Human-AI Workflows That 10x Productivity – with Iliya Valchanov from Team-GPT

This talk addresses the creation of human-AI workflows, emphasizing the importance of adapting to advancements in AI technology, particularly following the release of GPT-4. The speaker, who has a decade of experience in AI, provides insights into ten different workflows that can enhance productivity. He argues that understanding how to prompt AI effectively is crucial, but the context provided to the AI often matters even more than the prompts themselves. The session illustrates various applications of AI in generating content, automating tasks, enhancing marketing campaigns, and the development of AI agents that can perform multiple functions autonomously. By sharing real-world examples and tools, the speaker encourages the audience to embrace these technologies and integrate them into their own processes.

Hello. Hi everyone. Thank you so much. Thank you for the lecture before mine. This is a problem we currently have and it was really useful for us. Uh I'm really excited to speak here today. Thanks for the invite. And uh today I'm going to talk about hum uh building human AI workflows. We're going to explore about 10 workflows. We don't have much time so I'll jump into

it. I've been in the field of AI for about 10 years and uh mainly teaching people how to how to train models and how to use these models. Uh now sorry for the clicker what to expect from this presentation. First we need to get on the same page why AI is so important right now and then we're going to explore 10 workflows. I hope at least one

of them is uh useful to you. Okay, maybe this way. Okay, so first of all, Chpt, so I know that AI is the real deal for the last 10 years. And then each year more and more people were finding this out. The big catalyst was when GPT4 came out. That was in 2023. And since then you can see that for about a year new models were coming

out and nobody nobody cared like nobody nobody was using it. Nobody cared. And uh here is when GPT4 came out. GPT4 Turbo 404 mini here the price started going down and the usage started going up. 01 preview and so on. Right now we are here and it is expected that by the end of the year 1 billion people will be using chat GPT for their work whoever

so it's really really important that you start using AI before everybody else before the market saturates now why is this super duper important so in economics we call this a technology shock a technology shock is a time when with the same amount of labor people can do much more output. They can get much more things done. You know, important milestones were the industrial re revolution when electricity

came out. Just imagine pre-ele electricity after electricity times. Then we had the internet. Now is an inflection point. It's probably the the only one in our uh uh working lifetimes after the internet. Well, I was like five years old back then. But it's a very very important time in humanity and question is will you take advantage of the technology shock and start getting a lot more done.

Now this is already not so hard and we're going to explore 10 examples. Let's start with something super super simple. Um this was a gift. It's not moving. I have a couple of these but um we'll talk about it later. Now, when you plan a campaign, I'm sure everybody knows this one. We're starting really slow here, but uh you want to do a campaign, you want to

get Chad GP to tell you how to do it, and then there are different ways to to make it happen. Now, a good way to do it is to write this super super big prompt. And the better the prompt, the better the outcome, right? But the thing is a prompt like this is very hard to create and actually prompting is a lot of work and uh I'm

sure that many people are having issues with this. So the question is is there a better way? Is there a better way to create prompts? This is what we asked ourselves a couple of months ago. And the question was, okay, if I want to get uh this campaign in the best way possible, I first need to get this prompt as good as possible. And then there are

different ways to do this. Maybe you can create a prompt which creates prompts. Maybe you can use a product which creates these prompts for you. And this was a gift as well. But uh the thing is there are different techniques to do better prompting. Since I don't have a video, one of the best ones, one one of the best techniques is to say, let's say, I want

to create a marketing campaign. Ask me several clarifying questions. Ask me five questions about this campaign. And then the chat starts asking you questions. You start answering the questions. And you're without knowing it, you're building a very very good prompt. So the first thing so so the first thing to getting more out of the AI is to be able to prompt it right sounds logical however it's

not about it can't be all about prompts it can't be the future can't be us writing these long prompts and the question is what is the uh the right way to do it so it turns out and this is one uh so I posted yesterday I'm looking forward to speaking at the balcony e-commerce submit. I wrote this post. Actually, AI wrote it. Usually, I write my own

posts, but the promotional ones uh I give to the AI. And the question is, how do I get this? And how much do I prompt to get this done in a good way? In fact, uh I will send you the presentation later on. I know that it's hard to read, but this is my prompt. And the prompt is something along the lines of I'm speaking at the

Balcon E-Commerce Summit. I'll post this image with it. My topic is this one. And this prompt got me this amazing post which is very very in line with the way I write and what I'm trying to achieve. Now the the main thing is that there are products including ours but it's uh this doesn't matter. There are products which allow you to add context in a different way.

