Loading summary
Transcript95 lines
- [00:01]
Michael Stelzner
I've got some exciting news. We just launched the AI ticket to Social Media marketing World. It's like a conference within a conference. If you want to go deep in your AI learning, this is the live experience you've been waiting for. Imagine spending two days fully immersed in AI training. Grab your tickets now at social mediaexaminer.com Aicon 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 use AI. Today I'm going to be joined by Kyle Vollmer and we're going to talk about creating virtual assistants or AI assistants utilizing Claude projects. Now, you probably, if you've been listening to this show for a while, been hearing a lot of people sing the praises of Claude. Today we're going to talk about how to actually use their version of custom GPTs to create projects that, that allow you to do some really cool stuff. Also, by the way, if you're new to this show, be sure to follow us on whatever app you're listening to me on so you don't miss any of our content. We've got some great experts lined up to teach you more about how to use AI in your business. Let's transition over to this week's interview with Kyle Ballmer, helping you simplify your AI journey. Here is this week's expert guide. Today, I am very excited to be joined by Kyle Bulmer. If you don't know who Kyle is, is, he's an AI business educator and consultant. His newsletter is the Prompt Entrepreneur, designed to help business owners start and grow with AI. He's also the founder of the AI business, Breakthrough Academy. Kyle, welcome to the show. How you doing today?
- [02:09]
Kyle Vollmer
Hi. It's great to be here, Michael. Thank you so much for having me.
- [02:12]
Michael Stelzner
I'm super stoked that you're here. Today Kyle and I are going to discuss how to create AI assistants using Claude projects. Now, before we get on that rabbit hole and it's going to be a really discussion, I want to hear a little bit of your story. How in the world did you get into AI? Start wherever you want to start.
- [02:29]
Kyle Vollmer
Absolutely. Well, we can go quite far back, but I'll keep it relatively brief. So my journey with AI and my entrepreneurial background has been relatively mixed, I'm sure, like a lot of your listeners. So I actually come from an entrepreneurial background and a marketing background, and I've transitioned into AI over the last couple of years from an entrepreneurial point of view. I've started over 30 businesses. I've sold several, including a television station in Vietnam. That's a story for another time, another podcast. We won't talk about it now, but my main focus has been digital marketing, and I've run an agency in digital marketing for about a decade or so. I actually used your guides about 10 years ago when I was first getting started.
- [03:08]
Michael Stelzner
Very cool.
- [03:09]
Kyle Vollmer
When we were working out our kind of service offerings and stuff, we would go into your deep dive guides and be like, wow, this is, you know, fantastic stuff. So being with you today is really exciting for me. We ran this digital marketing agency for, I'd say, five to seven years, and it was going well. We became a little bit complacent, and in November 2022, we had the release of ChatGPT, or the public release of ChatGPT, the real turning point in artificial intelligence being publicly available. And at that time, I started experimenting with a few workflows, kind of messing around, using it to write Facebook adverts, for example, but never really crossed over and got really excited about it. I could see the potential and some of the potential risks for my digital market agency, but I didn't jump in feet first at that point. However, at the beginning of 2023, we actually lost two large clients. You know, these things happen. We'd become a bit complacent, and unfortunately, we'd put too many eggs in one, or in this case, two baskets. So we took a large hit to our incomes and to, you know, our lifestyles and also the company as a whole. So it was the coalescence of these two things, the emergence of AI in the background, making me think, there's something here, plus the push of losing two clients very quickly basically made me look at artificial intelligence a lot more closely. And for the last year, so this was the beginning of 2023, so the last year and a half, I've been really throwing myself into artificial intelligence. The main vehicle for that has been via Prompt Entrepreneur. That is my newsletter. The way that I learn things is I will tend to write. So I started to write publicly about what it is I'm learning and what can be used, especially by entrepreneurs and people in marketing in particular. So that became my focus. Every single day, six days a week, I was writing over a thousand words on artificial intelligence and specifically how it could be applied to actual Business problems because I saw there was a lot of newsletters and information about the latest model or the kind of the hype around artificial intelligence. But very few people were translating that into the entrepreneurial world, into the marketing world, and saying, okay, that's great, but this is what we can actually do with it. And that's been my focus, really, for the last year and a half.
- [05:28]
Michael Stelzner
Outstanding. So bring us up to speed, kind of where you are today. You've got your newsletter. What else are you doing?
- [05:34]
Kyle Vollmer
Oh, wow, that's a big question. So the newsletter has just hit a year. We had our first anniversary about a month ago. We have got 60,000 subscribers, which is fantastic. I also get to now fly around the world giving workshops on artificial intelligence to some pretty large companies and tech companies, which you would not expect. But every company right now needs this kind of information about how they can start to leverage AI in their company. I am also helping other people to provide my workshops around the world as well. And that's been very successful over the last few months as well. So it's been a very busy year. There's a lot of demand from businesses globally, in every single industry, every single niche about how they should be using artificial intelligence. So I've just been sitting at that border between artificial intelligence and entrepreneurship in business and helping to cross that divide.
