305 | Mastering AI Automation: Custom GPTs to Agents with Isar Meitis

1 - > Speaker: Hello and welcome to the Leveraging AI podcast, the 2 - > podcast that shares practical ethical ways to leverage AI to 3 - > improve efficiency, grow your business, and advance your 4 - > career. 5 - > This is Isar Matis, your host, and I have a interesting episode 6 - > for you today. 7 - > There have been growing rumors in the past few months that 8 - > OpenAI will cancel custom GPTs. 9 - > Now, those of you who don't know what custom GPTs are, well, then 10 - > you're not missing much.

11 - > But if you did not use any custom GPTs so far, then you 12 - > were missing maybe the most capable automation tool that AI 13 - > gave us until the existence of Skills and agents. 14 - > And many, many, many people, myself included, has multiple 15 - > custom GPTs which they were regularly using to run multiple 16 - > things in their business. 17 - > I switched completely to skills and agents, but I know many 18 - > people who have not. 19 - > Now, I haven't seen any real information from OpenAI about 20 - > custom GPTs going away, but if you're following what's 21 - > happening on X or on Reddit, then you would see that there 22 - > are many conversations talking about custom GPTs going away.

23 - > And if you are somebody like me, who until not too long ago were 24 - > running a lot of aspects of my business through custom GPTs, 25 - > then it is something that you need to be aware of and start 26 - > preparing for. 27 - > Either way, whether you have built custom GPTs that you're 28 - > using regularly or you don't and you wanna know how to build 29 - > different kinds of automations with AI that can run more or 30 - > less everything in your business, this would be a great 31 - > episode for you to learn the basic concepts because they're 32 - > the same across the board.

33 - > So let's get started. 34 - > I'm going to be sharing my screen, but those of you who are 35 - > just listening and not watching this, I will explain everything 36 - > that is on the screen. 37 - > So if you are driving or walking your dog or doing the dishes or 38 - > something like that, yoga, and you can't watch the screen, then 39 - > that's perfectly fine. 40 - > But if you do wanna watch the screen and you are able to do 41 - > this now, you can either do this on Spotify.

42 - > We're sharing the videos on Spotify as well as on our 43 - > YouTube channel, and there's a link to the YouTube channel on 44 - > the show notes. 45 - > But let's get started. 46 - > So there are several different levels on the ways you can work 47 - > with AI, whether it's ChatGPT or any of the others. 48 - > The very basic one is chat, right?

49 - > You can just go to the chat window of Gemini or Claude or 50 - > ChatGPT, et cetera, and just chat with it, and then you give 51 - > it instructions or questions, and then it does what it needs 52 - > to do. 53 - > You are heavily involved in the process. 54 - > Option number two is to create custom GPTs or projects. 55 - > These are a consistent setup with a bunch of instructions and 56 - > additional supporting information that becomes 57 - > reusable and allows the AI to follow an existing set of 58 - > instructions time and time again, without you having to 59 - > give it any additional instructions.

60 - > We're going to see examples shortly. 61 - > The third level, which is the newest of the three, and if you 62 - > want from an evolution perspective, we had custom GPTs 63 - > first. 64 - > That was something that OpenAI invented. 65 - > Then we got projects from Anthropic, which were then 66 - > copied by OpenAI to have projects inside of OpenAI.

67 - > And then we received gems from Gemini, which is a very similar 68 - > concept. 69 - > And then we got skills and agents and so on. 70 - > So skills was after, long after we had projects, but skills 71 - > allows you to make it portable and reusable automation 72 - > components, building blocks that you can use in multiple places. 73 - > You can use skills in other platforms.

74 - > So if you are using, as an example, ChatGPT in Excel, there 75 - > is an extension for that, or ChatGPT in PowerPoint, there's 76 - > an extension for that. 77 - > You can take your skills with you. 78 - > So if you have created skills for specific processes in 79 - > financial analysis, you can bring them from ChatGPT into 80 - > ChatGPT in Excel, and if you have a specific branding or 81 - > process or style that you're using and you created skills for 82 - > that, you can bring it into PowerPoint as well.

83 - > And then on top of that, there are obviously agents where you 84 - > can build agents. 85 - > And to be fair, I'll do a little distinction on what agents are 86 - > and what they're not, because there's a lot of chatter, and 87 - > everybody calls everything agents right now when it's not 88 - > really the case. 89 - > So agents are employees, right? 90 - > It's an entity that can do a full complete workflow and not 91 - > just one simple specific task again and again.

92 - > And they can do it end to end, and they can do it in a 93 - > consistent way while taking into account the information they 94 - > have and they do not have. 95 - > I'm not going to cover this today. 96 - > We're going to stop at skills because I wanna show you the 97 - > basic building blocks that you can use. 98 - > Uh, now, a skill is basically a very simple agent if you want, 99 - > and that's why a lot of people call them the same.

