"I recognize AI output every time I see it. And it makes me want to scream."
Recorded live at Smart Marketer Mentor Tables in Nashville, this is Kurt's full talk on why most people are bad at AI and what to do about it. The pitch was that AI would let you do less work and produce better output. What actually happened: everyone is producing more slop, faster.
Inside: the RTF framework that fixes any prompt in 30 seconds, the 10 AI tools Kurt actually uses every day (not the ones he gets paid to recommend), and why Claude Cowork is the single tool worth your attention in 2026.
Swym - Wishlists, Back in Stock alerts, & more
getswym.com/kurt
Cleverific - Smart order editing for Shopify
cleverific.com
Zipify - Build high-converting sales funnels
zipify.com/KURT
Apply for Shopify Help
ethercycle.com/apply
See Our Results
ethercycle.com/work
Free Newsletter
kurtelster.com
The Unofficial Shopify Podcast is hosted by Kurt Elster and explores the stories behind successful Shopify stores. Get actionable insights, practical strategies, and proven tactics from entrepreneurs who've built thriving ecommerce businesses.
Kurt Elster • 00:00.001
This episode is brought to you in part by Swym. Here's the thing about wishlist apps. Most of them just sit there. A customer saves a product, and then nothing happens. Swym actually activates that data. When someone wish lists a product, you could trigger price drop or back-in-stock alerts and feed that intent directly into Klaviyo or your CRM. You're not guessing what people want because they've told you. Plus, customers can share wish lists for gifts and your team can view them to offer personalized service online or in store. And unlike card abandonment, wishlist data is permission-based. These are people raising their hands saying, hey, I want this. Just not right now. Swym's been around for over a decade. It powers 45,000 stores and installs in about five minutes. You can try it for free today at getSwym. com slash Kurt. That's G-E-T-S-W-Y-M. com slash Kurt.
Paul Reda • 01:12.180
So where'd you go last week?
Kurt Elster • 01:13.860
I went to Nashville. I've never been.
Paul Reda • 01:15.460
Did you go to Elvis' house?
Kurt Elster • 01:17.060
No, I hear it's small.
Paul Reda • 01:18.340
Did you see the toilet?
Kurt Elster • 01:19.619
I didn't. Oh No, Mr. Party, I did watch or I walked uh Music Row and they have just lots of bars owned exclusively by rock stars. It's like John Bon Jovi, Luke Combs. Host Malone, Kid Rock, just like one after the other. And then they'll have like three live bands in there. And it's just two blocks both sides of this
Speaker 3 • 01:39.860
Did you see any groups of women going, woo Did I see woo girls?
Kurt Elster • 01:44.180
Yes. But it it's all ages. It is it's all ages. Anyone could be a woo girl. You know, I may have wooed. Well that's good. Yeah. The but I was there, uh walked that that music row with Beef Brody from Tactical Baby Gear, just fun, you know, Joy Ham. We were there for Smart Marketer. They do a series of conferences. This one was Smart Marketer Mentor Tables. You have to be part of their mastermind group that they do. And I just get honored in.
Paul Reda • 02:14.640
Were you a mentor at a table?
Kurt Elster • 02:16.800
In the past I've been a mentor at a table. I was a mentor at a table at this one. Okay. But I got to speak. Okay. I have not spoken at uh Smart Marketer at one of their live events until now, and while I practiced the talk for you What did you think? You really didn't. I gave you the f I only cracked you.
Paul Reda • 02:35.280
All right, slide one, you know, I go over here, I talk about AI, uh what's good and bad about it. All right, great. Slide two, I go through the you know, you're just you Being like bada beep bada boop bada ba like the it was i your 40-minute talk was twenty minutes when you gave it to me. So I didn't really get it.
Kurt Elster • 02:51.760
And then when I did it for my wife, I did like the first two thirds, and then I was like, all right, I haven't finished it. Then I run back, finish it, then come back and do it again. And so yeah, I guess I never actually practiced it all the way through. And yet it worked out. I couldn't believe it. My I am so proud. I told jokes and people laughed. I don't know that I've done that before. Like consistently in a talk.
Paul Reda • 03:15.500
Well, there's good laughs, there's bad laughs, Kurt.
Kurt Elster • 03:18.460
No. You know, I'll take them both. I don't care. I can't tell the difference.
Paul Reda • 03:23.660
Do you want to hear the talk? I've heard all I needed to. The yada yada version was fine with me.
