Leveraging AI

84 | AI-Powered Content Marketing: Building an AI Automated Content Funnel That Sells 24/7

• Isar Meitis, Aaron J. Steel • Season 1 • Episode 84

Share episode

Are you struggling with content creation overload?

In the high-speed world of digital marketing, the right content can turbocharge your business growth, but who has the time to craft that perfect post for every platform?

In this episode of Leveraging AI, Aaron Steele, founder and CEO of ENDGN shares his expert insights on leveraging AI to transform content into a lead generation powerhouse. We talked about how AI and automation tools are revolutionizing content creation, making it possible to generate high-quality material efficiently—leading directly to increased leads and business expansion.

In this session, you'll discover:

  • How AI can streamline content creation across multiple platforms.
  • The step-by-step process to implement AI in your content strategy effectively.
  • Real-life examples of businesses that have transformed their content workflows using AI.

Aaron Steele is the mastermind behind ENDGN, a company that helps business owners and influencers turn their content into a robust mechanism for generating over 3000 leads every month. Despite it being way past midnight in Australia, Aaron brings his vibrant energy and sharp insights to our discussion, making complex processes easy to understand.

About Leveraging AI

If you’ve enjoyed or benefited from some of the insights of this episode, leave us a five-star review on your favorite podcast platform, and let us know what you learned, found helpful, or liked most about this show!

Isar Meitis:

Hello and welcome to Leveraging AI, the podcast that shares practical, ethical ways to leverage AI to improve efficiency, grow your business and advance your career. This is Isar Meitis, your host, and we got a really fun and important show for you today. We all know that creating valuable content and distributing it in a strategic way, meaning the right content on the right platforms at the right time can drive a lot of value and can drive business growth and leads. Problem in that statement is that generating a lot of high value content in the right format across multiple platforms is really I'm consuming like anything from doing the research on pinpointing the exact target audience to creating multiple formats of the types of content that it will fit the right platform to distribute, distributing it across these platforms, scheduling it, and so on. It's just a lot of. the good news is that using AI and other automation tools can make this entire process significantly shorter and more effective, still achieving the results for which it's done, which is driving leads, creating additional influence and growing your business. Our guest today, Aaron Steele is the founder of CEO of Enjin, which is spelled E N D G N, but sounds like a car engine or a growth engine if you want, is a content lead generation ninja. And what he does is he helps business owners and influencers and content creators to generate 3000 plus leads every single month by Creating this social media posting machine. And so when I saw his stuff and the content that he's posting, I was very excited. And that's why he's here today. Aaron, welcome to Leveraging AI. Thanks for having me.

Aaron Steel:

I'm glad to be here.

Isar Meitis:

And I'll say one more thing. If you didn't pack. Pick the accent from the first sentence. He's in Australia and he's a real trooper because it's like after midnight for him and we're recording a podcast episode. so kudos for that. No, let's really dive right in. There's, I know there's a very structured process that you Teach people and help people put in place. Let's just follow your process step by step. if I want to create more content and have it very well targeted to my audience, what are the steps let's start at the beginning?

Aaron Steel:

Yeah, sure. So, it's, it's been a bit of an iterative process over the last, little while about, finding the best way to, to produce and scale that content. so I guess the important thing to, to note with. Content is, this is AI generated content. Like we're not trying to hide that, but I guess the key difference is that we do a lot of work. I guess you could say pre production work, and get lots and lots of context and training data to use in order to make sure that the AI content that we produce, actually sounds like you. And is. Modeled on best practices on what to post, how to post and that sort of thing. we're really taking a lot of the guesswork out of, that whole process and really just step people through. So what that looks like from a client's perspective is we have it's very important that we capture the true natural speaking voice of our clients so that it's not just polished writing, that they've, that they've used ChatGPT to answer questions in a form it's, literally having a conversation like this. whether it's live or like through the use of, voice forms where we pre record the questions and we get them to actually respond like this so that we can actually capture the transcript of how each individual talks, because, that's really what, we're all separate. what we believe separates us from other agencies out there, which are utilizing content, creation through AI is there. They're not really the focus. I feel like they're focusing more on the technique of producing lots of content as opposed to quality content. and so when you, once you go through the quite a lengthy process of answering like really in depth questions and my business partner, she is the master of this. like I focus more on the technical execution and the distribution side and she focuses on like the content quality vertical. And so she really spends. So much time and effort getting like extracting all the sorts of information out of people. so that we have so much rich content to work with when we're actually creating, the posts for them. So let's say once we've completed everything, we've gone through, just

Isar Meitis:

one second. Yeah. What kind of questions are being asked? So let's say I want to do this myself and I don't want to hire an agency. What kind of questions are being asked that I can. Practically ask myself and then answering the natural voice. So I have that information. So what are the kind of questions that you guys are asking in order to, I would say, qualify what is the tone that this person usually speaks out?

