Custom GPTs Secrets: How to Get Great Results Every Time

Summary

AI Explored Podcast Summary Episode: Custom GPTs Secrets: How to Get Great Results Every Time
Release Date: April 15, 2025
Host: Michael Stelzner, Social Media Examiner—AI Marketing
Guest: Saj Deeb, Filmmaker and Founder of Howfinity

---

1. Introduction

In this episode of AI Explored, host Michael Stelzner delves into the intricacies of Custom GPTs with Saj Deeb, a seasoned filmmaker and AI enthusiast. The discussion centers around optimizing Custom GPTs to deliver consistent and high-quality results, especially for marketers, creators, and business owners.

---

2. Saj Deeb's Journey into AI

[02:12] Saj Deeb:
"I think the best place is probably eight years ago when I started making YouTube content on a regular basis..."

Saj shares his evolution from a filmmaker with a background in video production to an AI-focused content creator. Over eight years, he built Howfinity and launched the Skill Leap AI YouTube channel, which now boasts over 249,000 subscribers. His foray into AI began in late 2022 with the emergence of ChatGPT, leading him to pivot his content strategy to focus exclusively on AI tools and their applications for businesses.

---

3. The Benefits of Custom GPTs

[10:03] Saj Deeb:
"Custom GPTs are a very specialized AI tool... they take the broad ChatGPT and train it to do a very specific task way better..."

Saj emphasizes that Custom GPTs serve as specialized extensions of ChatGPT, tailored to perform specific tasks more efficiently than the general AI model. He highlights two primary benefits:

1. Specialization: Custom GPTs excel in niche areas, providing more accurate and relevant outputs for particular tasks.
2. Efficiency: They streamline workflows by eliminating the need for repeated prompting, thereby saving time and increasing productivity.

---

4. Downsides of Custom GPTs

[13:36] Michael Stelzner:
"Create a custom GPT. Got it. Okay, so what are the downsides of Custom GPTs?"

While Custom GPTs offer significant advantages, Saj points out some limitations:

• Model Restrictions: Custom GPTs are confined to the ChatGPT model (currently GPT-4.0), lacking the flexibility to switch to other models like Claude, which some users prefer for different writing styles.

• Feature Limitations: Advanced ChatGPT features such as Deep Research and Advanced Voice Mode aren’t integrated into Custom GPTs, limiting their capabilities.

• Knowledge Base Constraints: The inability to connect live sources like Google Drive restricts the dynamic updating of information within Custom GPTs.

[15:32] Saj Deeb:
"From time to time, especially if you use things to do content writing, it is worth checking the system instructions..."

---

5. Creating Effective Custom GPTs

a. System Instructions

[24:26] Saj Deeb:
"System instruction is by far the most important..."

Saj outlines a four-step approach to crafting effective system instructions for Custom GPTs:

1. Define a Specific Purpose: Clearly state what the GPT is designed to do.
   Example:
   "This GPT generates highly clickable YouTube titles for YouTube creators."
   [26:33]

2. Establish the Tone: Specify the desired language style.
   Example:
   "Direct, compelling, and slightly clickbaity, but not misleading."
   [26:33]

3. Set Rules: Outline guidelines to ensure consistency and relevance.
   Example:
   "Keep titles under 60 characters, use power words, balance human appeal and SEO."
   [26:33]

4. Provide Examples: Include sample outputs to guide the GPT.
   Example:
   "10 AI hacks that will save you hours a week."
   [26:33]

Notable Quote:
[26:33] Saj Deeb:
"That's why it took like 30 minutes to really fine-tune these because you want to build them, give it instruction, see the output and then go back..."

b. Knowledge Bases

[29:30] Saj Deeb:
"Most of mine don't have a knowledge base because the system instruction has enough space..."

Knowledge bases supplement system instructions by providing additional context and information. Key points include:

• Document Management: Limit to 5-10 documents, each up to 3 pages or 1.5 million words, to prevent overwhelming the GPT.

• Content Quality: Use plain formatting (e.g., Microsoft Word documents) to ensure compatibility and effectiveness.

• Use Cases: Ideal for tasks requiring specific knowledge, such as creating Facebook ad copies or detailed marketing strategies.

[31:37] Michael Stelzner:
"Notebook LM is a Google tool that is free..."

Saj recommends tools like Notebook LM for managing extensive knowledge bases, especially when dealing with large amounts of dynamic data.

---

6. Practical Applications: Examples of Saj’s Custom GPTs

Saj illustrates the power of Custom GPTs through various use cases:

• YouTube Titles and Descriptions:
  Generates SEO-friendly and engaging titles and descriptions by processing video transcripts.
  [22:34]

• LinkedIn Posts:
  Transforms YouTube transcripts and search-based content into LinkedIn posts, streamlining cross-platform content repurposing.
  [22:34]

• Facebook Ad Copy (Marketing Navigator):
  Creates benefit-driven ad copy tailored for Facebook, utilizing a structured knowledge base containing target audience insights, unique selling propositions, and high-performing ad examples.
  [22:34]

Notable Quote:
[23:59] Saj Deeb:
"It was like, okay, give me another one, give me another one. And then I kind of just look at it and then copy and paste to Facebook from there directly. So it's a huge time saver."

---

7. Enhancing System Instructions

[41:17] Saj Deeb:
"I like that."

To refine system instructions, Saj suggests leveraging ChatGPT or Claude to analyze and improve existing prompts. Steps include:

1. Copy Existing Instructions: Extract the current system prompt from the Custom GPT.
2. Seek Improvement: Use ChatGPT or Claude to refine and enhance the instructions.
3. Implement Changes: Paste the improved instructions back into the Custom GPT.

This iterative process ensures that Custom GPTs remain effective and aligned with user intentions.

---

8. Team Plans and Sharing Custom GPTs

[44:09] Michael Stelzner:
"Let’s talk about teams and the benefit of Custom GPTs."

Saj discusses the advantages of utilizing Teams Plans for sharing Custom GPTs within organizations:

• Centralized Access: Custom GPTs saved in the workspace are accessible to all team members, fostering consistency.

• Controlled Sharing: Unlike public links, workspace-shared GPTs maintain privacy and security, safeguarding sensitive information.

• Ease of Use: Team members can readily adopt Custom GPTs without individual setup, enhancing collaborative efficiency.