- [06:24]
Michael Stelzner
Sweet. Well, I'm super stoked to have you here today. There are plenty of marketers and entrepreneurs that listen to this show and they have probably not yet embraced the concept of building an AI assistant, for lack of better words. Sometimes we call them custom GPTs. In this case we call them them projects because we're going to be talking about Claude. But why? What's the upside to this audience that's listening as to why they ought to maybe consider building an AI assistant?
- [06:52]
Kyle Vollmer
So I think AI assistant as a general term, it finally fulfills the promise that we were, we were given Back when we first got ChatGPT, the public release, a lot of people thought that we'd be able to use these AIs to save time on repetitive tasks, maintain consistency of output, to basically become an AI member of our workforce. And I think early on people would try with these tools and the quality just wasn't there. The AIs weren't able to consistently produce an output that could be used in a workflow. And therefore, I think a lot of businesses, especially early on, the people who experimented with the models early, they would bounce off, they'd be disillusioned because so much was promised and that Confidence was betrayed. So business owners and marketers tried the early models, didn't get what they were expecting and bounced off. Now, with AI assistance and what we're going to be talking about, Claude projects in particular, we are finally reaching that point where we can actually get consistent outputs so we can start to fit them into our business on a day to day basis. And it's actually becoming useful, I think for the first time for a lot of companies. There are obviously the custom GPTs which came before. What we're going to be talking about today has taken that kind of idea and run with it and has a much higher output, much more consistency and is able to match your tone of voice, which I know in marketing in particular is extremely important.
- [08:17]
Michael Stelzner
Yeah. So let's talk about cloud projects. Why cloud projects versus ChatGPT? Custom GPTs, what's your take on it? I've gone down the rabbit hole of cloud projects. I've not gone down the rabbit hole of custom GPTs, but I know you've also gone down the rabbit hole of probably both. So what's your take on cloud projects as to why?
- [08:37]
Kyle Vollmer
Taking a quick step back, Claude is actually a competitor to OpenAI's ChatGPT. Unfortunately, a lot of people don't really know about Claude I data recently, people tend to associate ChatGPT with AI. They're becoming synonymous in the same way that Google and search engines are synonymous now. And unfortunately that means a lot of people are ignoring Claude. So Claude is a model from a company called Anthropic and personally it's become my go to, it's now my workhorse. I use with pretty much every task. I know you've been using it yourself and you can probably talk to the quality of the output. It's also really good at coding and generally outperforms ChatGPT in many, many different roles. But people just aren't using it because ChatGPT has become the default. What we're talking about today though is something that's relatively new to Claude, which is called a Claude project. And it's a bit like a custom GPT which in OpenAI's ChatGPT is a way to encapsulate a set of information that you upload. So data, documents, blog articles, social media posts, whatever it is, you can upload a set of information and then have the custom GPT or, or the Claude projects work on that information for you, which is really powerful when we are building AI assistants because we can give them knowledge, we can give them a knowledge base.
- [09:57]
Michael Stelzner
I love it. Let's talk a little bit about kind of why cloud projects might be better. I mean, it's got some unique advantages and a couple limitations as well.
- [10:04]
Kyle Vollmer
Absolutely. And I'm so glad you brought up the limitations, because as with all of these things, we do need to address them before we run headfirst in and expect these AIs to do absolutely everything for us. There are still limitations. So probably one of the biggest, and it's a little bit boring, but Claude has a much larger context window than ChatGPT. A context window, without going into the technicalities of it, is basically the memory. How much information you can throw at this model before it starts to forget stuff from earlier in the conversation. If any of you listeners have used ChatGPT, they've probably got to a point when, if you have a very long conversation, suddenly ChatGPT begins to give you bizarre answers that don't correlate to the information you have previously provided. That's because it's falling out of the back of its memory. And unfortunately, ChatGPT doesn't tell you this. It doesn't say, hey, by the way, I forgot what we talked about earlier on. It will just try to fill in the gaps and that's when we start to get hallucinations. So CLAUDE has a much larger context window. We can throw a lot more information into it before we hit those problems. And that's really important when we're building assistance, because the more information about the task we can give it, the more information about how it should be performing, the tone of voice, the brand guidelines, for example, the better the output is going to be. So the context window, while quite a boring technical reason, is also very important in this particular case.
- [11:26]
Michael Stelzner
Yeah, and let's talk about what it cannot do, which I'll just go ahead and throw out there. It cannot search the web and this is important. Why don't you talk about that a little bit?