Aaron Steel:

Yeah. So we don't have a set question list. it's bespoke to every client that we work with. So we, based on what they have. what we know about them already. We'll guide the questions based on that, but it's, it really dives into, like their thoughts, feelings, passions, like what, how they see the world, how they see their business, how they see their customer, what the actual everyday life is of The customer. I can even bring up a few or sanitize them through examples of the sort of questions that we ask. in a moment, but it's, yeah, it's really trying to bring out that conversational tone. It's basically asking people

Isar Meitis:

about their business and having them respond about their business, but from a very personal perspective.

Aaron Steel:

Yeah, I would say we, we mesh the business side of things with also the personal brand. Side of things as well, so like we're we don't generally work with really big corporate clients. It's more sort of the small to medium size where there is definitely still a personal brand element to it a lot of the time. And so, yes, we can ask lots of questions about, Their business. But also we want to know about them as well. So because they're going to be the ones who are posting the car are going to be, the name and the face on the content, whether or not it's them posting it or not. we want to make sure that it's unique and it's personable and it sounds Them, and it's not just gonna get lost in a sea of Yeah. AI generated content. yeah, it's, yeah. It can be hard to, to answer what sort of questions, because it really depends on each client like

Isar Meitis:

no. but I think I, I understand you're talking about personalizing or humanizing the business, right? it's yeah. Hearing from the person about their passions and about their business through their personal lens. And then you capture that. So you capture that as a transcription of these recordings.

Aaron Steel:

Yeah. So transcribe it. Yeah. So literally word for word transcriptions of what they say. So we're looking for the, the phrasing, the tone of voice, the way they talk, the vocabulary, the, even the, we'll get the AI to intentionally perhaps incorrect sentence structure that they might use because everyone talks it's all jumbled unless you're very good at speaking, like we will talk a bit all over the place. and using lots of words like that, and we're not saying like we will get. the AI to write, and all that sort of thing, like filler words, but we want to really capture, the unique tone and quality of that person. so it definitely sounds like them because, I think, I'm sure you can as well. It's very easy to spot things that have been written by an untrained GPT 4, let's say. Absolutely.

Isar Meitis:

Yeah. Okay, cool. so we have a transcription, by the way, which tool do you use to transcribe the meetings or the calls or the,

Aaron Steel:

so we use, got a couple of different things that we're working with. the main one that we use is probably a voice form. So that's a, it's a bespoke tool or not bespoke. it's a tool that we, It's built for that purpose, but, they didn't build it for us. but it's, yeah, like literally, like any other kind of form, but you just click record and you record your responses and it captures the transcript live. and it can also, prompt them for additional. responses based on what they give you. So it has a, an AI component, which we're working with them on to, intelligently ask more questions based on what they give us. it's a really cool bit of software that we've, we've hooked up with, the last few months, which, is pretty exciting. Otherwise we just use standard, zoom recording or something.

Isar Meitis:

Yeah. And then transcribe it. And then, okay. So what do you do with the transcription? So what's step two?

Aaron Steel:

Yeah. So step two, we have, we've just moved over to Make, instead of, Zapier. okay. we have a transcription PDF. So everything goes into a giant PDF that we save. So all the, Existing information, we just compile it like the transcript is the largest part, but any like website information or anything that we can get, we compile it into a giant PDF. And this is To 300 pages. This PDF. yeah, like it's not small. and so obviously we need a larger context window, for that. So, anthropic Claude, is great, as well as, Google Gemini cause it has a million. Yeah, Gemini Pro 1.

Isar Meitis:

5. It's the verge of endless at this point.

Aaron Steel:

Yeah, it's well, it's definitely not endless. Not the way I've used it. yeah, so We take that PDF, and basically I'll get it to, I'll attach it to a form, which I'll submit with, the client's name, bit of background, here's the form, I submit that, and, from there, the make scenario or automation will, use Python to extract, the text from the OCR, because if you usually we're working with so much text that you can't just dump it into a spreadsheet, it's too many characters. so we have to bypass that limitation, by using that PDF. and then, yeah, so we use a bit of Python to, extract the text out. And then, I've just recently. cooked up a way to have, conversation history with Anthropic by the API. So at the moment as far as, yeah. if you're on the web browser and you're using, Claude Opus or something, obviously you've got a conversation history, but if you ever want to use it with the API, you basically, every time you prompt it, it forgets what you've said and you have to you waste a shitload of tokens. like I've got a massive deal with Anthropic this month because I've been running huge amounts of tokens through it. but, we've just in the last few days, successfully built conversation history into, the Claude API interface that we're using, which is just like passing a post request or something. We can. So there,

Isar Meitis:

there's a function within the API to recall? No, there isn't history.