[46:00] Michael Stelzner:
"Talk about the @ mention."

Utilizing the @ mention feature allows users to quickly access and integrate multiple Custom GPTs within a single chat, streamlining workflows and reducing the need to switch between different GPTs.

---

9. Conclusion

The episode concludes with Saj offering avenues for listeners to connect and engage with his work through LinkedIn and his website, SkillLeapAI.com. He emphasizes the importance of specialized Custom GPTs in optimizing business processes and enhancing productivity for marketers and entrepreneurs.

[48:26] Saj Deeb:
"SkillleapAI.com is the website where we have the library of courses..."

---

Key Takeaways

• Specialization over Generalization: Tailoring GPTs for specific tasks yields better performance and efficiency.

• Structured Setup: Clear system instructions and well-organized knowledge bases are crucial for effective Custom GPTs.

• Continuous Improvement: Regularly refining system prompts and leveraging advanced tools ensures Custom GPTs remain relevant and powerful.

• Collaborative Advantages: Utilizing Teams Plans facilitates seamless sharing and consistent usage of Custom GPTs within organizations.

Notable Quote:
[26:33] Saj Deeb:
"These are the most important things. I don’t want you to mess. That's when I notice people get frustrated..."

---

For more detailed show notes and resources discussed in this episode, visit Social Media Examiner Podcast.

Transcript

[Michael Stelzner]

Hey, before we start today's show, if you want to accelerate your AI learning, I have a solution for you. Become a member of our AI Business Society. You'll join me as we go deep with live AI training each and every month. Imagine crafting more persuasive content, creating stunning images and automating those time consuming tasks. It's all possible when you join the AI Business Society. Go to socialmediaexaminer.com AI and join today. Welcome to the AI Explored podcast, helping you put AI to work. And now, here's your host, Michael Stelzner. Hello, hello, hello. Thank you so much for joining me for the AI Explored podcast brought to you by Social Media Examiner. I'm your host, Michael Stelzner, and this is the podcast for marketers, creators and business owners who want to know how to put AI to work. Today I'm going to be joined by Saj Deeb and we're going to explore custom GPT secrets. If you have not created a custom GPT or you've created custom GPTs that are just kind of not working for you, today is the episode for you. Are you new to this show? If so, follow us on whatever podcast app you're listening to. We've got some amazing content coming your way. Now let's transition over to this week's interview with Saj Hadib, helping you simplify your AI journey. Here is this week's expert guide. Today, I'm very excited to be joined by Saj Hadeeb. If you don't know who Saj is, he's a filmmaker and founder of howfinity, a company that helps creators and businesses learn the the craft of video creation. His AI focused YouTube channel is at Skill Leap AI and he has more than 249,000 subscribers. His membership is Skill Leap AI and it's designed to help entrepreneurs put AI to work. Saj, welcome back to the show. How you doing today?

[Saj Deeb]

Hey, great, Mike, thanks for having me.

[Michael Stelzner]

Well, I said welcome back, but actually this is your first time on this show, so welcome back to one of my shows. So today Saj and I are going to Explore creating customers GPTs that will deliver quality results. Saj, there might be plenty of people in this audience that are not familiar with you, so I would love you to share a little bit of your journey, how you got into AI. Start wherever you want to start.

[Saj Deeb]

Sure. I think the best place is probably eight years ago when I started making YouTube content on a regular basis and I was in the video production world. I've had a video Production company for a long time. That's my background. And I was making a lot of how to videos for corporations, a lot of training videos. And I was like, well, maybe I should do that for myself. I saw a lot of people doing that on YouTube. Well, not a lot at the time, but at the time, to me, it seemed like a lot. Now everybody seems to have a YouTube channel. But. So eight years ago, I started posting how to videos on that channel called Howfinity, which recently got to a million subscribers. Took me eight long years, I think like 1500 how to videos to get there. But that channel, I was very consistent. It took me like a year to get to a thousand subscribers, by the way. So I don't know how I kept going. I was going on film productions, I had my production company, everything else was going on. But I still managed to upload pretty much every single day that first year, and I'm glad it did. And around 2020, with COVID when Zoom was going viral, I posted a video about Zoom, how to use Zoom. So all the stuff I was covering in the how to space was around the latest tech, whatever the latest tech was. And then that Zoom video got something like 2 million views in a. In a week.

[Michael Stelzner]

Wow.

[Saj Deeb]

In the how to space, you never go viral. So I was, like, very confused what was going on, and I kind of doubled down. And like, within that time, I got to 100,000 subscribers on that channel. I was like, okay, this is good. Because I was, you know, like everyone else. I was not working in my regular job. Everything production world was basically turned off for like a whole year. So we never went on a single shoot during that time. So it kind of worked out that the YouTube stuff took off. And then how I got into AI was in 2022, at the end of 2022, when ChatGPT came out, just because I was covering the latest tech for business, typically for business use case. But I was kind of all over the place. I started using ChatGPT myself just to test it out. I made a tutorial about it. I was like, wow. At first I got a little depressed because it just did such a better job at writing a script than I was doing at the time. And I was like, wait a minute, is this going to take over my YouTube channel? But then I was like, okay, I'll just cover it. I'll make tutorials. And then a few of the tutorials on the Howfinity channel around ChatGPT. And I think I was covering Dall E and Mid Journey at the Time that those were pretty big too. And then the funny thing, I had just gone on one of your podcasts and you had talked to me about being more laser focused because with howfinity, I was covering so many different things and I was like, we had brainstormed about what should I be focused on. I was focusing on more on content creation. And then with ChatGPT was like, okay, this is new. We just had that conversation. I was like, this is something I could go into because the disruption was so clear. The first time I used ChatGPT, it was so obvious that literally everything in my life, in every line of work I was in, was going to change. I mean, it's become obviously way more obvious now, but even the first week I used it. So I created that second channel, which is the Skill Leap AI channel now that is all focused on AI, just so I have a focus niche. And I was kind of uploading to both at the same time AI content to both. And then Skill Leap started getting much better traction because when you have a focus on YouTube, it's a lot easier to grow than general how to tech content. So I still, to this day, I still publish to both, but I really focused more on the Skill Leap channel, focusing on AI, more use cases, very practical tutorials, always cover the latest AI tools. And then from there I build Skill Leap. Skillipai.com is the community that I built that was like AI courses. And I made a really affordable was like 25amonth. You get access to all the AI courses. And then through that I was able to kind of build a separate business. We launched that back in June 2023, and that's pretty much what I'm spending most of my time on now.