100 - > But an agent could be a lot more sophisticated, have access to 101 - > more tools, be connected to more systems than just a simple 102 - > skill. 103 - > So these are the different options that we have in front of 104 - > us today. 105 - > Now, before we dive into the different options, I wanna kind 106 - > of show you the process that I went through, the evolvement 107 - > that I went through through this process, through the lens of 108 - > writing proposals. 109 - > So today, I use what some of you see on the screen, which is a 110 - > very sophisticated proposal pipeline agent that has multiple 111 - > skills, that is connected to multiple systems in my 112 - > ecosystem, and that creates amazing proposals every time I 113 - > finish a conversation, uh, that is relevant in which a proposal 114 - > was requested or discussed.

115 - > It does all of that automatically. 116 - > It updates my CRM, it updates my Google Drive, it updates my 117 - > outbox with a draft email while attaching the proposal. 118 - > It does research in the back end on the customer. 119 - > It does a lot of other great stuff.

120 - > But this is not how I started. 121 - > How I started was a custom GPT. 122 - > So now let's look at what a custom GPT is. 123 - > If I go to ChatGPT Initially, and this is one of the reasons 124 - > why people thinking that custom GPTs are going away, custom GPTs 125 - > lived on the left menu right above where the conversations 126 - > are.

127 - > And now you can see it's not there. 128 - > Like, if you look for custom GPTs, you won't see them. 129 - > So where are they? 130 - > If you look on the left side menu, you have where you have 131 - > New Chat and Search Chats and Library and so on, there's the 132 - > three dots that says More.

133 - > If you click on More, you will see, uh, GPTs, and if you click 134 - > on that, you will see all your GPTs over there. 135 - > So they made it less intuitive to get to or create, and if you 136 - > don't know they exist, they just will disappear on you, and you 137 - > won't even know that they're there. 138 - > That gives you a hint that this is not their focus and not the 139 - > focus of their product. 140 - > Now, how does this work?

141 - > You can see that this is called Multiply AI Training Education 142 - > Proposal. 143 - > That's the name of this custom GPT. 144 - > All I have to do is I have to go to Add a File and drag in a file 145 - > of a conversation I had with a potential client and once it 146 - > loads the file, all I have to do is click Go. 147 - > I won't type anything, I won't say anything.

148 - > I literally just upload the transcript of a conversation 149 - > that I had with a client that requested a proposal, and it is 150 - > going to work, and it is going to write a great proposal. 151 - > And you can see it's already started writing, and it has an 152 - > introduction, and the objectives and why AI training is urgent, 153 - > and why to work with Multiply, which is my company where I 154 - > provide training and education, and training formats recommended 155 - > for this particular company based on their needs, and so on 156 - > and so forth, and pricing and the whole thing.

157 - > It writes about a eight to 12-page proposal, 158 - > well-structured, and I'm winning a lot of business, so these 159 - > proposals actually work pretty well. 160 - > Now, this is how I started. 161 - > It worked very, very well. 162 - > When projects started arriving, I switched it to projects, and 163 - > we're gonna talk about later on why, but at first I wanna show 164 - > you just so that you see the different options.

165 - > So now if we go in this particular case, this is a 166 - > Claude project, but inside the Claude project, what you can see 167 - > is the same kind of thing. 168 - > I uploaded the transcript. 169 - > I asked,"Please create a proposal based on my brand 170 - > guidelines," and it created the proposal. 171 - > One thing you can see immediately inside the project 172 - > that we're seeing right now is that it has my logo on top.

173 - > It is using my brand colors. 174 - > It is using my fonts. 175 - > It is, from a structure perspective, has a much better 176 - > structure from a document perspective. 177 - > The benefit of doing this in the old school way without this is 178 - > that I can edit the file.

179 - > Here I cannot edit the file because it is a Claude artifact, 180 - > which is not editable, which is really annoying, but that's the 181 - > case. 182 - > Uh, but it does come as a fully branded output, which now you 183 - > can do with skills inside of custom GPTs as well. 184 - > So what we have right now is we saw that y - I can do this in a 185 - > custom GPT. 186 - > I can do the same exact thing in a project, whether a project in 187 - > ChatGPT or a project in Claude or a gem in Gemini.

188 - > All of them will do exactly the same thing, and they will write 189 - > solid proposals And then like I said Then I switch to the full 190 - > skill-based agentic pipeline that has all the other functions 191 - > as well. 192 - > So now I do not use the custom GPT or the projects, but it 193 - > doesn't mean they're not good. 194 - > It doesn't mean that I couldn't use them. 195 - > It just saves me more time and does more things in an automated 196 - > way, and so it saves me more time while creating a better 197 - > output.

198 - > And so this is why I'm doing this. 199 - > But what I wanna show you today is different options so you can 200 - > pick the one that is best for you. 201 - > So first of all, let's talk about what they are. 202 - > All of these things, custom GPTs, projects, and skills have 203 - > the same exact DNA, and that same exact DNA is they have two 204 - > components.