Aaron Steel:

Oh,

Isar Meitis:

there isn't? No. So how are you doing this?

Aaron Steel:

the, we're using, I'd have to ask my developer, but, Lang chain, I think they have a, a function that we're using to store conversation history and, so

Isar Meitis:

it captures, it's like a middleware built a line chain that captures the history of everything that was said.

Aaron Steel:

Exactly. I'll be honest and say, I don't know exactly how it works. I just, I don't think I'll

Isar Meitis:

get it. And I don't think the audience will either, but it was custom built for you. It's not like a tool I can go and use tomorrow.

Aaron Steel:

No, that's right. happy if people are interested, happy for them to get in touch with me and definitely, what is it we can make that work for people. But, and it's interesting it does Keep a conversation summary ongoing as well. And I was trying to, I was debugging with it and was asking, asked it for a conversation summary. And it was basically just whined to me about how the human kept asking for information without providing it. And cause I was trying to get like the variable through so that it could read the text and it wasn't working. And it kept Oh, we're being polite, but it keeps asking for the same thing. It was amusing how it had a bit of a suck to me about it not doing its job. but yeah, and so from there, once we, get that conversation history in which I haven't tested that heaps yet, because it's only been the last few days. We then run, multiple, Different scenarios in conjunction. So that'll split off into, it'll split off into one prompt, which says, okay, here's the text from the pdf. Here's all the context you need. He is very considerable prompt about how to write, say, a LinkedIn post. Here's some templates that we want you to model your, response on. and that'll produce a LinkedIn post. below that, it'll be doing the same thing for, say, an email newsletter, based on the content that we've got. And just go through the list of all the different content types basically, and Facebook, Instagram, Twitter, like we do tweets in there as well, and this is all, utilizing the writing style that we've given them like two, 300 pages of context of. So when you're actually, Producing these, articles and blog posts and Facebook posts that actually genuinely sounds like them, but you can't, you literally can't tell, like I've even got it to write YouTube video scripts for, YouTubers out there, excuse me. and. Yeah, you couldn't even tell that it was written by Anthropic, not him. cause it was, yeah. So I want to

Isar Meitis:

pause you just for one second, because there's two gaps, at least in my mind, that I want to understand how are you doing? Yep. So that 300 page document is basically just reference for whatever AI model. And because it's so long, it's probably has to be either, like you said, Claude three or, Gemini pro 1. 5, for those of you who don't understand what we're talking about, they, these AI engines have a limitation in the amount of memory they have in each chat or in each, API call. if it's on the API side and it's limited. Two it's called tokens, but a token is about 0. 7 words. So Open AI just actually increased theirs to 128, 000 tokens in GPT for turbo. That means about a hundred thousand words. Claude three has 200, 000 tokens, which is about a hundred and 50, 000 words. So it's just a lot more words and Gemini 1. 5 pro, which to get to it is not the regular Gemini, but like their backhand, yeah. Yeah. A tool that you can play with. It's like a sandbox to test their new tool. It's not the formal release, but it's available. And it's working, is a million tokens, which about 750, 000 words. So if you want to push a 300 page document through, then you need the larger context window. So that's like on that, but, so this is only the reference, right? This is only this is the style, the tone, the voice. This is how we speak this. Okay. How do you pick the topics? the content you're going to generate. Okay. So I understand now we have the reference on the tone, but what are the actual topics that you generate? And so that's question. Let's start with question number one. How do you start with the topics of the content?

Aaron Steel:

Yeah. So a couple of different ways. So for example, I haven't asked for permission to mention his name, so I won't, but there's a fairly well known YouTuber, in the AR space that I'm working with, and we take an extract, or sorry, every time we post a video, It'll automatically send the transcription to us and we use that transcription as the topic. So we've already got the context library of how he talks based on 150 video transcripts. that's a big document. And, we've already got the transcript from the most recent video. And we say, Hey, here's the existing history of how he sounds. Here's. The, the topic that I want you to talk about. and then we just push that through. The rest of the mechanism and it will create the content based on that video. But that's one way that we do it. So one

Isar Meitis:

option is basically you start with a one type of piece of content and really just use your engine to pun intended repurpose it to all the other different options. And you usually, I assume, start with whether it's that client or another with a live, sorry, without video recording, because that can be converted to basically anything. Yeah.