[Michael Stelzner]

Tell us a little bit about the YouTube channel, the AI channel, like what kind of content are you covering on there?

[Saj Deeb]

So for a while I was really focused on just ChatGPT content during 2023. But I started doing some stuff around Mid Journey and I was like, well, that's kind of a little confusing because Mid Journey is more for creators and I'm in the creator background, like, you know, from filmmaking background. My knowledge is in that world. But I'm also a business owner and I think I'm a business owner more than I'm a creative person. So I started pivoting towards just helping marketers and business owners. What I usually like to do, even though I. I do some of the creative work myself too, but I usually let the team, I think they do a better job. So I pivoted to teaching marketers and entrepreneurs how to use AI more than anything else. And I pivoted away from just ChatGPT and anytime I came across an AI tool, I would just set aside a few hours a week and I would test out every single AI tool that came out. Sometimes it would take a very long time to test out. Sometimes it would just take like 10 minutes. I'm like, okay, I'm not going to cover this. So I try to be the first YouTube channel when a new AI tool or a new AI model came out and I thought it was practical to make the first video that was like a practical walkthrough, not a news related video. I saw some people were doing news, I was like, no, I'm not even gonna actually, if a new model comes out and I don't have access, I'm not gonna make a video. I'm gonna wait till I have access and then I'm gonna make a video. Like top 10 practical ways to use chat GPT4, oh mini for marketing or whatever the next thing was. And, and then I think through that people find it really practical. Like this week my latest video is on this tool called Manus AI.

[Michael Stelzner]

I was going to ask about that because I have access to that and I don't even understand how to use it.

[Saj Deeb]

Yeah, it was, you know, it was going viral so I was like, I should make a video, I should make a video. And I said, well, I don't like making videos when I can't show you how to use it. And some of these demos nowadays, they look so fancy, but in reality they're not. So I was like, I'm going to wait. I finally got access, then I made the video and then I showed, I think I showed seven different use cases in there. They have limited credit so I was only able to do like four, four of them myself. Just today I gave it a text prompt. I said, make me a website where people could test their knowledge against AI. And it was like, that's it, it's like a one sentence prompt. It literally in an hour created the entire website for me. Front end, back end, all the copy and it was actually functioning. You could go there, you could pick an AI model and you could have an IQ test against it and it gives you the result at the end. It's like, I was so confused. Literally one text prompt. He generated all of that. The entire. It created something like 15, 20 different documents, including text, documentation, code, all the different IQ quiz. It's not connected to some of the APIs to make it a real application. But it's unbelievable.

[Michael Stelzner]

It's crazy. Well, it's absolutely crazy. And I was super excited when I discovered that you actually had an AI channel, because I had no idea, because I know you from the video world, right? So I'm really excited that you're here today. And folks, Saj is really good at what he does, so definitely follow him on YouTube. It is @ Skill Leapai. No, Saj, I feel bad for you because, like, there's something new almost every day, and I would imagine you can't possibly keep up with all of it. And we're here today specifically to talk about custom GPTs, which is part of ChatGPT. So let's talk a little bit about why should marketers and entrepreneurs use custom GPTs? Let's just kind of begin with kind of the benefits or upsides of utilizing custom GPTs.

[Saj Deeb]

Okay, so custom GPTs came out back in 2023. So it seems weird since I'm always on the latest, greatest AI tools to talk about it, but when we were talking about doing this, I was like, well, it's the thing I use every single day. I was thinking about from all the AI tools, probably 20 different AI tools I use on a regular basis, what's the one I use all the every day? And I haven't stopped using it since the day it came out. I was like, I don't think any workday I've not used one of my custom GPTs. And that's why I thought even though there's not been a huge leap in custom GPTs, they don't update it all that much is still super practical. And the reason, there's really two reasons why I think it's still very applicable to teach custom GPTs, and why people should, especially marketing, content creation and really any type of small business use case one is basically, custom GPTs are a very specialized AI tool, right? So you take the broad chat GPT and you train it to do a very specific task way better than the general version of ChatGPT or any other AI tool could do. And then all the other AI tools out there outside of the chat models, they're also designed to do one thing really, really well. Or with custom GPTs, you could train bunch of them. So what I've done with my company is pretty much every single thing that I do throughout the week. I thought about it and I thought about, is there a version of that where a custom GPT could help me do it faster or do it better or just optimize the way I'm doing it. And then I would go through that list of things that I was doing every single week, and then I would build one and then I would test it out. And you just tweak it one time, like, maybe It'll take like 30 minutes to really dial it in and then you have it for a long time. When I was prepping for this podcast, I was like going through back through some of my older ones. I was like, I actually haven't updated this custom GPT for like a year, but I still use it every single day the same way. It's worked perfectly fine for the last year. And the second reason is it's really designed to speed things up, right? So if you open up ChatGPT every single time, what people do, and I see this all the time when I work with people, they go and they try to find, oh, what was the perfect prompt formula again? How should, how should I start? Or, oh, this is not giving me the result. Oh, there was this prompt directory I actually remember that I downloaded. There was some PDF, let me go and copy and paste the prompt. With custom GPTs, you don't actually prompt it once you set it up. And I'll talk about some of the way I set it up once you set it up. I literally just copy and paste what I needed to do from like, let's say I'm transcribing or turning a YouTube transcription into a LinkedIn post, right? I just literally copy and paste the transcription, paste, press Enter, and it gives me the output. I don't have to prompt it, I don't have to describe the tone, no context. So it's a really fast thing that speeds up my day to day. So I was like, okay. Entrepreneurs really mainly care about buying time back and being more productive. And that's why I think custom GPD is the perfect fit for that. It's just the fastest way to get things done from all the AI tools that I use.

[Michael Stelzner]

Still today we have covered custom GPTs on this show multiple times. It's been a while. But really what we're going to learn today from SAJ is ways for us to really dial it in. Now, my understanding, and you can correct me if I'm wrong, because I tend to use cloud projects more than I use custom GPTs, even though I do have one custom GPT, you have to have a paid account with ChatGPT. Is that correct? Create a custom GPT.

[Saj Deeb]

True, yes. You could share it with people that don't have it, but to create One you have to pay for that.