- [11:34]
Kyle Vollmer
Yeah. This is a massive limitation, especially compared to custom GPT. So ChatGPT can access the Internet, whereas Claude will not. It doesn't search the web for you. So when you are providing information to your assistant, you need to take this into account. If you want an assistant which is specifically built to search the web and find information for you, Claude projects is not going to do that for you. It's not suited for this particular job. So it's really important that we think about the scope and what we are actually going to be using our assistant for. There are reasons why this is actually put an optimistic spin on this, but there are reasons why this is actually a benefit when it comes to creating an assistant, though, because when you give the an assistant the ability to access all information in the world, it's going to pull in irrelevant information, because what it is drawing in from the Internet is not necessarily going to be what's useful for the project. Whereas if you have a closed capsule of all your information that you have provided, you can tailor and make sure that the quality of the information you're uploading is high. So depending on the task, this can actually be a benefit. But obviously if you need your assistant to connect to the Internet to pull information from other sources, Claude projects isn't going to do the task for you. And it's a really important limitation to know up front.
- [12:50]
Michael Stelzner
Yeah, and we'll talk about this a little bit later. But the good news is you can copy and paste big chunks of information and you can take screenshots, PDFs. We'll get into all that stuff a little bit later. All right, so let's talk about the kind of assistance we could actually create with something like Claude.
- [13:07]
Kyle Vollmer
Sure. This is always a difficult question to answer. Whenever I'm asked by a business, what can I do? Generally, the answer is like, pretty much anything is, it's a very wide use technology. It's a bit like asking, what can we use the Internet for 20 years ago? It's like, well, yeah, kind of a lot.
- [13:23]
Michael Stelzner
Yeah. Well, let's explore some creative ideas so people can at least wrap their heads around it.
- [13:28]
Kyle Vollmer
Let's go for that. So I've actually come up with a brief framework. I'm calling it the Spun Framework. So it's scope, priming, uploading and narrowing. Scope is so important up front. So the S for scope. Scope is about working out what you want the assistant to do. Like, we already know it cannot connect to the Internet, so we're going to remove that from the type of tasks that we would be building an assistant for. Instead, we might want to do something like have it help us write a newsletter. So I've started to actually use Claude projects to write my newsletter. Whereas I've spent the last year, despite being an AI guy, despite being extremely pro AI, I have avoided using AI because the quality wasn't high enough. Now with CLAUDE projects, I'm able to give it enough information, I'm able to prime it in such a way that it's actually really good for becoming a co writer for my newsletter.
- [14:18]
Michael Stelzner
We also use Claude to do the Social Media examiner newsletter. And we train, basically we train it based on Content that we provide it, for example, the news section. We'll give it some information and we'll have it write summaries of news to be in our news section of our newsletter. Also, we use it to write articles from all these podcast interviews. So this actual interview from this podcast, when you go to social mediaexaminer.com 24. I mean a 24, you will actually see that we used Claude. We trained up Claude on our best content and trained it up on how to take the transcripts from this and write the actual articles, but keep going.
- [15:00]
Kyle Vollmer
So here's a question for you. Why do you not use ChatGPT for that? Is it the quality or some other reason?
- [15:07]
Michael Stelzner
Because we've been able to train Claude on our writing style and we feel like Claude is a superior writer. As you mentioned earlier, it's always been a better writer. And I think, Kyle, it's because the data set that it's trained on is, you know, the Washington Post, all the Amazon data. Let's not forget Amazon is one of the major investors inside of Anthropic. I feel like they've got a better data set to train on English language writing.
- [15:30]
Kyle Vollmer
And I 100% agree. It's hard to objectively say it's a better writer, but everyone I talk to, especially writers, as soon as they use Claude, it clicks for them. They're like, oh, wow, yeah, no, this is much better. And especially when you upload as you have done your own content, it's able to mimic and replicate the style much better than ChatGPT ever would. ChatGPT has a bizarre robotic ChatGPT style and it's very recognizable, especially if you use AI a lot, you know, immediately, oh, that's ChatGPT. Whereas Claude sounds human.
- [16:01]
Michael Stelzner
Now, I will tell you a little side note that I will sometimes take Claude content and have CHAT GPT enhance it. And actually, Chat GPT is excellent at editing something that's pretty close, but not 100% there. And lately I've been taking the Claude content and asking ChatGPT to kind of just simply the word enhance it, and it just makes it slightly better. And then I take it back into Claude and I say, Claude. What do you see is different about this?
- [16:28]
Kyle Vollmer
That's fascinating. Have you tried to use Claude to edit Claude?
- [16:32]
Michael Stelzner
I have not yet tried that because.
- [16:33]
Kyle Vollmer
The problem there is that it might see its own results is already very good. So going to a different AI allows you to get that differentiation.
- [16:40]
Michael Stelzner
I've done a lot of back and forth between Claude and ChatGPT and a lot of the work that I've done and I find that they kind of enhance each other. It's like having two different sets of eyes on the content. They find different things. It's fascinating.
- [16:51]
Kyle Vollmer
Okay, cool.
- [16:52]
Michael Stelzner
Okay, so other stuff you can do with AI that we haven't mentioned. There's a million other things.