Aaron Steel:

Yeah, so yeah, there's, I guess there's the ideation stage, which where it's an ongoing development, I suppose so, but you can like the very early days, the way I had this set up was, I would just get ChatGPT to say, I'd say to a look, I'm an AI entrepreneur, ask me 10 questions so that I can answer them like I'm on a podcast. And then I just recorded myself and each of the answers to those 10 questions was a different topic. And that's what I pushed through. And That was the content right there. Now I am a bit more sophisticated than that, but not much. I, yeah, like we'll get, still using AI a lot of the time to generate content ideas, but using, more like book style, titles rather than just a blank title. and it will generate the content based on the hook and the existing content, but. Before we get to that stage, we also have, I've got a robot that will every day, based on our context and like it has an understanding of what we do, or what the client does, it will generate, different keywords and it will then go and scrape the Google search results as well as like a variable, keyword like opinion or news or, whatever, there's a few different things that we use, and it will take the contents of the search results that it gets back from those, and then compile keywords and, topics from those. So it's an ongoing thing that happens every day, basically, then it creates more content topic ideas. And then it will reflect on that and then ask, provide questions. To respond to about those as well. So that's like an ongoing thing that, it's a dynamic that depends on what we tell it to search for basically. But, yeah, so it's pretty unique.

Isar Meitis:

What tools do you use in that process? Like for this curation of ideas?

Aaron Steel:

Yeah. So I'm still using Make for that. So I, plug in, browse AI, which is just like a web scraping. interface. And so you have a whole stack of pre trained robots as they call them. and so you can say, Oh, I just want to have it like a Google search robot and you just put in the keywords that you want, and tell it how often you want it to go and collect those search results and it goes and does it, or you can, connect it on Make and it will, integrate with the rest of your processes. So I had that scheduled for every day and then it just pushes the search results through based on what ChatGPT or Claude or whatever we're using in that particular use case, tells it to. So, yeah, it's pretty, pretty handy tool. there's, yeah, I'm sure there's ways we could build like a bespoke word scraping application, using Python, but I'm happy with this at the moment.

Isar Meitis:

So for those of you who don't know what Make is or what Zapier that was mentioned earlier, these two platforms are probably the biggest, most commonly used automation platforms in the world today. And they're basically the glue between everything and everything on the internet. So you can take data from any, it's not any platform, but they have literally now tens of thousands of platforms they're already connected to. So you can take data from one source, let's say Google search and push it into ChatGPT and then get the output from ChatGPT and put it in Google sheets. And then take the data from Google sheets and aggregate it and send it to a PD. Like literally anything to anything, your CRM, your email marketing, automation tool, like whatever tools you have, most likely are already in there. And you don't need to be a developer to actually make those connections. It's basically telling you the type of data you want to bring first name, last name, topic, idea, content, like whatever it is you want to pull. And then it knows how to send it correctly between all these different platforms. So it enables people who are not developers to build really cool and sophisticated multi step automations. like Aaron just mentioned. So you can take the data from, you tell it what to search at what frequency with one platform. You take the content from that, throw it to ChatGPT. It knows how to summarize it. It puts it into a word document that is saved to a Google drive that can then trigger something else. So you can build all these things that can dramatically reduce the tedious work that you otherwise would need to do just by copying stuff and moving it around and converting it to different formats.

Aaron Steel:

Yeah, absolutely. And just to make note of, I agree. Yes, you don't need to be a developer, but you really don't need to be a developer full stop. And this is still a bit of a controversial opinion. But, I've built a lot of stuff with Python and I am definitely not a developer. And it's just by working with Multiple, language models like ChatGPT and Claude and just saying, this is what I want to do. Like I literally sketched out, the diagram, like I've got background in, business analysis and all that. So I'm not completely clueless with how to represent. Idea and technical ideas. But, I sketched out a diagram, like a data relationship model of how I wanted, the software to work and said, this is gave it to Claude and said, this is what I want. Can you code it up in python for me? And then, got it to walk me through how to host it on my own. My device and then put it up into Heroku and, and then I got 90 percent of the way there. And then I said, okay, I'm, I don't want to, I don't want to do this anymore. Gave it to a developer who was like, Oh, and fixed it. And, yeah, so like you can go a long way and. The gap between developer and non developer is getting smaller and smaller with these tools. It's I