[Michael Stelzner]

Got it. Okay, so what are the downsides of custom GPTs? Because I know that there are a whole bunch of advanced features that ChatGPT has that aren't necessarily built into custom GPTs. Right?

[Saj Deeb]

Yeah. And one of the ones is the one you brought up, the fact that you can't pick a different AI model. So if you like Claude more for its writing style. Right. That's not something that you could get inside a custom GPT. That's a ChatGPT product and it's only going to use a ChatGPT model to do it. So I know a lot of people because CLAUDE projects came out after custom GPTs, and I noticed a lot of people switched over because at the time they liked the style of the writing. But I feel like ChatGPT, the underlying model, has gotten so much better at writing before, you had to really get creative with your prompt to get it to not write. Like kind of a spammy salesperson. A lot of times the default writing style was really strange, but nowadays I don't have a hard time either getting it to do it, or by default, it does a pretty good job. That's why I stuck with it over cloth projects. And to confuse everything even more, ChatGPT also released something called ChatGPT projects, which is a ripoff of Claude projects, which was like a copy of ChatGPT GPTs. It got really confusing, but. So that's one of the main downsides is that you can't pick a model, and not only can you not obviously pick a different model from Claude, you can't pick a different GPT model either. So the reasoning models, I really started using those more and more. Like the O1 model, the O3 mini model, the ones that, you know, kind of think through the steps before they give you a response. Well, those are not something that ChatGPT could use inside of custom GPT, so I really hope they roll that out.

[Michael Stelzner]

Just to clarify real quick, my understanding is the custom GPT defaults to whatever the latest default project like, as of this recording, is chatgpt. Is it 4.5 or is it 4.0? I can't remember what the newest one is.

[Saj Deeb]

This 1 defaults to 4.0. Still 4.5 is out, but it's not.

[Michael Stelzner]

Inside of the custom G. Yeah, so that's important, right? I'm assuming once 4.5 becomes the standard, then all of a sudden all the projects are going to work with 4.5, but I also think it could potentially break some of the older projects, Right. I mean, I've heard stuff like this. So this is important for you to cross check these things, right?

[Saj Deeb]

That's totally true. That's the one thing I did notice when all our original custom GPTs were built on GPT4 and then GPT 4.0, which has pretty much been the main model for like a year that most people use, that one did break some of ours, not a lot, it wasn't like a huge leap. But I think four or five is also not going to break many things because it's just better at writing. So it's not like a whole different thing that's going to break things. But from time to time, especially if you use things to do content writing, it is worth checking the system instructions, which I'll talk about here. It is something to pay attention to. Yeah.

[Michael Stelzner]

And then what about Deep Research and Advanced Voice mode and some of these other features that are part of ChatGPT? They are not included with custom GPTs. Web research is though, right?

[Saj Deeb]

Web search is. So they have basically four functions that every time you build a GPT you could pick and choose from. Search is one of them. And I have a good example of one where I use search all the time and I have that on as a function. The other one you could use Dall E. I don't really use the Dall E custom GPTs, but you could turn that on to generate images and help you with thumbnails and things like that. They have an analysis tool that you could turn on, which I also don't use because I think Claude does a much better job with data analysis. So I don't really have any custom GPTs these days that do analysis. And they also have this canvas for writing, which I'm honestly not a fan on. I don't even like it inside of ChatGPT.

[Michael Stelzner]

Yeah, they call it Artifacts inside of Claude, but yeah, I know, exactly.

[Saj Deeb]

Yeah, I like them in Claude a lot more because Claude could give you visual presentations of like apps and dashboards and things like that. So that one I also don't turn on. So Deep Research is not there. Which what you could do, what I've been doing is using Deep Research and then taking the findings of Deep Research and adding it inside of projects or inside of custom GPT as a knowledge base, like as part of my knowledge base, which I'll talk about. So that's one way you could do it, but it's not part of it. But one big thing is that I don't know why they don't have. Once I talk about the knowledge base when you could give it upload like, you know, a bunch of documents as your source, as your knowledge source. It's kind of static. Like it's whatever documents you upload, you can't connect it to your Google Drive, for example, and then have those updates happen in Google Docs and then come into your knowledge base. So it's not useful for that. So we use notion for things like that. Or we use inside of ChatGPT. You could connect it to Google Drive. It's missing that function inside of custom GPTs right now.

[Michael Stelzner]

Okay, so we might have sent a lot of people thinking that custom GPTs aren't valuable because of all these limitations, but really what we're saying is that ChatGPT has a million different features and functions. They have not yet integrated a lot of these into custom GPTs. So why don't we talk about like what kind of custom GPTs we should make given these limitations. Because despite these limitations, there's still a lot of value. Obviously that can happen with custom GPTs. So why don't we explore some of the examples of some of the things that you've done with custom GPTs just so people can understand how powerful they can be.

[Saj Deeb]