So instead of prompting you can give it other stuff. In our case we give it good examples. So I I have trained it um which is a wrong term but makes the point across. I give it a bunch of my posts and then with a very simple prompt I can get a post which looks exactly like the way I write. So this is very very important because

it shows us that the prompt doesn't matter at all. actually uh scientific research shows that with good examples there is nothing better than good examples when trying to get something done. So if you provide a couple of examples and the simplest stupidest prompt like do something, it will have a better result than this big ass prompt. this prompt will always perform worse than a tiny prompt with

a couple of examples of what you want it to look like the output. So, so it's not about the prompt. Actually, everything you get out of the AI, it's about the context you give it. and think about context as examples, images, PDFs, videos, your code base if you're coding your brand voice, links to stuff, uh your Google Drive, notion, all of these things. So, the more stuff

you you stuff it with, the better it's going to perform. Now, uh you can't give videos as context right now, but you will be able to give videos as context in like two years. Like, this is inevitable. uh the thing is we will be able to stuff more context and the question is can you find the right context to give it so that it can uh get

something done. So then I go to the next one, detailed descriptions. I'm going to use the description for this event. Uh I got this email. I was out of office and they asked me, you have to describe your talk. My talk was not ready yet. So I didn't know what I'm going to write uh talk about and uh I wrote this short description which is completely AI

generated. How did I do it? Well, I I used again our product but this is irrelevant for this. So I said my topic is going to be building human AI workflows. I gave a link to the Balcon e-commerce conference and I gave a notion page which is our preparation for e-commerce companies and then I went to the chat with this context in mind. I sent a screenshot

of the email. I didn't prompt it. I just sent a screenshot of the email saying can you send me a short description? So the the prompt is already in the image, right? And then I said, well, these are my ideas and it came out with um the description I sent you. Took me one minute to do it because this was already set up for me for a

different reason. Now the most important thing here is I didn't even give it an instruction. So in the first example, we prompted heavily. In the second example, we prompted slightly but gave the right context. In the third example, we didn't prompt at all and it worked. Now, this will be coming better and better, but it's all about the context, right? Uh uh by the way, this works

for products, bio submissions, applications. I have done like hundreds of submissions in this way. I I no longer like think about these. Now the thing is I if you know the context if you know what you're trying to achieve if you know what what what are the final results maybe there are even better ways to do this and u this was a video as well but um

uh this was showcasing one tool we created for ourselves with Maggie we were um so what we wanted to do was we wanted to watch YouTube videos very fast so you have a one-hour lecture from last year uh balcony e-commerce conference whether to enter the Hungarian market. I want to know what these guys are going to say. Now, we have built and since there is no video,

I'll just show you. We have built a tool where I paste the video link and then it takes the video, summarizes the video and I can ask questions to the video. So, basically I can simply uh take a onehour video and say what is this video about? What are the important bits? watch what are the challenges to entering the Hungarian market or whatever and then I would

get the exact points in the video where I can click watch this bit and go go further. Now this is done again with the same technology. So how we do it is we take the transcript of the video as context and we just search where it is. Now you would say that's cool but I don't have a product team right like uh but no problem generative AI

is much much bigger than prompting with the same technology you can build websites tools apps very quickly very easily. So over here I built the Balcon E-commerce Summit 2026 website. It um it assumed it's going to be in Belgrade uh but this is untrue I'm sure. Uh so this website is actually live right now. So how did I create it? It it again it was a video.