205 - > They have a set of instructions that tell them what to do, and 206 - > there is context, meaning knowledge files and additional 207 - > information that the AI may need in order to create the output 208 - > that it needs to create. 209 - > So in the proposal example, this could be a proposal draft, or if 210 - > you want a template that it can use. 211 - > They could include a winning proposal or a few that has won 212 - > me business to show it what good looks like. 213 - > It could include descriptions of the services that I provide or 214 - > the products that I sell, so the AI understands what it can pick 215 - > from when it writes the proposal.

216 - > It could include links to my website or testimonial websites 217 - > or testimonials on LinkedIn that it can pull from in order to use 218 - > when it's writing proposals. 219 - > So all of these things are included as reference material, 220 - > as context for each and every one of the three. 221 - > It doesn't matter if it's a custom GPT, a project, or a 222 - > skill. 223 - > So they're very, very similar, and that's why when you think 224 - > about, let's say you do have 20 or 30 custom GPTs, and you're 225 - > afraid maybe OpenAI will take them away, converting them into 226 - > projects or skills is a very small effort because they use 227 - > exactly the same things Before we dive further, let's talk 228 - > about instructions.

229 - > AI reads instructions from multiple places every single 230 - > time you talk to it, especially in these kind of places. 231 - > So first of all, there is the system-wide, if you want, the 232 - > account-wide always-on instructions. 233 - > In Claude, you can change your claude.md, which is the file 234 - > that Claude reads at the beginning of every single 235 - > conversation, and you can tell it exactly how to work with you 236 - > across the board in every single interaction it has with you.

237 - > In ChatGPT, there is the custom instructions that live inside 238 - > the settings on the bottom left corner, and you can go and tell 239 - > OpenAI how to work with you, again, on every single 240 - > conversation. 241 - > Then inside a project, if you created a project, there are 242 - > project-level instructions. 243 - > Uh, same thing with a custom GPT. 244 - > So you can give it instructions, and again, I'll show you in a 245 - > minute where exactly you do this in the system itself.

246 - > Then there in the Skill, there is Skill details, which is where 247 - > it writes the instructions for the Skill, and there are two 248 - > levels of memory in addition. 249 - > One of them is the general memory, which remember things 250 - > about you, and you probably noticed that ChatGPT or Claude 251 - > already knows stuff about you. 252 - > It knows where you work, it knows what you do, it knows your 253 - > hobbies, it knows everything you talk to it about, but it's not 254 - > at a very granular level.

255 - > On the granular level, a, it is connected to the specific 256 - > project. 257 - > So every project is a little bubble of context that remember 258 - > things about that particular project. 259 - > And a project could be, well, a project, that kind of makes 260 - > sense, but it also could be a specific client. 261 - > So you can keep a specific client information in a project, 262 - > have a separate project for every client, and have all your 263 - > conversations about that client in that project, and the project 264 - > is gonna remember more and more details about that particular 265 - > client, about the people who work there, about proposal you 266 - > sent them, and so on and so forth.

267 - > Uh, and so there's these two levels of memory. 268 - > There's the general memory and the project-specific memory. 269 - > So all of these things are places that the AI will reach 270 - > out to, depending on what kind of conversation you're having, 271 - > in order to get additional information to be used Now, what 272 - > I want you to remember that applies to each and every one of 273 - > those things here is that AI is an intern. 274 - > It's not just an intern, it is the best intern on the planet, 275 - > right?

276 - > It will do amazing things. 277 - > But just like you won't bring an intern into your room and say, 278 - > "Hey, I want you to write me this report," or,"I want you to 279 - > create this proposal," or,"I want you to create the summary 280 - > of one, two, and three," because it will probably fail. 281 - > And if it will fail, you will know it is your fault because 282 - > you didn't tell the intern exactly what to do, and the 283 - > intern doesn't know you, he doesn't know the company, he 284 - > does not know the industry.

285 - > He or she is an intern. 286 - > And AI is exactly the same thing. 287 - > So the instructions you're gonna give it, regardless whether you 288 - > give them in a custom GPT or a project or a skill, et cetera, 289 - > are the SOP, the standard operating procedure on how to do 290 - > a specific process. 291 - > The knowledge files that you're gonna give it is the binder 292 - > you're going to give the intern in order to know everything they 293 - > need to know.

294 - > So when you sit with the intern, you're gonna show them,"Okay. 295 - > Here are the previous proposal we gave to this client. 296 - > Here's information you can find about this client. 297 - > Here's where on SharePoint you can find information about the 298 - > services that we provide.

299 - > Here on these Excel files, you can run the pricing and get to 300 - > the..." Like, that's what you're going to do, and you need to do 301 - > the same thing for the AI, and then it's going to do an 302 - > amazing, amazing work for you every single time. 303 - > Now, the project, if you want a ChatGPT or a Claude project, is 304 - > a dedicated workstation, right? 305 - > It's, it's the table that has the computer and all the files 306 - > and everything you need in order to get access to the right 307 - > information.