Yeah, and I think that's a great point. It's a lot of times I notice people want one AI tool to do everything. They're like, well, what's the one you use all the time? I'm like, there is really not one. I mean, custom GPTs is one thing I use all the time, but the reason why I like them is for very specialized tasks, I don't get them to do like everything that I do. I use ton of different AI tools for different things. So the two ways basically you could train a custom GPT. And these are some. I'll give you a bunch of use cases that I have for mine. They have this GPT builder, right, where you could just have a conversation with the GPT builder and it'll help you kind of create it. And the other way is with system prompts that I'll talk about in a bit. But the one thing that I do with them is, as I mentioned, is like coming up with my day schedule. I'm figuring out, hey, let's go with YouTube. I have to come up with YouTube titles. Well, I'm going to create a GPT for that. To help me optimize my YouTube titles, I have to somehow come up with a YouTube description. Incredibly time consuming, especially since I've made like 1500 videos. You want it to be SEO based, but you want it to also be organized in a nice way. So a lot of people look at the description on your videos. Well, I have a GPT for that. It literally takes a transcript of a YouTube video, which YouTube automatically gives me. I copy and paste, then I get a transcription for that. I literally have one called script refresher where I upload an old transcript and it gives me a new transcript because a lot of times I'll make a new version for a video I've already made. Right. How to use ChatGPT. I've made that in 2023, in 2025. It can't be the same exact transcript. I start with that script refresher, it gives me a new one, I edit it, update it and I'm ready to go. That's just like on the YouTube side, just three. I use almost all the time. The title one I obviously use pretty much every single time. And then I also use ones for generating content for other platforms. The one thing custom GPTs are really good at is getting them to give you content when it's repurposing other content or if it's trying to find content from search and repurposing it. So I have one for LinkedIn posts and LinkedIn. It's really hard to have time for like a hundred. I saw you posted on LinkedIn that how do you find the time to post on, you know, all these different product pages and company pages and personal pages and then all the social media platforms? Right. So having these AI tools, especially these little GPTs, I design one for a very specific platform. So one is just for LinkedIn posts and then I also have one that turns a YouTube transcript into a LinkedIn post and another one that I have where it turns search based content into a LinkedIn post. So it goes and finds the latest AI news, turns it to the LinkedIn post. So you could see how little difference in those GPTs. I don't try to make the same one do two things. That's what most people do. They're like, I'm going to build this custom GPT that is my social media marketer. Well, that's like a hundred things. It's just not going to do a good job. It's only going to Write well for LinkedIn. But that same one can't also make you a good YouTube description. You have to make them a little bit separate. And I've dialed it down. At first I was trying to build too many and I was like this is not a practical thing thing. So I was like, what are the real serious things that I do very frequently that I need this for? Everything else I still use chat, GPT or Claude with a prompt, you know, but the ones that are things that I do almost every day, like YouTube titles, descriptions, posting on social media. If I'm coming up with new ad copy, I have one for my Facebook ad copy that I've trained with a whole set of knowledge base and all kinds of different things. That one I use all the time.

[Michael Stelzner]

What about you? You mentioned a marketing navigator. Remember we were talking about this marketing. Tell us a little bit about that.

[Saj Deeb]

It's funny, I called it marketing navigator, but I almost exclusively use it for Facebook ads.

[Michael Stelzner]

Oh, it's the same one.

[Saj Deeb]

Yeah. So that one I've trained over time with my Facebook ads because that's where I spend a lot of time doing my Facebook ads. And then what I've done is I gave it a really detailed set of instructions that I'll kind of talk to you about, like how I do these instructions. And then I give it five different documents as the knowledge base. For example, one of them is just my previous ad copy and what performed well. And I've obviously analyzed that over time. And then I gave it that and some other examples of ads that I liked to give it as an example. So that one is one of my favorite because it literally has bunch of research that if I was, for example, let's say I have to come up with a new ad copy for some platform for Facebook, I would go to ChatGPT. But if I start from a blank ChatGPT chat, I would have to come up with a good prompt, I would have to give it examples, I would have to upload previous things. I would have to go do a bunch of back and forth. This one literally is like one click, it gives me a copy and then I go from there. It was like, okay, give me another one, give me another one. And then I kind of just look at it and then copy and paste to Facebook from there directly. So it's a huge time saver. Like, I assume most people that are, you know, doing marketing, any type of business owner, it's all about saving time. This will save you a ton of time once you build a few of these out.

[Michael Stelzner]

Perfect. Okay, so what I heard you say is I've got a lot of custom GPTs for very specific tasks like titles, descriptions, refreshing scripts for making new versions of videos, LinkedIn posts, search news into a LinkedIn post, Facebook ad copy. With this marketing navigator, you mentioned earlier, the system instructions. So talk to us a little bit about the importance of setting up the instructions for the chat custom GPT.

[Saj Deeb]

So, yeah, when you train these custom GPTs, as I mentioned, the normal way to do it that I don't use anymore is you just literally have a chat with the custom GPT creator and it creates that system instruction for you by itself. Right? You could go look at it when you edit that GPT and it has that in the background by itself. These days, that's not what I do. When I go to create a GPT from scratch, I just go into the edit page of it where you could give it like a name and you could give it a description and you could give it a system instruction. And I manually type in the system instruction there on my own. I don't let it, like, pick a system instruction for me. So the way I come up with system instructions, I kind of like thought about it in, because since I make a lot of videos, I think of things in like, formulas, like, what's the best way you could give it a system instruction? And for people that don't know system instructions, basically AI chatbots, you have your prompt that's talking to it as a user. But then the system instructions could train the chatbot to talk to you in a very specific way or have access to a very specific knowledge base or a tool like search. So those are the things when you set it up, that's in the behind the scenes of it. And system instruction is by far the most important. So the few different things I do when it comes to like, really creating a good system prompt is the very first thing is like, you have to give it a very specific purpose, since again, these are little mini GPTs with a specific purpose. So for the YouTube one, mine starts with this GPT generates highly clickable YouTube titles for YouTube creators. That's literally like the first sentence of the custom instruction. And then the second thing is I give it what tone do I want? And the tone for this one I have direct, compelling and slightly clickbaity, but not misleading. And I found, like, this takes a little bit trial and error, right? So that's what I had. I think the first time I tried to do it, I said make it clickbaity, but then he would just say nonsensical titles that you shouldn't have on YouTube but you do want people to click. And I found the best way to kind of describe that was slightly clickbaity but not misleading. It took a few tries to get that to work. Right.

[Michael Stelzner]

I like that.

[Saj Deeb]

Yeah. That's why it took like 30 minutes to really fine tune these because you want to build them, give it instruction, see the output and then go back, be like, no, that wasn't quite right. Or you'll see it uses too many emojis. Okay, go fix the system instructions. The third one is I give it a list of rules. So the one for YouTube titles, I said keep the titles always under 60 characters long, use power words, use a balance between human appeal and SEO. And that one, you know, I do a lot of SEO type of videos too. So that kind of helps me with that. Avoid misleading clickbait. I added that again because even with that other one, it wasn't quite getting. It was still a little clickbait. So I added this to the third part of it and I said, give me multiple variations. Sometimes I'd be like, every time, give me three titles so I could like easily choose or give me five titles. And then the fourth part is I give a very specific example on things that I've done before that worked well or what I think will work well even if I hadn't done it before. So the one video that I had go pretty viral. I think it got like a million views in ChatGPT related topic. And it was. I created the perfect Chat GPT prompt formula, right? I didn't really create it, but it was a video about that prompt formula and that title worked well. So I gave that as an example, 10 AI hacks that will save you hours a week, things like that. So I gave it multiple different. I think this one only has like five different examples of what it should be, what an optimal title should be. And then I pretty much, that's the system instruction that I had. And I followed those four steps every single time I create a system instruction. What I did, I actually built a GPT is publicly available too, that builds custom GPT system instructions. It's like, I think like it has like 35,000 messages that people have and it's pretty highly rated. You could rate public GPTs. It has like a 4.5 star. And with that one, you just give it a crappy system instruction and then it like turns it into. Follows this formula and turns it into a really solid system instruction. And you could always ask ChatGPT to help you come up with a system instruction. Give it all the context on what you need it for and it does a good job and you copy and paste from there.

