and take.

Now, what do you do about that?

If you give it a one-line prompt, it might work.

It probably works.

If you're in the first bucket, if we're in the first half of my talk here, and we're just talking to the assistant and just generating something, I'm there to give feedback.

If I don't like the code, if I don't like the result, I'll tell it.

It will readjust.

But what if I give this to my intern?

What if I give this to my grandma?

What if I give this to a coworker?

Gonna have a harder time.

Even worse, what if I automate it away and never look at it again?

What if I insert a prompt into one of my automations and it just runs in the background and the next time I check it's in like five months?

Becomes a bit of a problem.

There's no way to double check it, right?

Like the point of automation is just setting it up and ignoring it.

What do you do here?

You extend the prompts.

You flesh out all the context that is relevant to it.

So for example, I'll give you a quick example.

And this is basically what prompt engineering is.

It's understanding what capabilities you have within these tools, what context might be relevant to make your tasks, your instructions more effective, and then being able to communicate that context.

In other words,

When a new employee comes into your office, what would you do first?

You would introduce them.

You would give them a guidebook.

You would introduce them to the other workers there.

You would spend some time with them, or somebody in your team would spend some time with them, educating them, making them accustomed.

It's completely normal in a company.

When you work with ChatGPT and you use this framing of a new employee, a lot of people go in here and are like, write me an email.

It's like, about what?

Who are you?

Excuse me?

What's even going on here?

This thing is set up to do its best job possible and to not resist you.

So it's going to write you that email, but it doesn't have any information whatsoever.

The opposite end of that spectrum is when you give it all the info.

You give it a prompt where you tell it.

hey, I'm Igor.

I'm this age.

These are my goals.

These are my values.

This is the way I work.

These are the projects I work on right now.

This is what I want you to do.

And then you give it information on the company.

You give it information on your writing style.

You give it information.

Heck, you give it the last five emails that you wrote so it knows.

If you flesh out all of this, it starts becoming longer than a free word or a one-line prompt.

But what you get and what you unlock is consistency.

So if you're using it by yourself, and if you're there to feedback the responses, great.

No need for that.

Honestly, a lot of people sell it in a way where it's like, I'm going to teach prompt engineering.

I'm going to teach you how to write these prompts.

And then you can do it out on an everyday basis.

And it's going to take two hours instead of one, which is so counterproductive.

On everyday use, the way I use it, the way I showed you is you just start typing, you start talking to it.

It's great if you learn about prompt engineering, you become aware of some of the hidden drawers and compartments that hide within LLM like this.

Great, you can start choosing that.

Nevertheless, my everyday interactions, I don't copy paste big prompts much.

I just start typing and I start talking to it like a human and that works.

If you're doing it for others, if you're automating it, you want to flesh out the context.

You want to give it examples.

You want to tell it about the tone.

You want to give it all relevant context, just like you would give to an employee, especially if you're automating stuff.

You're not going to be checking that stuff.

So how do you do it?

Prompt Engineering for Consistency and Automation

This is going to be the end of my little talk here.

We built another GPT.

This one is a bit more intricate.

This one took quite a long time to develop, but basically,

Also here, you can just say we attached everything to the knowledge base.

So you can just say download document, and you can look at the entire prompt engineering in the background.

But basically, this is simple.

You can go just to quick build, and this will help quick build a prompt for you that is ideal to delegate or to automate.

So this you wouldn't be using if you're using this for yourself.

This you would be using if you're going to put the prompt in automation or if you want consistency.

If you're going to run the prompt 50 times and you want it to be as consistent as possible, that's where you want to be using this tool.

So I'm just going to say to my marketing department,

Or maybe let's say sales department.

That's even better, right?

Maybe I'm the type of boss that once a month just writes an angry email to my sales department, firing them up a little bit.

And I want them to be kind of consistent.

I don't know, just random example I came up with right now.

Then this is preferable to just having.

So let's have a look at this.

Everything it writes in here is preferable to just having a one-liner, where it's like, the original prompt would be something like this.

Write an email to my sales department telling them that they are not hitting their numbers, right?

This would be the basic prompt.

This is the fleshed out version, okay?

With no work on your part.

It includes all the techniques that you would want in this context.

like placeholders for five-shot prompting.

It includes all the relevant context, like the fact that it's going to be a draft that you might want to work with an outline, include a motivational message that emphasizes the importance of teamwork and collective efforts.

These are all things that are kind of non-obvious, but very useful if you're going to be using this consistently.

And all you have to do is you have to copy this.

And when you use it, make sure to customize the variables.

So everything that's in brackets like this, you need to customize, or you can also delete it.

That's OK.

So here you have an example.

Here's an action plan.

Utilize the CRM.

Just double check all of this, and you're ready to put this into automation.

This works so much better than simple one-liners if you're going to be building automations with LLMs in the middle.

And with that, I believe you're a bit more equipped for this new age than you've been 20 minutes ago.

You know how to craft personalized assistance for others.

You know how to use this on an everyday basis for yourself.

Don't stress out over intricate prompt engineering or whatever.

Just talk to it like a human.

And you know how to craft prompts in one or two clicks that will stand the test of time and the test of repetition.

And with that, free new skills.

Conclusion

I hope this helps, and thank you.

Finished reading?