- [16:56]
Kyle Vollmer
Sure. With assistant. So social media, post creator, promotional email drafter, legal document reviewer, proofreader, ghostwriter. Basically, I tend to, when I'm working with clients, suggest go through your business processes first. Find the tasks that are repetitive, laborious, or take up a lot of time. And these are the things that you should be focusing on automating using AI. And this is where you should build an AI assistant. You start from the business task and then you combine it with AI rather than just try to shove AI in wherever. Although a lot of companies are still trying to do that. They just want AI and they're not necessarily sure where they want that. But I would just say start with what is already quite repetitive and has taken up a lot of time and see if you can build an assistant that will help you there.
- [17:45]
Michael Stelzner
Yeah, I'll just add a couple of quick things in here. As a marketer, also, if you can train it up in your products. For example, I've trained up a project on social media marketing world. I've given it all the, the sales page data, screenshots of the sales page, emails that we've written in the past. And I created a project so it kind of understood what social media marketing world was. I gave it a lot of background on like who the primary customer is and all that kind of stuff. And I know I'm getting ahead. So anyone who's listening, if you are in marketing and you have a product that you're marketing, which you almost always are, you could quite literally have a cloud project trained up to be kind of your second set of eyes, your creative advisor on anything you create related to marketing related to that product. So let's just maybe make that an easy way forward for any marketer listening right now. So the first part of your process is deciding on a role. Right. Does this mean when we're setting up a project that we should set up a project to be an expert on our product and have no role, or would be better to have a specific role? What's your thoughts on that?
- [18:50]
Kyle Vollmer
I think very, very important is one of the biggest mistakes I see when people try to build AI assistants, whether it's custom GPTs or Claude or doesn't really matter, is that they will try to build an assistant that does absolutely everything. It's really important to build an assistant that does one task really, really well because this is going to be apparent later. The data that we upload to it and the instructions that we give it, the more precise and specific they are, the better the quality of the output. And it's very tempting to try and create a marketing assistant that will do all of our marketing tasks or a product assistant that will do everything to do with product. Whereas we really want to keep them specific so they get higher quality outputs.
- [19:29]
Michael Stelzner
So that's really important Then one distinction from what I just said is, hey, if you do like five different tasks related to your product, you might want to actually have the same data set in five different cloud projects. That's really what I'm hearing you say, right? Instead of having one project do all those tasks so that you can get that role super specific. Like, like Facebook ads expert, email marketer. Right?
- [19:51]
Kyle Vollmer
Absolutely, yeah. They should be individual projects either with similar information about your company, about your brand, about your tone of voice. There's going to be a base layer that's going to be similar for each of your projects, but then with additional information specific to that project. Specificity is always key here. And trying to do too much, trying to have a jack of all trades assistant will lead to disappointing results.
- [20:12]
Michael Stelzner
So what else do we need to do when we're priming?
- [20:14]
Kyle Vollmer
Sure. So for priming, I tend to use what I call my risen framework. So it's R I, S E N. So first is the role that is defining the specific job of the AI assistant. This is just saying act as. And we're giving it a certain specification of what type of job it should.
- [20:32]
Michael Stelzner
Like an email marketer, for example.
- [20:34]
Kyle Vollmer
Absolutely, yeah. Because the thing about AI is it can act as anything. So we need to limit that scope. We need to tell it, okay, I know you can do everything, but I today I want you to do this, you're going to act in this particular role and that helps to focus. After that is instructions. So providing detailed guidance on how to approach the task. I like to think of this as if you had hired a new employee, someone onto your staff or maybe a freelancer. You're not just going to give them a task and say off you go, like work it out. You're going to give them detailed instructions, you're going to give them a step by step guide so that they can go and do that work. It's like giving a standard operating procedure to a new, higher, we need to give the same level of detail to an AI, otherwise we can't expect quality results. And I think this happens a lot with AI where people will give it an extremely generic task like write the company report for 2024 and then be disappointed when they get a bad output. But in reality that's because you didn't give all the information that you would have given to a human. The AI is not going to do any better than the human. So instructions is the next step, just giving detailed guidance on how to approach the task. The next part is steps, so that is specifically breaking down the instructions into clear, sequential steps. So literally A, B, C, D, E, F. These are the steps I want you to walk through in order to complete this task. And then I like to finish off with the end goal, E. This is clarifying and telling the AI what success looks like. We're actually saying this is what a successful output looks like. And potentially even this is what a non successful output looks like. We can give it examples as well, saying this is good version of a good high quality social media email. Yeah, an email. Whereas this is a poor one and it's going to be able to learn from that.
- [22:20]
Michael Stelzner
And then isn't there a last part too narrowing?