[Michael Stelzner]

Well, I'd love you if you're able to look up the custom GPT and how we could find it. And I'm going to iterate what I heard you say there are four different steps that we need to take when we are creating the system instruction or system prompt. Number one is give it a very specific purpose, you are blank. Number two, discuss the tone of the language that you want the output to be. Number three, set up a list of rules, for example, I want to see multiple variations. And number four, give examples of the ideal kind of output that you want from it. Is that correct?

[Saj Deeb]

That's exactly right. Yep.

[Michael Stelzner]

Did you find your custom GPT and how do we find it?

[Saj Deeb]

I could share the link with you. Actually it doesn't have a good name.

[Michael Stelzner]

Okay, if you want to give me the link, I'll make sure we add the link to the show notes. Okay.

[Saj Deeb]

Okay. Yeah, I'll share that with you.

[Michael Stelzner]

Okay, cool. So what do we need to know about the knowledge base? Let's talk about that.

[Saj Deeb]

Okay, so this is, I think a lot of people get confused with the knowledge base because for some reason you could give it an incredible amount of information. I just actually looked it up just to see where their info because they changed how much context you could give it. And as right now you could give it 20 different documents as your knowledge base. Each one could be 500 megabytes and each one could be 2 million tokens, which is like a million and a half words, like 1.5 million words per document. So just because it allows you to do that, I noticed people are like okay, let me give it millions of words as a knowledge base. And it almost always breaks the GPT. You just can't figure out anything that you're doing. So what I found and usually how I tell people when you create it based on a knowledge base and that marketing navigator I have, I'll share the five different documents I have for that one. So I always start with five and never more than 10 documents. And I always make sure the documents are kind of stripped down if I had a much bigger document to start with. So let's say I had a Google Doc that was like 30 pages. I would actually run that through the regular chat GPT, give it three page version of that. I'll just ask ChatGPT, hey, summarize this to just a three page document or like a thousand word document, whatever. And then I use that, download it and I give that as a system instruct or as a knowledge base. Just because I found it does not do a good job if you do more than that. Now if you do need to have more than that, I still will try to figure out how to do it with a small amount, five to 10 documents, each one up to three pages and then go up from there after you do some testing because you'll notice at some point it breaks for massive knowledge base When I need a massive knowledge base, I use NotebookLM for example. It could have 300 different sources and it does a much better job than like a custom GPT would do. So if I'm trying to analyze any information, I just picked the right tool for the job.

[Michael Stelzner]

Let's just note the Notebook LM is a Google tool that is free and I would imagine it could take a big doc and make a littler doc for you. Is that kind of what I'm hearing you say?

[Saj Deeb]

Yeah, you could start with any source. You could have YouTube video as a source, PDFs as a source, any Google Docs as a source, and the free version of it lets you have 50 different sources. The paid version, they have a paid version that's inside of Google Workspace lets you have 300. And you could mix and match all these sources as your knowledge base. So that's why it's one of my favorite tools in the world of AI in general. But when I have to talk to my own knowledge base, I use notebook element. I just don't use custom GPT. That's why when I teach people custom GPTs, I want them to think about these very specific use cases that are repeatable. Not can I make it work for this thing? No, there is a better AI tool for that purpose.

[Michael Stelzner]

Okay, let's just back up a little bit and talk about what is the role of the knowledge base. Just so everybody understands why it's so important. Because we've already talked about the system instruction has examples inside of it, right?

[Saj Deeb]

Yep.

[Michael Stelzner]

So why should we even attach extra documents?

[Saj Deeb]

I usually think of it in like three steps every time I create a custom GPT. Most of mine don't have a knowledge base because the system instruction has enough space there where you can give it some examples. Like the YouTube title I mentioned had some examples. But when it comes to ad copy, it will take up way too much of the system instructions for me to give it all the other things we talked about, plus bunch of examples and plus other things I wanted to have specific to my company information. That's where my knowledge base comes in to kick in more information as it needs it. So I notice every time I use a custom GPT, it almost always Looks at my custom instruction first it tries to give me a response based on that. Then it goes to your knowledge base and it pulls from that. And then if I want search to be the third thing, I turn on search. So go through my stuff first. Then if you need more information, go through search. So it's like three different things where I could pull in knowledge and then in your system instructions you could actually give it that order. First, go through the system instructions. Second, go through the knowledge base Third, do an Internet search and to see if you need to find more information.

[Michael Stelzner]

Oh, I like that.

[Saj Deeb]

Yeah, it depends on what you're building this function for, but that's where the knowledge base comes in. I think maybe from the 20, 30 different GPTs I've built and I probably use like 10 of them pretty regularly. I think only one or two of them has a knowledge base, including the one I call Marketing Navigator. This is my knowledge base for that. It has five documents. The biggest one is three pages, but most of them are like one and a half pages. The first one is. And I could give you this, basically this. The system instruction starts like this. You are an AI trained marketing assistant specialized in highly converting Facebook ad copy For Skill Leap AI. Your goal is to create engaging, benefit driven ad copy that appeals to entrepreneurs and marketers. And then I give it its key responsibilities. It's briefly about the target audience and some of the selling points, but it's very brief. So what I did with the system or with the knowledge base I give it. The five documents are the target audience breakdown. It's a full page all about the target audience and what they are. The second one is just a unique way of positioning this like a unique selling proposition for that. I have a whole page on that and I just work with ChatGPT to kind of create these documents in a conversational way. So I don't like type these out even myself, but I do make sure they're exactly what I want. After a little question answer from ChatGPT, the third one, the future and benefits of the product. But our unique selling power position and the target audience has the pain points and all the different things that I like to put first in an ad and then lastly I would put the future and benefit part of it. Any best practices that I have for YouTube ad copy, that one is like a document I created with the help of ChatGPT, just general Facebook knowledge. It has that in his training data by default. But I like to emphasize some other key points. And the fifth part is the most important. Part which is bunch of the ads that perform well. And I don't like to put that in the system instruction because they're just too long. Especially when like the long form ads. I would have to give it three examples. The system instructions will get diluted. I like to keep the system instructions pretty tight. You could have like 8,000 characters in that box. But it has the same issue as the knowledge base. It's going to go all over the place. These are the most important things. I don't want you to mess. That's when I notice people get frustrated when it's like no, it's not following my system instruction. And I look at their system instruction, it's like 7,000 words or you know, as long basically they maxed out the box. I'm like, yeah, it's getting confused here. You got to narrow that down and put the extra stuff in the knowledge base as best as you can.