Isar Meitis:

agree with you 100%. I'm just like you. I've never written a line of code in my life and I'm doing really cool stuff short. Like I'm not, recreating a FIFA 2025 or something like that, but for daily small applications that do stuff that I need, I create code with Claude and ChachiPT regularly. And when it fails, I go back to it and tell it what happened. I'm like, okay, this is what happened. It's giving me error 0, 3, 6, 1, blah, blah, blah, blah, blah. And then I tell it this oh, okay, try this code instead. And then it's giving me a new piece and I don't have a clue what, where the code is good. I'm like, I don't know how to read it, but I'll try. And usually within four or five, six attempts of back and forth, it will build something that's actually working and doing what I need it to do.

Aaron Steel:

Yeah. And the next iteration of that workflow though, is, Agent or agentive workflows, which I've been working with as well. So instead of your I writing in the, going back and forth between The code and the language model, you can set up a big, like a smart, expensive language model, like Opus, Claude, sorry, Claude Opus, and giving it the instructions of what you want it to do, and then it passes that task to a smaller developer model, sorry, a smaller language model, and then they can bounce back off each other until, and they're able to test it, like you can equip it to test, The code and, you can, it can test it and review it and do all that stuff itself like very fast. And then it will spit back out once it's done the working code that you would have had to have gone back and forth. manually, and now it's taking care of that process for you. And that applies not just code, but like content and all kinds of things. Like the agenda of workflows is the next big thing. I think, I'm not the only one who thinks that as well.

Isar Meitis:

no, absolutely. It's going to be probably as big as a jump as ChatGPT was when it first came out, when real agents will work properly. But let's go back to our thing. I think I, I threw us on this, let's go back. so quick summary, you train the model by giving it a lot of background on how the person talks by literally letting the person talk. So that's step one, step two, you figure out what is the content, what is the content you want to create either by content that person is already creating or the company's already creating, or by doing an iterative process with some automation that goes to the internet and looks for specific topics that are relevant to the industry, the target audience, the time, the month of the year, like whatever other information is relevant to your industry and your audience. And then you have the ideas for the content. Then you feed that into. All of this together into Claude or GPT, 1. 5 pro, and then it spits out multiple pieces of content, that are relevant to whatever it was created. Can you share the exact prompt you're using in those tools? can you open it up and say, okay, this is what you do, because I think it will give people a very good context on the how, not just the what. Yeah. Yeah,

Aaron Steel:

no, totally. I'll, if you give me, forgive me for just opening my browser up just quickly, I'll read out some of what's going on. and the, just while I'm doing that. So the process I've set up an interface, let's say app. engine. com. So app. endgn. com. I've built like a, a SaaS platform where you can go in there and you can do all that. The exact process that we just discussed yourself. You can upload your context yourself, your writing style yourself. and you can edit your prompts and edit your writing strategies, templates, everything that we do behind the scenes. And you can connect up, your social accounts to it and you can just push content out, that way. and, SAS. Yeah. Yeah. Awesome. Like it's pretty, pretty hands off. like the clients that I work quite closely with, I will build a similar kind of, Yeah workflow with for them, but it'll be specifically for them with their own prompts and all that kind of thing. So it's We're bypass the sass part and just hook it straight into their socials But and then yeah, there's like a an approval Kanban board that all the posts land on which you can go in and just drag it over to approved Okay, so I've got a Make scenario here called blog generator, which, So this generates a, so this is, I won't give too many details cause this is for a specific client. Yeah, just go beep

Isar Meitis:

every time there's a name of a person. That

Aaron Steel:

would be fine. Or a business. Yes. So we go, this will generate, five day, like a five day challenge, a blog post, free assets, like a newsletter, a quiz and social media posts, all in one. So a five day challenge, let's pull up the prompt for that. Alright. You are an expert ghostwriter. Your job is to first conduct an extensive review and analysis of the content of this document, taking care to note the unique style of writing, speaking, turn of phrase, content expertise, and manner of speaking, and then it gives the variable for the PDF content. put that in quotation brackets. Now that you have become the name of the client and understand their style of writing interest experiences and verbal mannerisms, your job is to create a five day challenge related to their business aimed at coaches and consultants doing, two to$5 million in revenue. Who are at a growth friction point challenge should educate participants of how lost profits are impacting their business. And so it goes through some pretty context specific stuff about the client. I'll skip through that, but basically, yeah, most importantly, right at exactly how name of client speaks drawing on the unique story insights and way of speaking to build connection and credibility, make it interesting, exciting, and highly valuable and actionable. The goal is for participants to think if this is free content, the paid workshop must be amazing. And then in capitals, do not sound like an AI, do not respond with anything other than a requested text, no preambles, no conclusions, no pleasantries. I do that because. AI always has a habit of saying, okay, sure. Here's the, I'm like, no, I don't want that. And I don't want the, like here's the thing that I've just written for you. Like it always tries to do that. I'm like, no, I don't want that. Like just give me the text. Yeah. Cause it'll find its way into the post otherwise. And it's you have to then cut it out with a, like the, Structured text extraction module with GPT and it's just a pain. So I'm always yelling at it and being mean to the the language models, but they're pretty useful. Question,