- [22:23]
Kyle Vollmer
There is a last part in the Risen framework which is narrowing. However, when we are building an assistant in particular, we can actually do this afterwards. So there's going to be a process of feedback which we can talk about in a moment, where once we have built the basic assistant, once we have got some outputs from it, we're going to have a conversation with it again. Like if we were talking to an employee, they've gone away, they've produced our piece of work, they brought it back to us. We're going to sit there and correct it. We're going to tell them what's good about it and what's not good about it. That's the narrowing process and it's about giving feedback to an assistant so that it can learn exactly what it is you want.
- [23:02]
Michael Stelzner
Many of the top experts you've heard on this show will be speaking at Social Media Marketing World 2025 and with your AI ticket, you can attend at a very economical price. You'll discover practical AI workflows and advanced AI automations that will increase your productivity and save you time. Imagine getting live and in person training from Matt Wolf, Chris Penn, Jeff J. Hunter, Rachel Woods, Molly Mahoney, Brian Piper, Jeff C and many others. Isn't it time to enhance your career by fully embracing AI? Grab your tickets now at socialmediaexaminer.com Aicon we're going to get into the uploading of all the data and stuff here. But how long of an instruction set are we given when we talk about the role, the instructions, the steps and the end goal? We talking about a paragraph, multiple paragraphs.
- [23:56]
Kyle Vollmer
So it would depend on the complexity of the tasks. However, if we come back to the point we just made, we want our task to be a single focused task. So if you find that you were giving it 20 pages of instructions, chances are that's multiple tasks that could be broken down over multiple assistants. So if you do find yourself in that position where you are uploading like a whole standard operating procedure or a whole manual for, you know, customer service or customer relationships, chances are you're throwing too much at it and it's going to give you a poor output. So really we're talking about about less than a page worth instructions. That should be sufficient to provide the priming for the assistant, at least to get started.
- [24:35]
Michael Stelzner
Okay, so a couple quick ground rules before we get into some of the specifics related to Claude here. Number one is you have to have a paid account with Claude and that's currently, I believe, $20 for a professional account per month. That's the only way you can use projects. And it can get a little confusing when you first set up a project and I found this out the hard way. So let's talk a little bit about from your perspective. What do we need to do when we set up the project so that we're setting ourselves up for a good output?
- [25:03]
Kyle Vollmer
So the first step is going to be working out the scope. Then we're going to put together our priming instructions. We're then going to upload content into Claude. Now the, I believe the difficulty you're referring to here is there's a complexity in the UI where it's not quite clear where you should be uploading your documents. I think CLAUDE could certainly improve on this. When you create a project, there is a space for the files that are uploaded to the project. So these are your, you know, your blog articles, your brand guidelines, your tone of voice documents. They can sit at the project level. However, CLAUDE also allows you to upload those any files at the chat level, but it doesn't necessarily translate the information between the project level and the individual chat level. It can draw information from the project level down to the individual chats, but it can't take information from the chats up to the projects. It's a bit of a limitation. It Takes some getting used to and I think they could do a lot better on their design to make sure that's clearer.
- [26:03]
Michael Stelzner
Yeah. And just for anybody who has a cloud account or chatgpt account and you're used to just operating in a single thread in Claude, there's this thing called project knowledge which is off to the right and that's where you can set up custom instructions and add content. So presumably these custom instructions are, are what we were just talking about in the prior role. Right. Like this is your acronym that we were just referring to. Right. What's your role instruction steps and end goal that would be set presumably in the custom instructions, not in the actual chats. If you put them in the chats, you're going to learn the hard way that when you create a new chat inside the project, that chat doesn't know what you talked about in the other chat. It only knows what's at the project level. And that's kind of an important discovery that I made just before getting on this, this, this interview today and that's really important for everybody to understand. So let's talk about how we can train the model, like what kind of information we've briefly gone after a little bit of it I've mentioned briefly, but you know, how much information do we really need to put into the project?
- [27:04]
Kyle Vollmer
Generally the more, the better, the more information we give it, the more it can work with. That said, it needs to be very relevant and high quality data. Just throwing information at the model for the sake of it doesn't really help. We want to make sure that anything that we uploading is going to be relevant to the task that it's doing. So there is a very understandable impulse to want to give it absolutely everything about your business every single time, regardless of the task. That may not help though, because we want to keep it focused. So really this comes back to the question of what is it we want our assistant to do? And then again thinking if we were talking to a human being and if we were providing them with instructions, if we were providing them with the relevant kind of knowledge base, the relevant data, what would we give to them? So we wouldn't necessarily give them the annual reports, if we're asking them to do social media posts, for example, it's just not relevant to the task at hand.
- [27:57]
Michael Stelzner
You talked when we were prepping about pre processing information and do you have any wisdom or insights as to kind of how we should discern what to put in there and what kinds of things you normally put in there when you're, you're creating a project.