308 - > And the skill, if you want, is a laminated SOP card. 309 - > It's something you can take with you if you want a suitcase with 310 - > everything you need, so you can go to other offices and do the 311 - > work over there, and you'll see why I'm saying that once we talk 312 - > more about what a skill is. 313 - > But it is basically a project or a custom GPT that is fully 314 - > portable and can work from everywhere without going 315 - > specifically to that folder. 316 - > So let's dive a little deeper to the first two options of a 317 - > custom GPT versus a project.

318 - > And again, if you'll see that they're very, very similar and 319 - > that you can switch between the two very, very quickly. 320 - > So if custom GPTs will go away, your first immediate line of 321 - > defense is switching your custom GPTs into projects. 322 - > And as you will see in a minute, there are benefits to actually 323 - > using projects over custom GPTs. 324 - > But both things will do roughly the same thing.

325 - > They will get an input, they will run through a process, and 326 - > will give you an output, and they will give you a consistent 327 - > output every single time So we looked at the example before of 328 - > the proposal, but what I wanna show you right now is in 329 - > addition to showing you the outcome, I wanna show you how it 330 - > actually works. 331 - > So if I go back to the custom GPT on the top left corner, 332 - > there is the name of the custom GPT that has a dropdown menu, 333 - > and you can click on Edit.

334 - > If you click on Edit GPT, which is exactly the same screen 335 - > you're going to see if you create your own, which I'll show 336 - > you in a minute how to do, you're going to see a similar 337 - > screen. 338 - > What you're going to see is on the left side, you're going to 339 - > have the name, the description, the instructions, the inputs, 340 - > the knowledge base, the actual graphs you attach, and some 341 - > other stuff such as recommended model that you can pick and 342 - > capabilities that you can choose.

343 - > And if you really want, you can add code on the bottom. 344 - > And there are conversation starters where you can add 345 - > buttons to allow other people to understand how to use your 346 - > custom GPT. 347 - > But let's dive into the most important things. 348 - > The name just gives it a name, tells it what it does.

349 - > So in this particular case, it's Multiply AI Training and 350 - > Education Proposal. 351 - > The second thing is the description, writes proposals, 352 - > blah, blah, blah. 353 - > The description doesn't really matter. 354 - > It's for you, and if you share it with others, for the people 355 - > you share it with.

356 - > The most important part is the instructions. 357 - > The instructions is where it tells you what it's actually 358 - > doing. 359 - > And you can see you're an expert proposal writer. 360 - > Your goal is to write clear, easy to read, and follow an 361 - > attractive AI training proposals.

362 - > And then it explains what the inputs are going to be, and it's 363 - > going to explain exactly what the process is to take those 364 - > inputs and analyze them, and then it tells it exactly what 365 - > the output needs to be. 366 - > So this is how it runs. 367 - > It references two separate files in the proposal, in the 368 - > instructions. 369 - > One is the Multiply AI services brochure, where it can take 370 - > information about what are the services and describe what's the 371 - > value in doing or taking these services from us.

372 - > And the other is a master template AI for training and 373 - > education of AI. 374 - > Why does it need that? 375 - > Because then it knows how to draft it from a formatting and 376 - > flow and structure perspective. 377 - > And this outline is really a very full comprehensive outline 378 - > of a proposal I will never write.

379 - > It describes every single thing that I deliver, which nobody 380 - > ever orders all of them, at least not all at once. 381 - > And so What it tells it in the instructions, it says,"Each 382 - > proposal that you write will include one or more of these 383 - > training options. 384 - > You need to only include the components that were discussed 385 - > with the prospect based on the transcriptions and/or emails I 386 - > will provide you." Right?

387 - > So this is what it does. 388 - > It gets access to a lot of different options, and it knows 389 - > how to pick just the ones that are relevant based on the 390 - > conversation we had. 391 - > And that's it. 392 - > That's all you need.

393 - > That's the entire magic, and it will know how to write proposals 394 - > if you wrote the instructions correctly, and we're gonna talk 395 - > about this in a minute on how to do that. 396 - > In a very similar way, if you go to the project, the project has 397 - > the same thing. 398 - > So if we look at the project level, you'll be able to see the 399 - > files that are connected to it, as well as the... 400 - > You can see brand, uh, guidelines and, uh, clarifying 401 - > client proposal details, and so on.

402 - > And you can see here that it has instructions. 403 - > So inside the instructions, if I click on them, it is the same 404 - > exact instructions that we've seen before. 405 - > I literally copied and pasted the instructions. 406 - > So if we go back to the previous section, you can see it looks - 407 - > it's, has the same exact thing, the overview, the knowledge base 408 - > requirements, the core workflow, like all the different things, 409 - > uh, that were there before are also here.

410 - > And then you can attach different files to here, uh, 411 - > that you can attach, so it can use as references for the 412 - > proposal. 413 - > So very similar process, different tools, but 414 - > instructions and knowledge files, and then you can just 415 - > drag in the transcript or emails or whatever you define for it as 416 - > inputs, and it will know how to do the work By the way, to 417 - > create new custom GPTs inside of ChatGPT, you go, as I mentioned 418 - > before, to the three little dots where it says More, you click on 419 - > GPTs, and on the top right corner there's a Create button.