Isar Meitis:

question, two questions about this one. So this is running it on Claude. You said, So this, So

Aaron Steel:

that's,

Isar Meitis:

the Claude,

Aaron Steel:

IPAS three prompt.

Isar Meitis:

And the follow up question, now that I know that it's running on Claude and I know what you're sending is, does it actually do it every time? Like one of the things that I see, like I have a lot of GPTs that are running for my business and so on. So it's a similar thing, right? It's a, GPT is just a fancy recurring standout standard prompt, right? That's what it is. with some background information, the. The problem that I'm running into is that every now and then it will not do exactly what you tell it to do, at least the way I see it on ChatGPT, I've never tried to run something regularly on Claude just because GPT is currently only run on ChatGPT. So do you see it that sometimes it will write that extra stuff or these kinds of things?

Aaron Steel:

it's pretty good. Usually I haven't encountered too much of that. like I have, I do put in place. like that identity of workflow as well. So I'll get it to check its own work, after this so that, and I'll say, here's the criteria, make sure it doesn't have the preamble, make sure it doesn't do this. It doesn't do that. Does this, it meets this criteria. and if not loop it back and get it to do it again. So

Isar Meitis:

you send the outcome of this through make to a different conversation. Yep. Verify that it's really doing what it's supposed to do.

Aaron Steel:

Yeah. And that's brilliant. I use a less expensive model than Opus. yep.'cause yeah. Opus is not cheap. and yeah, like I'll get it to, to loop back on itself a few times if it needs to or and basically there's eight approval gates, and if it approves it the first time it goes off and it goes to get scheduled, it has eight chances to get it right basically before, it gets to the very end. If it still can't get it right after eight sort of, pieces of constructive, Feedback. it still just goes through to, the scheduling part and just flags as we couldn't get it right sort of thing, but that hasn't happened. So that's probably just being over engineered, but so I want to,

Isar Meitis:

I want to now explain to the people what we're just saying and as far as pricing and so on. So each and every one of these models, when you run them through an API, which is what Aaron is doing. So he's not going and copying stuff into the chat that you usually go to. He's sending it through an API call, which has multiple tools that can do it. Or even through, Make and Zapier and N8N, like all these tools can talk to the APIs. there's a tool I really like that you may or may not know that is actually pretty cool. That's called, OpenRouter. And what OpenRouter enables you to do is just like a hub of all these large language models for APIs, and you can connect just to OpenRouter and then call more or less any large language model through an API. So you have one API connection, and then you can call, by name, whichever large language model you want to use. The other thing that it shows you the pricing for each one. So to put things in context of what Aaron said, that Claude 3 Opus is very expensive. If it's the most expensive tool out there, both in the content that it's getting in. So you sending it content such as those 300 pages, as well as in content coming out. and again, to put things in perspective, an open source model, a good open source model, like Mistral 7B, which is the name of a large language model costs$12 and a half cents for a million tokens. So we spoke about what are tokens. So 12 and a half cents for a million tokens. and, Claude three Opus is$75 for the same US

Aaron Steel:

dollars as well. So it hits us even harder.

Isar Meitis:

So it's just to put things in perspective, that's obviously about what, like 700 times more expensive than running it on Mistrall, the quality is there. And the context window is there. but like Aaron saying, there's ways to trick it. if you want to just test the content, you don't have to run it through a$75, a 10 Million tokens, like option. You can run it through cheaper options. And the trick here is always to optimize quality and cost, right? So you want to have the right tools for the right thing that do a good job. It doesn't have to be the best job. You just have to be good enough job to get the job done in this particular case, good content that drives leads to a business. If you can do that for$12 and a half cents instead of$75, a good way to do it. So picking up different steps and running it through different of these lenses in order to get the good outcome is a brilliant way to optimize both for quality and for costs. So kudos. Yeah. Great idea.