- [28:08]
Kyle Vollmer
Sure. So the main difficulty is going to be running up against the upload limits. So one very useful thing is there is a progress bar which shows you how much of the knowledge base has been filled up. So every time you upload it saying, okay, 10%'s gone, 20%'s gone. That's quite useful. If you are uploading, let's say, information from a website, you'd want to go ahead and strip out all the HTML and CSS and any kind of metadata that's connected to the actual information. You want to provide the model. Just because we need to keep that memory clear and anything that's not relevant, we just strip out. So that kind of pre processing, just to make sure that the data only contains what's useful for the model, I think is going to help because it allows you to fit more relevant data within the memory.
- [28:50]
Michael Stelzner
Yeah. I also ran into some limitations with file sizes.
- [28:53]
Kyle Vollmer
10 megabytes.
- [28:54]
Michael Stelzner
10 megabytes per file. Yeah. Which can be really easy to hit if it's a massive screenshot, you know what I mean? So you might have to take those screenshots and compress them or save them as JPEGs, or copy and paste text instead of a screenshot. So you can put a lot of data in there. Right. It's just the file size could only be 10 megabytes at a time. So you can put multiples up there or is that not true?
- [29:16]
Kyle Vollmer
Correct. So 10 megabytes at a time and you can upload five files at a time, but you can then upload another five. Another five. So it's all going into the same memory pool. So you can get a significant amount of information in there over the last year. For my newsletter, so when I'm using my newsletter Assistant, I have 300,000 words worth of information that I'm able to draw on from there. And that's why I'm able to get very high quality results because I have so much data that I can work with.
- [29:45]
Michael Stelzner
Okay, so let's review kind of where we're at right now, if you don't mind. We're starting to create a cloud project and we have identified, we've gone in, we got a cloud account, we created a custom project, we've identified a unique role. For example, we're just going to stick with this email marketing role where you're going to help create emails. And we've identified the kinds of steps, right. The instruction steps and end goal. Now, if we don't know what kind of the instructions and steps should be. Do you recommend having Claude and or ChatGPT assist you in identifying the steps? If the steps aren't obvious to you, you've just done it by routine, instinctively.
- [30:25]
Kyle Vollmer
Absolutely. So I would recommend for this going outside of your project so that we don't confuse the project by getting too meta with it. We want to ask how we can best give information to a project, so that's better to do within a separate chat, just to separate out so it doesn't get confused about what we're talking about. But absolutely. So I find now, whenever I have a question about pretty much anything, any process, I no longer go to Google, I will now just ask an AI how it is. What are the steps steps to get this done. So in this case this is exactly right. You can ask it how best to condense image files, for example, how best to chop up a database so you can upload it into 10 megabyte chunks and it will give you instructions about the best ways to do that.
- [31:11]
Michael Stelzner
Yeah. And the truth of the matter is, from my experience, that whether you ask it in a project or out of a project, it's still clawed, so it's still smart. It's going to do its best to answer whatever questions. Like, for example, if you want to send 10 emails over 10 days and you wanted to help you decide what the topics ought to be, it doesn't really matter whether you do that inside the project or outside the project, but I would imagine if you did it inside the project, it'd be better because it would know at least what the product is and maybe some of the common. It could read the data that you've given it and try to discern what some of the common objections might be to your product, where it would have to guess on it if it was outside the project. Right?
- [31:51]
Kyle Vollmer
That's true, yes. No, absolutely. In that case, it can help you to create the priming instructions based on the information you've given it already. So, yeah, you can say my end goal is to write a series of 10 welcome emails for my marketing campaign for my product. You have the information about the product, give me some potential frameworks I can use for these 10 emails, or give me five different variations of 10 email campaigns and then you work from there.
- [32:16]
Michael Stelzner
Yeah. And I've done that exact thing where I said, here's the start date, here's the end date, we want to do multiple emails on the end date date. And I just said, what do you recommend? And it came up with a really good plan. And then I started working each one of those emails kind of one at a time. And I find it really interesting. So this is where we built the project. We've trained it on the basic data. Now we're about to like start using it. And I guess this is where we get into this narrowing concept. Right? So talk to me a little bit about how once you're starting a project, how we can kind of refine and iterate to make it better for what we want.
- [32:54]
Kyle Vollmer
So it's already got the priming instructions, it's already got the uploaded data. So the first thing we do is we just ask for an output. So if we are doing an email campaign, for example, we'd go ahead and ask for the first email in that campaign. And then we are going to, as a human, we're going to review it and then we are going to provide feedback on its output. And this is how we start to train the model to get better at knowing what exactly we want. Because we're going to go back and forward. We're going to say that first paragraph is a bit lengthy or you need to punch up the ending. And we're going to start giving that feedback in human language to the model and it's going to be able to use that to redraft. We're going to give more into it, more feedback, third draft, and then we're going to continue to iterate again and again. So sometimes my newsletter, I will go back and forth about 20 to 30 times, each time giving feedback as if the AI assistant is my ghostwriter.
- [33:47]
Michael Stelzner
Talk to me about how you do that, because there's a million ways you could do it. But I'm curious what your process is. Right? Let's say you're having it write the first email and there's something about it that's off. What do you tell it specifically to do?