[Michael Stelzner]

Okay, so a bunch of questions here. First of all, the typical documents that you would include, I heard you say, are the target audience your unique value proposition for your product or your unique selling proposition. As you said it features and benefits. That's the third thing. Best practices. And then of course a bunch of high performing ads. Now I would imagine if you're savvy, which I know you are, you could use Chat GPT to help you develop these documents. Right. For example, you could use deep research to help you come up with the best practices. Have you done stuff like that?

[Saj Deeb]

That's exactly what I did for that one. So what I do is I give ChatGPT a lot of context about my business. Nowadays you could just turn on search and have it go crawl your whole website. So I'm like go to this website. I'm trying to make a document for my target audience to train a custom GPT to do things a certain way. And it will crawl my website. It will go through the pages that I put in the URL, I'll put in the URL, I turn on search and then it goes and pulls in all that information. And then I say ask me as many questions as you need to to help me write a document that I could use for this use case. And it'll just. A lot of times it'll ask you 10 questions at a time. So I say ask me one question at a time, wait till I reply, then ask me the next question. Because that way I found it. For some reason, every time you say ask me questions it will just ask you 15 questions in a row all at once. And That I find it. I think you're a fan of the voice mode too. So a lot of times I do this part on the go in the advanced voice mode of ChatGPT and I just have a conversation. It's more dialogue, one on one kind of a thing where it's asking me one question at a time and then I refine that document. And the one other thing about Knowledge base is you can have multiple different document types. But again, what I found works the best is the plainer the document styling, the better. So if you could just do all Microsoft Word documents then like PDFs that have different formatting and tables and things like that, it just does a better job. That's inside of the ChatGPT OpenAI website. They say the same thing. The more formatting you have, the more you could cause the knowledge base to not work properly. So I almost always download these as a simple word doc.

[Michael Stelzner]

Okay. I got a bunch of questions. Got to remember them all. So first question is you mentioned links. So if you turn on the search in a custom GPT and let's say you have a sales page that's constantly being updated or iterated, can you in the system level instruction just say learn about the product by visiting these links and will it do that each time it runs it?

[Saj Deeb]

That's actually a really good idea. You know, I didn't. I'd not build one like that. So I always thought about given at the Knowledge base and the system instruction as like the static data. But I feel like with search you could bypass that. I have to try that. That's a really great idea.

[Michael Stelzner]

Yeah, because you kind of triggered that a little bit when we were talking on that little last discussion about how you're using ChatGPT in general to help you prepare these documents. Okay. Yeah, that was the first thing. It's worth an experiment is what I'm hearing you say everybody with the system level instructions giving it some links and then giving it specific instructions to go out and read the content on those links to kind of fill its brain up and see if that works. That's an interesting experiment giving it the examples of the high performing ads. Now I know that Facebook ads typically have video, images, all that kind of stuff. Are you dealing with screenshots? What are you dealing with in order to give it the copy, if you will, for the high performing ads?

[Saj Deeb]

Yeah. So this one that I built was specific to making the copy just a copy portion of the Facebook ads. I had built out this other. I kind of put together this PDF document for making Facebook ads. And that one is when I started using that technique, when the vision capability of ChatGPT came out, where I would literally take screenshots of my. I would go to the Facebook ad library, I would look in under, you know, a lot of people were making AI kind of course content at the time. I would find my favorite ones, I would screenshot it and then I would just use the regular ChatGPT to extract all the information and to help me create some of those knowledge based documents that I mentioned and some of those examples. So I did not build those inside of custom GPT. It's just, it would be too limited for it to be able to do that. So I kind of use that as two steps. Just use regular ChatGPT, use the vision, take a bunch of pictures of your favorite ads, have it analyze those ads, give you the key points of those, give you the best hook that it could find, even come up with a formula to repeat those type of hooks that you like best from your own and from the other ones you found. And I sort of build a little bit of a system like that. But then they went into my knowledge base. So when the GPT was built, it had that. Now when it comes to creating the visuals of the ad, I usually use a whole different GPT for writing the script. A lot of my ads are me. Since I make YouTube videos. I, you know, I have a little bit of a, even in the Facebook side, people recognize me from YouTube. Yeah.

[Michael Stelzner]

So it helps me special advantage. Yeah, exactly.

[Saj Deeb]

Yeah. So I have my face in the ads, which I know some people don't like to do that, but so then I write a script, same kind of way. I train the GPT for the copy, but it's a different GPT for the script copy. And that one, it's not specific to YouTube or to Facebook because my YouTube ads are the exact same. Typically there's this is like short format, 62nd, 32nd, 62nd ads. And those work really well. I mean on YouTube they work a little bit better for me because I could run it to the warm traffic of my YouTube subscribers. So little benefit there.

[Michael Stelzner]

This is another crazy question which I don't even know if you've tried, but a lot of people that might be listening that already have GPTs might not have the best system instructions. Have you thought about asking ChatGPT to an or Claude to enhance the instructions and then for lack of better words, creating a new variation of it and testing out whether or not the new instruction set is better than the old instruction set. Do you recommend stuff like that?

[Saj Deeb]

Yeah, 100%. Yeah. Either one. You could use ChatGPT or Claude. Just go to your existing GPT, click edit, go to the page where it shows you the system instruction, copy and paste that, and ask ChatGPT to analyze and improve that. It literally will do a much better job. And the one thing is, if you come up with exactly what you're looking for, the problem with the GPTs a lot of times is you're just not giving it enough very clear information. And that's the general thing that I've seen across. Anybody that uses AI, they wanted to read their mind a little bit too much. Like it's just not quite there yet, where it just knows what you want. It has a limited memory, it totally forgets all the conversation you had with it for the most part. But you're like, no, I talk to you every day. Why don't you remember? I need this.