Aaron Steel:

That's perfect. Yeah, that's why I'm so excited about getting Conversation History working with the Anthropic API as well, because one of the reasons I was getting hit so hard with API costs was because if I had, if I'm wanting to generate 10 posts with 300 pages of context for each post, I was getting hit for all the tokens every single time I ran it. Now that I've got a conversation history component, I can just run that context once and then just like you do on the web browser, just keep hitting the same conversation. and it remembers the context and it doesn't need to be reminded or, starting a new one every time. So it's significantly cheaper. then, the way it's set up by default. I'm pretty excited about that.

Isar Meitis:

Makes perfect sense. Yeah. yeah. Okay. so now we have figured out the tone. We have figured out what the content that we want to create, we have created the content, we check the content. And then you said it goes to scheduling. I assume, again, that's a connection through make that goes to a scheduling software, whichever, whichever one the client uses or that you use, I don't know which one you're using.

Aaron Steel:

I use Metrical, because they allow us basically from an agency perspective, they allow us to connect up to 25 brands on the price point that we're on. which means each brand can connect basically all of their socials. And, we built a little integration so that when we, because everything goes over to Airtable when it's like ready to be published. when an Airtable record or a post gets, approved, it will get pushed to a auto list, which I love in, in Metrical. So my like, I'm not, I've got Hootsuite as well, but I don't use it anymore because you still had to specify the schedule. every time we wanted to post it, whereas in metrical, I can just set up predefined times and times of day frequency. Days of the week, and it'll just act as cues that I can just send a hundred posts to, and it will just post them at the predefined times, which is great. It's one less thing I have to think about. so yeah, that's the tool I use. I'm not sponsored by them or anything like that.

Isar Meitis:

Yeah, no, it's I love this kind of advice. this is real life, real people, real use cases. it's the best advice there is.

Aaron Steel:

Yeah, that's, and then, yeah, like it's, that's. that's the written component of the post. I also have, an image component. So LinkedIn, I will, collect like 150, 180, roughly, like selfie type images of myself or the client, or, that can be more photo shoot type images and just, have those set up in a Google drive and, I'll just get Make to collect, produce a random number and then like use that number to determine which photo it pulls and then just post that photo with the post on LinkedIn. and that's a really basic use case for it. But, yeah, we also can, use tools like, create a mate, which is like video publishing. Which with dynamic text and images and videos, I'm, I use that to create like carousel style posts on LinkedIn. and the other one that we use for still images is, switchboard. Canvas or switchboard. ai. I'm not sure what they're actually called, but, switchboard. ai I think is the name of the actual site. but again, that is like dynamic images and texts. if I'm the way I would use that is I go in and I set up these templates and you can actually connect to Canva within Canvas. Switchboard AI and do the design in Canva, then just delete the text out of the design, pull it back into switchboard and then like position dynamic text boxes where the original text was, and then you can pull it out from Make or, it doesn't integrate with make directly, so I'd have to use the API, but, in Zapier, You just pull whatever text you want dynamically straight onto the image. and you can do like quotes or whatever. and it's, yeah, it's a really good tool which doesn't get talked about enough. but it makes the whole image component of socials automatic as well. yeah, a little bit of a very, very cool.

Isar Meitis:

I have a question for you. Did you try using one of the image generators APIs in order to generate images based on the text in the actual, yeah, post. Yeah. So I do this manually now. I've never tried to do it through the API, but I would use let's say I use ChatGPT to have the whole conversation on a post I want to write or a presentation I want to make, I will ask it for recommendations on what might be a great graphics for this kind of audience with this exact post. And then it gives me three ideas. I'm like, Oh, I like number two, go create it. Or I like number two, but I want to add this and that, then they will go and create the image for me. So did you try that like at scale through the API.

Aaron Steel:

Yep. So I've got two ways that I do that. So the LinkedIn way that I, I do it, or not, it doesn't have to be LinkedIn, but like I was trying to be a bit less creative on LinkedIn, but I am just connected to the, unsplash API, which is like a stock images, site and would get, like GPT 3. 5 to do a quick, summary, get it to review the post that was written and just say, I need you to spit out five keywords. and then I would just pass those keywords through to unsplash, get it to select a random page. So I don't get the same image every time. and then just pull that, stock image based on the content of the post. Post through. so you're using,

Isar Meitis:

you're using stock photos and you're not generating AI images on the fly per post?