- [33:59]
Kyle Vollmer
So I always recommend you just use natural human language. The AIs are smart enough to understand that. So even if you don't necessarily know what's off with the email, you can say, I don't like that there's something wrong with the first paragraph. And you can give it, you know, hints like, oh, it's, it's too long, it's too short, or it's too salesy, or it's not salesy enough. And it's going to be able to create new drafts based on that. So in the same way you would talk to a person who is drafting up your emails, you can give similar levels of feedback to the AI. And it's going to be able to work out what it is you want. It's much better, though, because it's more patient. So if you ask for 20 redrafts, it's going to keep redrafting for you.
- [34:36]
Michael Stelzner
It's never going to get upset. Yeah. And give an example. First output I got out of an email in a project was very formal. It didn't match my tone of voice. And I said, make it more casual in the voice of Michael Stelzner, assuming it knew who I was. And it got really casual. So then I had to say, okay, find a happy medium between the two. And when it found the happy medium between the two, that third iteration was much closer to kind of what I was looking for. So I think this is really important for everybody who feels like, you know what it's like if you've ever outsourced to a writer, right, or you've been a writer yourself, as I've been. If you don't like it, you might be like, oh, this thing sucks. Right? That might be your natural disposition, but you gotta keep working it and you gotta keep working it. And then I've been able to say, okay, whenever you sign these emails, sign it this way. Michael Stelzner, comma, founder of Social Media Examiner. All on the same line, right? And I just find that I have to keep. The more I tell it, like, don't use these kind of words. Use an emoji in the subject line. It's just one of those iterative processes that goes and goes and goes. But I've even found, and I'm curious if you've done this, maybe there's a sentence or a paragraph that seemed completely off the mark. I've asked it to rewrite just that one sentence. I mean, I don't know how far down the rabbit hole you go with this editing when you're doing this.
- [35:58]
Kyle Vollmer
So I will, as I said, for a newsletter, I'll go through about 20 or 30 different rounds. So some will be small things like you missed off the introduction, some will be larger, like, I want you to reshuffle around these sections. So going back and forth, what I would then do when I am happy with the draft, I will copy and paste it out into my email software. I will manually make changes, I will rewrite it. I will get it ready for publishing to my standards, that will go ahead and publish that. That's fine. I will copy the final version back into my Claude assistant and say, hey, this is what I landed on. So this is including my human Edits. Learn from this and it will go through and compare to the last version with the final published version and it will summarize, okay, these are the changes you've made. I will include these in the future. And this is why using the AI assistant is really powerful, because it allows us to have these running conversations. So when I go back for my next newsletter, it's going to have remembered all of that information. Whereas if we are doing this in individual chats, like on ChatGPT or even Claude, we're going to start losing that information between the different tasks. But within a project, we get a much larger context length, we get a much larger upload pool as well. So that allows us to have this continuing conversation and to make it a true AI assistant.
- [37:23]
Michael Stelzner
Well, and it's my understanding also that with Claude, I think they have a teams. I'm almost certain they have a teams version where you could potentially share your projects with other people inside the company. I feel like I've heard that before. I don't know if that's true. Have you heard that to be true?
- [37:41]
Kyle Vollmer
Yeah. So it's not a different level. As soon as you have the premium membership, you can create projects and projects can be shared within your team, you can publish them.
- [37:49]
Michael Stelzner
Oh, okay, so wait, is this with any paid membership then? Yeah, you mean with the professional. Oh, interesting. Because I don't have it. I'm the only one in my company. Well, I have other employees that have their own cloud accounts, but I haven't upgraded because I felt like it was a little more expensive to have the teams account. Just a little bit more expensive. And for whatever reason, I'm being cheap and I'm not doing that. But what I'm hearing you say is you could share projects with other people.
- [38:11]
Kyle Vollmer
The projects can be shared. Yeah. So they're really useful within the team. So once you've created your assistant, you can share that.
- [38:16]
Michael Stelzner
Huh. Very fascinating. Is there anything that you've discovered, have you also used ChatGPT custom GPTs?
- [38:23]
Kyle Vollmer
Yes, but coming back to the technical reason, the context length, it doesn't allow you to upload enough information for it to necessarily capture your tone of voice. So I've never had success with them. I've always been disappointed. Also the fact that, as we discussed, the written output from Claude is just so much higher. So because personally, most of my tasks are due with the written word, so newsletter, social media, I just find Claude so much more superior. And that's why, despite having written my newsletter for over a year, I've never really used AI in the workflow up until now. Up until. Until the release of CLAUDE projects. And that's why I wanted to talk about it today, because it's. It's such a powerful tool that I don't think enough people are using.
- [39:06]
Michael Stelzner
Have you discovered anything that CLAUDE projects can do outside of what we've already discussed? Like, is there any cool little. I'm just curious if there's any little tips or anything that you've found that's unique to projects that. That's maybe not just inside of claude, generally speaking. And if you haven't, that's totally cool. I just.