420 - > If you wanna create a project inside of ChatGPT, they live 421 - > right above your chats, and you can see I have many, many, many 422 - > of them. 423 - > And in here, next to where it says Projects, if you put your 424 - > mouse over it, there's a plus button that will allow you to 425 - > create a new project. 426 - > So on both platforms, it is relatively easy to know where to 427 - > go. 428 - > Again, other than custom GPTs that are now hidden under the 429 - > More menu because I think OpenAI wants you to now build projects 430 - > and/or agents or skills.

431 - > So now let's continue with our flow. 432 - > Another example, by the way, that we have here, which I'm not 433 - > gonna dive into, but I will talk about in two seconds, is 434 - > converting data, raw data, like really large Excel files. 435 - > So if I open this, you will see that it has, multiple columns, 436 - > like dozens of columns, and then thousands of rows in this 437 - > particular file. 438 - > And this could be a sales report, this could be financial 439 - > analysis, this could be a scraping of customer, uh, 440 - > pricing, like whatever the source data is, and it turns it 441 - > into a very detailed report in the end that shows, in this 442 - > particular case, sales, and it has the table of content, and it 443 - > has an executive summary, and it has a business overview, and it 444 - > has a lot of other stuff that shows graphs and charts and 445 - > different information.

446 - > By the way, all the information in this one is completely fake. 447 - > It's based on a random generated, data, but the concept 448 - > is perfect, right? 449 - > You can see a very detailed... 450 - > Again, those of you who are seeing, those of you who don't 451 - > have to believe me.

452 - > It is a well-branded, well-structured, well-organized 453 - > report with graphs and charts and analysis of all the data 454 - > from the source file. 455 - > And this could be built either as a custom GPT or as a project 456 - > or as a skill, so that doesn't really matter. 457 - > So now between custom GPTs and projects, let's look at a quick 458 - > comparison table. 459 - > The purpose of both of them is roughly the same, right?

460 - > Is to bo- be able to have a conversation. 461 - > But the project has another benefit. 462 - > The project is more of a workspace or, like I said, a 463 - > context bubble where you can have free conversation. 464 - > So while the custom GPT is built to do a recurring task and again 465 - > and again and again, the project can do the same thing, but can 466 - > also allow you to just have any open conversation inside the 467 - > project while taking into account the memory of the 468 - > project and the information that was attached to it, while a 469 - > custom GPT doesn't do it, or at least doesn't do it as good.

470 - > You can attach files to both of them. 471 - > You can give instructions to both of them. 472 - > The amount of data you can give into each and every one is 473 - > roughly the same, depending on the specific plan you have. 474 - > There are Three big differences.

475 - > One is memory. 476 - > As I mentioned, custom GPTs, every single chat is a new chat. 477 - > It doesn't know anything about what happened in the previous 478 - > chats with that custom GPT. 479 - > Inside the project, there is a project memory where it's going 480 - > to learn more and more information as you have more 481 - > conversations inside the project.

482 - > The other benefit of running the conversations inside a project 483 - > is the conversations live inside the project, where you can see 484 - > each conversation that happened inside the project, one under 485 - > the other, while in a custom GPT, the conversations appear in 486 - > the regular conversations of ChatGPT, which means they're 487 - > gonna be hidden between the other 10,000 conversations that 488 - > you're having. 489 - > So if you wanna be able to see the same report that you created 490 - > last week, it will be extremely easy to do in the projects 491 - > because it's just gonna show there as the previous line item, 492 - > and in the ChatGPT, you'll have to do search and filter and find 493 - > the right one and check it, and so on.

494 - > So that's another benefit. 495 - > The only real benefit of custom GPTs is that you can add code at 496 - > the bottom of the custom GPT, which means you can connect it 497 - > to an API of external tools. 498 - > This is something you cannot do in projects right now. 499 - > That being said, with Skills, you can now do similar things.

500 - > And that being said, I probably have over 50 custom GPTs that I 501 - > created and used, and I created code in two of them. 502 - > So it's not something that I've done very commonly, and I doubt 503 - > that a lot of people did. 504 - > While it is a benefit, it is not a huge benefit. 505 - > Once I understood the benefits of projects, I stopped using 506 - > custom GPTs completely.

507 - > I converted the main ones into projects, not all of them. 508 - > And then I said again, I converted more or less 509 - > everything into Skills and Agents How do I create all of 510 - > them? 511 - > Whether I'm creating a custom GPT, a project, and/or a skill, 512 - > I'm always creating them starting with a regular chat. 513 - > I'm going into a regular chat, either in Claude or in ChatGPT.