Aaron Steel:

I've done both, so yeah. The, I also would use, I use Dali a little bit, but I think I've moved, over, yeah, I've moved over to Midjourney. Yeah. and playing around with stable diffusion as well. But yeah, same kind of thing, using the content to generate posts. And then I just have a retry. function where if I don't like what it's come up with, I just click a button or drag it into the retry column or something and, it'll Get rid of the picture and come and do another one. yeah, I've done multiple different ways, to generate those images. I guess it, it depends on what the client for looking for. I think people are still a bit scared of doing AI images, but like mid journey version six, you basically can't tell anymore that it's

Isar Meitis:

incredible. I want to. Summarize quickly, because you just mentioned a word that makes a lot of sense to me, and you mentioned it before, and I want to add a little more context about it, about the Kanban board and the columns. So what reading between the lines of what you're saying, there's a Kanban board that really shows the human. Operator, what's happening in the process that all these AIs are doing in the back end. And I assume that some of the process is done through make that can move items from one column to the other. And some is a human in the loop like confirmation phase that saying, Oh yeah, this image is awesome for this move forward, or this is shit. Go get me another image because it doesn't make any sense having a rabbit when I'm talking about a business, whatever, something. And then we'll go and get a different image and then you can get to review it again. So I think the beyond every magic in each and every one of your steps, which I really find magical, I think the Also building a framework that allows for human intervention in those different steps while you can still monitor what's happening and sending backwards or forward, is really important because these models as good as they are, they're getting better and they're amazing and probably nine times out of 10, you don't have to do anything. The one time out of 10 may make you look like an idiot in front of your audience, but you may not want. And or maybe not nine out of maybe not one out of 10, but maybe one out of 30, but it doesn't matter if you're posting once a day, that's once a month. It's going to look like an idiot. And so having that human monitoring option, which is you, again, I think using a Kanban board is brilliant because it makes it very easy. You just look at this column and you see the final product and saying, Oh, this makes sense. Move forward. That's it. It takes you three seconds to monitor every single and then the machine keeps on working and doing its thing. Yeah. This was Amazing. I'll let you, if you have any summary, anything you want to add, anything else we need to know about this process, that we didn't share yet.

Aaron Steel:

so I guess the, the other element of what we do, is once we've got all the content set up, we then will, Basically facilitate and automate the whole sales process as well. So once you've got the content that is attracting leads and traffic via SEO and, social media, we also can integrate. Chatbots into emails and messenger and all that kind of thing to facilitate the sales process and make sure all the leads are being followed up on and nurtured, and, using agents to, book in, sales calls for people and all that. So basically, To give you a real life example, like I'm working with a law firm, where, everything's very manual for them at the moment, and we're working with them to automate everything up until the point where the client walks in the door, for the meeting that's already been booked. It's no longer going to be a case of them having to respond to emails and pick up the phone and respond to Facebook Messenger and respond to everything. And, client pieces go missing and, forms don't come in and that's all going to get completely taken care of and they still have the opportunity to intervene if they want to. but AI will use their knowledge of the legal firm that we're working with to respond to questions, taking care to not give legal advice, but, facilitate the whole process so that they're, the owners can, The partners can focus on, the value added task, which is actually the real human interaction, when the client walks in the door. So that's like the engine, like it is very much about that whole process. Like we, we want to make everything as, as smooth and easy as possible for people. yeah, that's very, it's a pretty holistic package,

Isar Meitis:

Amazing. Aaron, this was absolutely fantastic. Like really what you built is incredible and it's definitely the way of the future, right? Like I think every business will eventually move to work like this. You're just providing it now versus sometime in the future and you're providing it as a SAS already built, ready to go solution, which I think is amazing. If people want to follow you, learn from you, work with you, what are the best ways to do that?

Aaron Steel:

Yeah. probably finding me on LinkedIn, is probably the best way to do that. so my LinkedIn profile is just Aaron J. Steele. I haven't posted on there for a little while, like contrary to, what you might think. I've just basically had some personal stuff going on, which I've prioritized instead, but, and I've been busy building a lot of things here. But, and if you're interested in, the. Engine SAS application. You can go and sign up for a free account. It's just app dot E N D G N dot com. and yeah, feel free to reach out to me on LinkedIn. I'm always happy to have a chat.

Isar Meitis:

Awesome. This was great. Thank you so much. Have a good night, And I appreciate you taking the time and sharing all this amazing information with us. Yeah, it's my pleasure.

People on this episode

Head Shot

Isar Meitis

Host