It's a whole website speakers venues all of these things. So how was this created? Create a simple website for Balcon e-commerce conference and then I I gave it the current website and it created the new website with completely new stuff. Now the the moral of this story uh for for this workflow is if the context is known and the context here is create a website the tool

in this case lovable it know it has seen all the websites he's seen all the code bases and it's very good at creating websites and apps which means that okay ch is just the beginning and then there's a myriad of of apps that are doing completely different things which are very very powerful Now the next thing was um and I was like okay what if you work

with images? I guess you guys work with a lot of images. You've heard that there is a new image model. Uh yeah I don't have time to give you more context but to show you what this model is capable of. I was on the flight back home and I was I took this Iran Iran and I said create an image brand this Iran with the photo from

the balcony e-commerce conference. This is the prompt and this is the iron. So the product ch is already doing these things with the stupidest prompts ever. Right now as you can see my face is a bit different. So it changed my face as well. But this is because chpt is not meant to to do this right. It's it's charg um uh there is this product photo room

that released this product. There is an image of this uh cup and then it puts it in real life situations with real models. So now you can see that products which before you needed to do photo sessions and all of these things this costs nothing. This is almost free to use and you can now put this in real life situations but with something more than just GPT

like you can't use the just openg because they don't care about your product images right they care about showcasing the tick now what if you're doing a lot of ads well there are tools for this tool and uh this one is let's say this is your competitor's ad well there is this tool which takes your competitor ad and brands it like your product, puts your product over

here and then it creates a a 100 different variations of ads. You can just provide it with your competitor's uh link and it will find the ads, search the ads, take the ads and then create variations of these. You can do this every day on auto repeat with these uh products. This product, by the way, is called Icon. They just launched a couple of months ago. Uh

I haven't used it but I have um played around it uh a bit. Now full automations this one you know you know for sure. So you've definitely used some form of Zapier make if public non or so on. So the thing is uh these workflow automations before they were dubbed with this name if this then that and the idea was if this happens then do this other

action. These were simple workflow automations. Right now the hot thing on the block are AI agents. And how are agents different? Well in a workflow you can chain a bunch of steps. So you can do a bunch of steps one after the other. But some of them could be subjectively decided by the AI. For instance, go to my competitor's website before meant scrape the website. Now it

can mean go to my competitor's website, understand what it is about, understand the key value propositions, extract them and then create a website for myself. So this is why we call them agents because they can do a bunch of stuff autonomously which before were not possible. Now for the purposes of this lecture, you can take all the previous steps, chain them together and you have an AI

agent, right? And this this is what many companies are doing. I know I know many companies that are already doing this. The there are issues with the agents. They're not dominating everything yet, but they will eventually. This will be the way um stuff are getting done. Now, how do I know the all of these things? Well, this is my job. My job is to work on this

product and um build a future of work product. Now for us it is extremely important to build these new experiences. So we're constantly thinking about okay right now prompting is a thing but what what will it be in one year? What will it be in two years? What are the real problems that need to be solved? So far we have about 50,000 users and bit by bit

we realized that actually and this is still a secret we uh have not announced this yet um we most of them are marketers so actually marketers are getting the most out of our software and they're coming from SMBs agencies and startups. This is kind of the the distribution. So we we kind of decided okay so marketers they love our product even though we were constantly focusing on

enterprises so why not stop focusing on enterprises and start building for marketers. However we maybe are good at marketing but we're not really good at at it. We don't do it every day. We build products every day. So what we're looking for is we're looking for marketing teams who are really excited about the technology who want to build uh something. So who who have workflows they want

to automate and if this is you by the way please come and send me an uh DM interested. The thing is we are looking for design partners. We're looking for someone to bounce ideas off. We want to solve your workflows because this means we solve it for everybody. And we're going to choose a couple of companies that are excited about it and start working with them to

build these new experiences which currently do not exist. Um if you're interested, please um let me know. And we're at time I guess. So thank you so much for this and see you around.