514 - > When I do this in Claude, I do this in Claude Cowork. 515 - > When I do this in ChatGPT, it doesn't matter. 516 - > You can do this in the regular ChatGPT and/or in Codex, and you 517 - > can explain what you wanna do. 518 - > You can give it the files, and you just work through the 519 - > process.

520 - > You give it the data. 521 - > I said,"Okay, let's clean the data. 522 - > I want you to understand what's in the data. 523 - > I'm gonna give you this kind of data every single time."

Now you 524 - > iterate through the process. 525 - > You explain exactly how to get to the outcome you wanna get to, 526 - > whether the final report or the proposal or the analysis or 527 - > whatever it is that you're trying to do. 528 - > Once you get to that outcome, you can ask the AI, again, 529 - > whether ChatGPT or Claude or Gemini or any other, and say,"I 530 - > want you to turn everything we did right now, just the stuff 531 - > that worked, not the stuff that didn't work, into a X."

This 532 - > could be instructions for a custom GPT, instructions for a 533 - > project, or a skill, and it will know how to do that. 534 - > If it builds a skill for you, it will package it and will give 535 - > you an Install button as soon as it's done. 536 - > If it is a custom GPT or a project, you will have to then 537 - > take the instructions and paste them into a custom GPT you 538 - > create or a project that you create. 539 - > What you should tell the AI is which files you're gonna give 540 - > the custom GPT or project as references, so it can use it in 541 - > the instructions that it's creating.

542 - > So let's look at a quick example. 543 - > In this example, I was building a custom GPT that writes 544 - > questions for sessions of the courses that I teach. 545 - > Uh, those of you who do not know yet because you're new to the 546 - > show, I teach AI courses. 547 - > I either teach them online to people who can join or do 548 - > private workshops for companies.

549 - > Either way, I need tests, and I need ways to check that people 550 - > are learning the process. 551 - > And instead of doing this manually, I'm using a custom GPT 552 - > to do this. 553 - > Again, not anymore, but this is how I started. 554 - > So what I started here is I said,"Hey, I need your help in 555 - > creating a new custom GPT.

556 - > I'm going to explain to you what the needs are, and I'm going to 557 - > give you some examples, and then I will need your help in 558 - > creating the instructions for the custom GPT. 559 - > Is that okay?" And it said, "Absolutely. 560 - > That sounds great," blah, blah, blah.

561 - > And I said, Okay, so I'm creating a course about AI, and 562 - > what I have is the transcript of all the different sessions What 563 - > I will give you is examples of the transcripts of one session 564 - > and the questions that were written for it by humans, so you 565 - > get an idea of what kind of questions we are looking for. 566 - > And then I would like you to do is I would like you to explain 567 - > these questions in a way that the GPT will be able to create 568 - > questions when I give it new transcripts, and so on and so 569 - > forth.

570 - > You understand the point. 571 - > We went back and forth several times. 572 - > You can see if I scroll down, those of you who are watching, 573 - > it's a very, very long conversation. 574 - > It created the first set of instructions.

575 - > It didn't work well. 576 - > We tested it again. 577 - > We iterated, we compared it to other questions. 578 - > We basically went in several cycles of fixes, and you can see 579 - > it here about iterating.

580 - > It most likely won't work perfectly the first time. 581 - > You iterate several times, and once you're done, you ask it to 582 - > create the instructions, and then you can create the GPT or 583 - > project or skill. 584 - > Then you still have to test it. 585 - > There is a pro tip here about reusing the code.

586 - > What does that mean? 587 - > If you're doing financial analysis, as an example, every 588 - > time you're uploading an Excel to ChatGPT or Claude, what it's 589 - > actually doing it is writing Python code. 590 - > By clicking inside the thinking little thing, uh, when it's 591 - > blinking, when it's doing its work, or with retrospect, you go 592 - > back and click and expand these sections where it was thinking 593 - > or creating or doing something. 594 - > Not where it's providing you the answer, but the section where it 595 - > was thinking.

596 - > It's usually in a more, uh, grayed out, smaller font. 597 - > You click on that, you can see the code. 598 - > So if the AI wrote code that actually did the process 599 - > perfectly, you do not want it recreating that code every time 600 - > you run the process because of two reasons. 601 - > One, it's gonna cost you tokens to write the code, so it's gonna 602 - > cost you more money.

603 - > Two, one in every X number of times, let's say one in twenty, 604 - > it will not write the code perfectly and you're gonna get 605 - > the wrong outcome. 606 - > So what you can do is you can take that snippet of code and 607 - > tell AI inside the instructions to use that code for that step 608 - > in the process. 609 - > This way, it just uses the code that is there. 610 - > It doesn't have to recreate it from scratch, and it will run 611 - > consistently every single time.

612 - > so now we understand what is a custom GPT and what is a 613 - > project. 614 - > Now let's talk about skills. 615 - > So what is a skill? 616 - > A skill is basically the same kind of thing, only it's 617 - > portable, and the fact that it's portable make it a lot more 618 - > reusable in several different ways.