[Michael Stelzner]

Yeah, yeah, that's important, Right. It's going to act like it's talking to you for the first time every time you run it, right?

[Saj Deeb]

Almost every time. Right. ChatGPT has a very limited memory, which mine is always capped out. I have a message on top of my CHAT GPT memory full, and for some reason sometimes saves the things I don't want to memory. And I kind of gave up on trying to get the memory to be a functioning thing. But if you do that, if you just do what you said, take the system instruction, have ChatGPT or Claude improve it, and then copy and paste it back in. Nowadays it does a really great job because it formats it in the exact right order on how it should be. That whole formula I talked about, that's like kind of how OpenAI, every time OpenAI releases a new model, by the way, they give you a way you should kind of prompt it. For example, with ChatGPT, you needed a lot of information, but with the reasoning models, it needs a whole lot less information. So the whole prompting formula kind of changed. With custom GPTs, if you just follow those four steps that I talked about, if you look at what ChatGPT gives you as an output for your system instruction, you'll see that it's actually following those same things. It'll follow if it gives you a bunch of emojis, by the way, which it does almost every time, just ask it to remove it because the system instruction doesn't need any emojis.

[Michael Stelzner]

Okay, let's talk real briefly about teams and the benefit of custom GPTs. Talk about that real quick.

[Saj Deeb]

I actually pay for some reason for two separate chat GPT subscriptions. We have a teams plan which requires you to have two people in a subscription. You can't have a teams plan with just one person. But the one benefit you get with the teams plan if you have two people. I know some people even because they want some of the benefits of teams, have it for themselves with two accounts, but with teams, every time I'm the guy that makes custom GPTs for everybody in the team just because I'm not going to go tell them to figure this out. Like I've been doing it for a couple of years. So I go build it. And when it comes to saving a custom GPT, you have five different options to save it. You could save it for yourself. So any you don't want to share. But my favorite part is saving it in our workspace which with the teams plan, you save it in the workspace, then every single person in that workspace, like our Skill Leap workspace that we pay the multiple subscription for, gets my GPT just shows up on their account on the left side. So then it was like, hey Andrew, I just built this custom GPT. I know you're doing the LinkedIn post now. Can you just use that from now on? And then he'll just use that from. I don't have to go explain to him how we want to do it. I just do it and he just takes it over from there. And the GPT store is another way. I've shared it before. I have a few different public GPTs that have done pretty well, but I think they sort of gave up on the GPT store. So if anything, I would recommend people upgrade to the teams plan if they want to share this. And the reason why I like it, you could technically share with just a link, but if you're having sensitive information in your custom knowledge base and custom anything that you have inside of there you don't want other people to get access to. I like the teams sharing option because it's private. Even if someone gets a hold of that link that they can't really access that GPT in any other way.

[Michael Stelzner]

Well, and talk about the at mention.

[Saj Deeb]

Oh yeah, as I mentioned, like once you get past like 10 GPTs, they all show up on the left column inside of ChatGPT. And I've tried to like name them in a way where like I'll put SL Marketing Navigator. So I know it's for Skill Leap because I have different companies and I'm like, oh, this is the wrong one. And I've trained it based on our knowledge of what skill leap is. So if you made copy for hotfinity, it wouldn't make any sense. It doesn't know what happens. So I would put HWB and then I started getting confused. I was like, I can't ever find the right one. This is now taking me more time to figure out what I'm doing than I just using chat GPT. So inside of the chat you could type in the ention sign and it pulls in all your recently used GPTs and he's searchable. So I'll just type in SL. Well, the YouTube ones they say SL, YouTube and whatever the thing is, those come up right away. I click on it and then I just do the title one first. Then I do the description one all in the same chat.

[Michael Stelzner]

And you don't need a teams account to do this. This is a function of any chatgpt.

[Saj Deeb]

That is not a teams account related. Right? That is, anybody could do that one.

[Michael Stelzner]

Okay. Oh, so what I'm hearing you say is that if you've got custom GPTs that you've created or that you regularly use, you can bring them into an existing chat for lack of better wor that what I'm hearing you say?

[Saj Deeb]

Yeah. The one thing I noticed though, it doesn't do, I actually tested it because I was curious with the deep research. I was like, can I start a deep research and then bring a GPT to then utilize that deep research? It doesn't do that for some reason. Like he ignore, ignore the deep research even though it was part of the same chat. But what I use it for is let's say like that YouTube process is actually multiple GPTs. Right. I have title description. You could even make one to give you like three words that go in the thumbnail. That way you know how to design the thumbnail. You can make one for that. So every time I want to come up with one, a lot of times I do it for sponsored CTA too. So if I want the CTA sponsorship to have a certain format, it's real quick. So in the same chat I'll just use the admission sign like three different times. And then I don't have to open up three different GPT since it's kind of the same workflow. That's sort of how I use it.

[Michael Stelzner]

Saj Adib Wow. We have given a lot of insights to everyone that's listening that is using custom GPTs or is about to use custom GPTs. There's going to be plenty of people that want to connect with you. Where's the best place for them to connect with you? And if they want to work with your business, where do you want to send them?

[Saj Deeb]

I think the best place would be LinkedIn. So my name's Saj Adib, I'm on LinkedIn, I'm on X under the same name. So that's a place where you can connect with me personally and skill. Leapai.com is the website where we have the library of courses. We have a custom GPT course, but pretty much everything, anything new that comes out, we try to be the first to kind of master and make a course for it and make it more applicable too, for entrepreneurs.

[Michael Stelzner]

Daj, thanks so much for coming on the show today.

[Saj Deeb]

It was a pleasure, Mike, thanks.

[Michael Stelzner]

Hey, if you missed anything, we took all the notes for you over at social mediaexaminer.com a49 be sure to follow this show on your favorite podcasting app. And if you've been a listener for a little while, would you give us a review or let your friends know about the show. And do check out our other shows, the Social Media Marketing Podcast and the Social Media Marketing Talk Show. This brings us to the end of the AI Explored Podcast. I'm your host. Michael Stelzner will be back with you next week. I hope you make the best out of your day and may AI help you become more successful. The AI Explored Podcast is a production of Social Media Examiner.