- [39:23]
Kyle Vollmer
Just.
- [39:24]
Michael Stelzner
I kind of feel like it's this. I mean, first of all, the fact that it can be private or public is kind of intriguing. I did not know that. Are there any other things that you've discovered inside of cloud projects that we've not yet discussed?
- [39:35]
Kyle Vollmer
No, it really is just the quality that is astounding. And it's so hard to communicate to people until they try it. They actually go and try claude. I'm not affiliated in any way, but I do bully people into trying claude. So I think it's not a. Necessarily a hack or a tip or anything like that. Just. Just the underlying quality of the model is just so high that I think it's very important that people try it out.
- [39:58]
Michael Stelzner
There is something that I just discovered, quite literally during this interview, which I did not know was there. There's a capture screenshot button inside of Claude and I'm in Chrome, and it allows you to pick any particular tab that you have open. Now, that's kind of cool. Were you aware of that?
- [40:13]
Kyle Vollmer
Yeah, yeah. So you. Could you try that on your images that you're trying to upload? Although I know that you're trying to capture a whole page sometimes, which I don't think it's going to be able to do. I know it can't do that. It can undo the straight screenshot, but.
- [40:26]
Michael Stelzner
The fact that it is even there, that allows you to capture a screenshot is kind of maybe a little hack, because I didn't even realize that there was that little camera thing there. I've not yet tried using the mobile app. Have you found any advantage? Are you using the mobile app for Claude at all, or no?
- [40:42]
Kyle Vollmer
Yeah, it's a very slick app. I actually do most of my newsletter writing on the mobile app, which sounds absolutely wild, but it's such a nice interface that it's entirely possible to work with projects on a telephone. So I often write my newsletter on the bus, which not many people can say, I'm sure.
- [40:59]
Michael Stelzner
Wow. How are you doing that on your phone? I mean, like, there's not a lot. I mean, are you speaking to it? It doesn't have a verbal interface like ChatGPT, does it? Or does it?
- [41:06]
Kyle Vollmer
No, it doesn't. No. So I'm primarily doing the rounds of revision. So because it takes so many revisions, I might over a few days be going back and forth, back and forward over a particular newsletter issue. So just having it there on my phone and being able to pick up a chat that I've been having on the computer anywhere and to kind of go through its latest draft and say, maybe you can change this, maybe you can change that. Being able to do that on the fly, on the move is really powerful.
- [41:31]
Michael Stelzner
Are you doing anything inside of a Google Doc and then bringing it into Claude, or are you doing the whole darn editing of your newsletter completely inside of Claude?
- [41:39]
Kyle Vollmer
So I do the final edit in my newsletter tool in Beehive, but up until then, everything is in Claude.
- [41:44]
Michael Stelzner
Wow, Very fascinating, Kyle. Thank you first of all, for answering all of my questions and hopefully opening people's eyes to what can be done with Claude projects. If people want to discover more about you, where do you want to send them?
- [41:57]
Kyle Vollmer
Sure. So my social media, I'm primarily active on Twitter and TikTok and I'm IAM Kyle Ballmer on both of those platforms.
- [42:05]
Michael Stelzner
Is it I am Kyle Ballmer or just Kyle Ballmer on the.
- [42:09]
Kyle Vollmer
I am sorry.
- [42:10]
Michael Stelzner
Okay. Yeah. Cause I wanted to clarify because you said I am sorry. So it's a.m. kyle Bulmer.
- [42:16]
Kyle Vollmer
That's it.
- [42:16]
Michael Stelzner
Okay. All right.
- [42:17]
Kyle Vollmer
That's a good clarification I can tell you. Working social media.
- [42:21]
Michael Stelzner
And what were you going to say about your website?
- [42:23]
Kyle Vollmer
Sure, the website. So because I got so fascinated in this topic after we first decided to do it, I've actually written up a whole guide on how to go through this process. So I know we've covered it in kind of broad strokes today, but if anyone is interested, I've put together a written guide. It's about 4,5000 words about how to build an AI assistant. So I'm going to make that available on my website. So that is aibusinessbreakthrough.com SME for social media examiner.
- [42:52]
Michael Stelzner
Awesome. Kyle, thank you so much for coming on the show and sharing your insights and wisdom with us today.
- [42:57]
Kyle Vollmer
Thank you. And thank you for having me.
- [42:59]
Michael Stelzner
Hey, if you missed anything, we took all the notes for you over at socialmediaexaminer. Dot com. Be sure to follow this show on your favorite podcast app. And if you like what you heard, what you let your friends know about this show, you can tag me. I'm Telzner on Facebook, Stelzner on LinkedIn, and iKestelsner on X. 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. I'll 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. Don't forget to get your AI Ticket to Social Media Marketing World 2025. Become an AI Enhanced Marketer. Grab your tickets now at social mediaexaminer.com a icon.