619 - > You don't have to go to the project or the custom GPT to do 620 - > the work. 621 - > You can just tell AI to do something, and it knows which 622 - > skills it can pull. 623 - > It can mix and match skills together to do much more 624 - > sophisticated processes, and it's just much more powerful 625 - > because of that, and they are the building blocks for more 626 - > advanced agents afterwards. 627 - > So there are a lot of benefits for skills.

628 - > But a skill is basically the same thing we talked about 629 - > before. 630 - > It is a set of instructions and knowledge files just packaged in 631 - > a way that the AI knows how to pull it when it needs to pull it 632 - > So how does it know? 633 - > How does AI, whether ChatGPT or Claude or any of the other tools 634 - > that are using skills today, and today it's more or less 635 - > everything, including coding platforms and so on. 636 - > The way they know how to use skills is because every single 637 - > skill has a short one-paragraph description in the beginning 638 - > that explains what it does, and this is how it's formatted every 639 - > single time, and you don't have to know about it, and you don't 640 - > have to care about it because you don't do this.

641 - > The AI that creates the skill will create that first paragraph 642 - > for you. 643 - > But every time you start a conversation, any conversation 644 - > with any AI, it is going to upload all the descriptions, 645 - > that first paragraph of all the skill it has. 646 - > So in this case, you can see there's a skill called Proposal 647 - > Coordinator, and then there's a description, Master Orchestrator 648 - > for the Multiply Proposal Pipeline, which I showed you 649 - > earlier the screenshot for, with all the different skills in it 650 - > and all the different components and the connection to my tools 651 - > and all the other stuff that it does.

652 - > So this is the orchestrator that manages the entire process. 653 - > By the way, talking about the courses that I teach, we have 654 - > been teaching the multi-agent orchestration course for the 655 - > past few months, both to companies as private workshops 656 - > as well as open to the public, and it is extremely successful, 657 - > and people and organizations are doing absolutely magical things 658 - > immediately after the course. 659 - > So if you're interested in learning how to combine multiple 660 - > skills together to create processes like this one, but 661 - > like any other process, literally any knowledge work 662 - > that you have in your company right now can be automated with 663 - > this exact process.

664 - > Just be adapting it to the needs that you have. 665 - > So if you wanna learn that, we are currently selling the last 666 - > few seats of our August cohort. 667 - > We sold May, June, and July already. 668 - > August is the currently open one, but it only has a few 669 - > seats, if any.

670 - > So if you wanna do that before September, this would be a great 671 - > time to click on the link in the show notes and jump straight 672 - > there, or you will have to take the course in September. 673 - > If you are in a leadership position in an organization and 674 - > you wanna do this privately to your team, please reach out to 675 - > me on LinkedIn or via email. 676 - > There's a link to book time with me in the show notes, and I can 677 - > explain to you exactly what are the pros and cons and, what is 678 - > the service and so on.

679 - > But back to this. 680 - > In this particular case, this is an orchestrator skill that 681 - > manages the entire process, and if I would have opened this in 682 - > whatever tool that I'm doing, I would have gotten the entire 683 - > instructions, which are long, detailed, and complicated. 684 - > But because it knows how to read this, then if I'm going to say 685 - > something,"I need your help in writing a proposal," or if I 686 - > will say,"I need to write a quote for this and that client," 687 - > or if I will say, run their proposal pipeline, all of these 688 - > things it will understand in a regular chat anywhere in Claude.

689 - > It will know how to pull the right skills and use them, and 690 - > this skill will know how to call the other skills so I don't have 691 - > to package them all together, and so on. 692 - > So it understands from the context which skills to use 693 - > based on that initial paragraph that it reads from. 694 - > So because you can build multiple skills and they can 695 - > call one another, in addition to writing proposals, you can do 696 - > really cool things. 697 - > In this particular example, this is a finance team example.

698 - > There's a really big, large file that I showed you before. 699 - > It's again a fake file that I use for different examples, but 700 - > it's a really big, large fake file with multiple data points 701 - > and multiple rows and columns. 702 - > And if you are in a financial team, and this is a financial 703 - > report, you need at the end of each week, each month, each 704 - > quarter, whatever frequency, create multiple reports, like a 705 - > trend report, a variance report, a regional report, a labor, a 706 - > WBS breakdown report, a movers and changers report up and down, 707 - > multiple reports that you need to create.

708 - > And it takes a lot of time, and it generates more Excel files. 709 - > Well, what you can do is you can create multiple skills that will 710 - > create different outcomes whenever you want. 711 - > So in this particular case, it's something I've done for a real 712 - > client. 713 - > Instead of creating just the Excel reports that they were 714 - > creating previously, we also created a skill that creates a 715 - > document, like a detailed report with analysis and graphs and 716 - > charts and explanations like I showed you before.