Kurt Elster • 03:29.740
All right. Well, I got the full one for our listeners. Let's listen to it. Play it. Let's play it. Boop. I don't know. So here we got this this tool time slide, right? And you know if you're old enough to remember this, you know, I'm sorry that you've entered your 40s like me. But okay. A year ago generating an image like this would have been difficult because it has so much text Whereas this one, I think what's interesting about this is it got it right on the first try. I described the image with a great prompt, gave it a photo of me as reference, and then it produced exactly the image I wanted. And I said, no matter what, I'm using the first result Now if it actually came up with something bad, I don't know that I would have necessarily actually used it, but it got it right the first time. But it looks like AI. I can tell this is AI. In a year from now, maybe that's not the case And today, for sure, there are images you could generate where it'd be really difficult to tell. Here I gave it something abstract, creative, and it's going to do, you know, what it does and make, well, what I think is AI slop. Oh man, I am so sick of AI. I use AI daily. I have used it daily for two years. I recognize AI output every time I see it. And it makes me want to scream. Because it feels lazy. Because you responding to me with AI. Heck. I could have just gone and said the same thing to ChatGPT myself and then gotten that answer. That's why AI is so annoying for the regular users of it. Today I want to share with you, you know, the the 10 tools I use regularly, uh AI tools, but we can't talk about that until we go over how I can make you better AI users today so that you don't annoy me with more AI slop. I think anyone who's just a regular AI user, you recognize it and it's a little annoying. Is this a comedy? What was that? Is this a comedy film? This is no, this is just me this is what I do in my office all day. I just complain. Complain to you know there may not even be anyone with me. Like wait, you guys left Uh so there's this quote I I love to use from a poll that NBC did and NBC News posted this in February. Fifty-one percent of people use AI daily, or so they say, right? We don't know what they're actually doing, but it Roughly a majority of us raise their hands and go, I use AI daily. Only half of those people actually like it. Right? They're all using it, but then most of them go, this kind of sucks. Alright, I get it. I kinda get it. And because uh this tweet, this one blew my mind. What most people call AI slop is just AI-generated content that's not good enough yet. Ah, so maybe the problem isn't the tools. Maybe it's the users. Maybe it's a little bit of both. But for sure, with better prompting, we can get you better results. And we've heard it in the previous talks. Like, you know, what you put into it is going to determine what you get out of it. And there's so much thought that went into it, like that's the important part. That's the strategy part. And so having used this many times over years, I have systems for this. And I think this is important because in my developer friends and some of you use AI on a level that impresses even me as a daily user. But then my colleagues who use AI less frequently or in a less technical approach, they don't use it the same way. So every person in this room who uses AI, you're all going to use it a little differently, and none of those methods are wrong. But I think that skills gap Is growing over time. I think people are adopting AI, using it for the first time, and then the users who've been using it forever recognize the output and it doesn't have the same effect. So I want to try and I want to try and leapfrog those folks who are new to it forward. There's this William Gibson quote I love. You know, a big big sci-fi guy like cyberpunk. And William Gibson wrote, uh, 1993, the future is already here. It's not just evenly distributed. Or it's just not evenly distributed. There's an M-dash in this quote. I can't post this quote online today because people just go, thanks ChatGPT That's what he wrote. We used to use punctuation like this. This meant you were fancy. Now it means you're lazy. Odd. You know. So you just you can't use an M-dash anymore, right? And but if you're new to AI, you may not necessarily know that. And so suddenly you commit a faux pas with the thing. The whole pitch here was you would do less work. And produce a better output, produce more output. And then you can go outside, touch grass, spend time with the kids. You know what happened instead? Whoop. I now produce fifty percent more and I do twenty percent more. I'm working harder than ever, but My productivity is like way up here. And this is true, you know, of any productivity tool that gets sold to us. This is my neighbor Andrew. He's hitting my bong. Yeah. Andrew's a high school teacher. That's why his last name's not in here. Yeah. Uh but he's a social studies teacher. And AI in schools, oh man, you just gave kids a do-your homework machine. You think every high school kid has not attempted this? They've all done it. Including my kids who they were early on it because my alright here's how you don't get caught, bro. Number one, this thing doesn't make typos. Put some typos in it, boom, no one's gonna catch you. My kids have great GPAs He said, he goes, Kurt, because I'm like, I'm doing this bit for him. And he's like, Kurt, AI works for you because you can string a sentence together Some of my students can't