717 - > We created a executive summary of that report in a PowerPoint, 718 - > and we created a dashboard that shows live data that you can 719 - > filter and change and select and go through different aspects of 720 - > the data in a interactive dashboard. 721 - > Each and every one of them was a separate skill. 722 - > They're all being fed from the same data source. 723 - > So this is the benefit of using skills, and all you have to do 724 - > is tell it what you want, and it will know how to pick the right 725 - > skills or run all of them because there's an orchestrator 726 - > that knows how to create all of them all at once.

727 - > So what are these skills' superpowers? 728 - > First of all, they're more flexible. 729 - > They can be used anytime. 730 - > You don't need to open the project or the custom GPT in 731 - > order to do them.

732 - > They can be combined with other skills to create agents or to 733 - > create multi-skill processes with one skill calling other 734 - > skills, and so on. 735 - > And you can run it in other apps, like I said in the 736 - > beginning. 737 - > You can run skills inside of Claude and skills inside of 738 - > ChatGPT, extensions inside of Excel as an example. 739 - > So if you created a skill that knows how to do a specific 740 - > financial analysis, you can now take it into ChatGPT in Excel 741 - > and do that in Excel itself without ever leaving it and 742 - > continue to work the way you're used to while enjoying those 743 - > skills.

744 - > This is something you cannot do with projects or with custom 745 - > GPTs So we talked about three out of, if you want, the five 746 - > things that you can create with AI today that can help you 747 - > automate your work. 748 - > The first one is just a regular prompt in a chat, which we all 749 - > know and like and use a lot. 750 - > Uh, to be fair, I switched completely to the agentic world. 751 - > The amount of times I use regular chat is negligible.

752 - > I'm not saying it's not helpful. 753 - > I'm just saying using a more agentic environment like Claude 754 - > Cowork or like ChatGPT Codex, or now Copilot Cowork, which is 755 - > basically a copy of the Claude Cowork just connected into the 756 - > Microsoft universe, is a lot more powerful. 757 - > You can create projects which are reusable automations. 758 - > Again, you can create custom GPTs.

759 - > I do not recommend doing this anymore. 760 - > Just create projects and that's it, because all the agentic 761 - > universe are gonna be built on top of that. 762 - > And then it can be the skills that are more flexible, can be 763 - > combined with others, and so on. 764 - > The two layers above that is creating entire workflows, so 765 - > agents that will do more sophisticated things like the 766 - > one I showed you in the beginning that connects to many 767 - > components in my tech stack, including my CRM and including 768 - > my, email platform and my marketing platform and my data 769 - > behind the scenes.

770 - > For my case, it's Google Drive, but it could be Notion or it 771 - > could be, Microsoft or any other kind of solution where you save 772 - > files. 773 - > And you can create entire applications, meaning use vibe 774 - > coding in order to connect a lot of these things together, which 775 - > in many cases are not necessary Something to think about though 776 - > before you develop any of these things is if you're running on a 777 - > personal account, then these things will most likely force 778 - > you, as you build more and more of them, to go into the higher 779 - > levels of licensing.

780 - > So you won't be able to use - I mean, you will be able to use 781 - > the$20 a month, tool, but not to do a lot of these things in 782 - > parallel. 783 - > So you will be forced to upgrade to the next level, either 50 or 784 - > 100 or$200, depending on the platform and how aggressively 785 - > you use these tools. 786 - > Uh, if you are on an enterprise account, many of these things 787 - > charge by tokens or credits, and every company does it 788 - > differently, and they do it confusing on purpose.

789 - > But in general, every time you're using agentic 790 - > capabilities, it is going to cost you and/or your company, if 791 - > it's not you paying, additional money, so you need to be aware 792 - > of that, and you need to be aware and start learning how to 793 - > reduce the amount of tokens these tools use in order to 794 - > reduce the cost that is going to be associated with using these 795 - > more advanced tools. 796 - > There are many, many different ways which we're not going to 797 - > jump into.

798 - > The easiest one to save money is to go to not the frontier model. 799 - > Models from six months ago and even a year ago are good enough 800 - > to do most of the tasks you're doing today, and so, you don't 801 - > necessarily have to use Fable 5 or GPT 5.5 or whatever the case 802 - > may be. 803 - > You can use a model from six months ago, pay significantly 804 - > less, and still get a solid output.

805 - > That's it for today. 806 - > I hope you found this helpful. 807 - > Again, I'm not sure OpenAI is going to sunset custom GPTs. 808 - > It definitely seems this way.

809 - > The rumor mill is pushing hard in that direction, and the fact 810 - > that they've hidden it very, very well, and the fact that the 811 - > API that did this through an API, the Assistants API, has 812 - > been sunset, announced to be sunset earlier this year. 813 - > Most people deserted it already, and it's gonna be completely 814 - > stopped by August of this year. 815 - > So, but there are, again, great other options in the shape of 816 - > projects and Skills and agents.

817 - > And as I mentioned, if you want to learn how to do this at Skill 818 - > and combine these things together with your tech stack to 819 - > build any kind of automation for any kind of knowledge work in 820 - > your company, don't hesitate and come join our courses or our 821 - > private workshops. 822 - > But that's it for now.