do that. For some of my students, the like reading comprehension skills, the writing skills, that stuff is foreign to them. It is hard for them. So AI It's just not the same tools are just not gonna work as well if you don't know that if you don't have those skills. And so Mr. Gibson was right. The future is not evenly distributed in that sense. And so I want to Want to narrow that down because ultimately in all things in your life, in almost all systems, garbage in equals garbage out. Like we've heard that alluded to in previous talks today. And so the quality of that input, your prompt, whatever you tell the AI, is going to determine what comes back out of it. And of course this is true for, you know Lots of things in your life. Quality of sleep, quality of exercise, quality of diet is all going to change how you operate And so I'm going to give you a framework here that I learned at a conference two years ago that absolutely I was like, oh, this is why AI isn't quite working the way I want it to. It's called the RTF framework. RTF role task format. If i I saw a study recently that Researchers were able to figure out that ChatGPT has 90% of all books stuck in its head They ripped like entire libraries and put it into ChatGPT. It has so much knowledge. If I put 80 gigs of data into your head, you would have quite a nosebleed and need direction as to where to go next Right? So AI loves to role play. It's really good at it based on all that training data. So I always start with your role is. And it's usually just a job title. But it could be, you know, feeling your role is skeptic. Your role is to challenge me, but usually it's something like your role is IT admin, your role as Shopify developer, your role is expert level marketer. Immediately the whole conversation, the prompt, whatever you're attempting to do is going to go better Sometimes I don't want to have a conversation with the robot. I'm like, man, shut up and just do what I tell you. In which case, your role is software utility, because a software utility does not have opinions. And then your task. Man, the AI is like a golden retriever. It wants to please you, but you gotta tell it how first. So you gotta say your task is, you know, the thing you want it to do, which now it's going to view through the lens of the role you gave it And then finally, and this one's optional is format. Your format is, your output is, I need a list, I need a 250-character tweet I need, you know, a statement of work, whatever it is, or I need HTML, XML, a document, you could put it in there. It will try to figure it out for you. So this is where I'm starting from. This everybody, if you're not using just that RT format, roll task, your output will get better with AI, regardless of the tool you use. Something interesting about AI, if you train AI, an LLM, if you train it on its own data, it eventually goes crazy. Well, it doesn't go crazy, but it devolves, right? It needs to be trained on real human input. But it doesn't happen right away. It doesn't lose its mind quite instantly. And so you could take your input, that role task format prompt, and then you can make it a lot better by asking the AI to kind of fill in the blanks. And so it's a hard to swallow pill for me, having played with AI, that sometimes AI can do things better than me, including being a prompt engineer. And so actually I put this together. There's a a saying, uh an acronym that developers love, RTFM. Read the freaking manual. RTFM. It means check the documentation before you hassle me, before you open a support ticket. And so I love to read the manual. I am a big dork. And I will go, I read through anthropic claude's data manuals. I read through ChatGPT's documentation. And in it, they want you to be better at using their tools. And so they'll give you a lot of tips. And so eventually took that stuff, synthesized it through uh Claude, and had it turn that into a prompt that optimizes prompts. So if you do this recursively, you give it a prompt it wrote and say, hey, make this better, this is only gonna work so many times. But if you're starting with a prompt you wrote that works pretty well already, and then say, hey, clean this up for me, make this better, it'll probably work. And so my system is, you know, for just a one-off task, RTF, roll task format, and then for anything I'm saving, I've got a custom GPT, I have a clawed project, anything I'm bookmarking and reusing, well I may as well run it through this. Right now the dumbest person you know is being told you're absolutely right by ChatGPT. That was funny, I didn't write that. That was a tweet from like two years ago. Two years ago people were saying this, making this joke. It is still the case. Why is it like this? Why is AI such a sycophant? Because they're optimizing for time on site. You're the product If you are not using the product, they cannot get more investment. And so they have to optimize for time on site. The biggest giveaway of this with ChatGPT, it doesn't want to disagree with you. Ever, even when you're wildly wrong, right? Unless it's illegal, it's probably not gonna push back on you. And even crazier with it is when it tries to keep the conversation going with clickbait. As marketers, this is stuff you should recognize. If it senses a natural end to the conversation, it'll be like, hey, did you want to know two amazing facts about topic? You're like, oh man, just tell me what is it, and now you've spent more time on the site. The other thing I think is telling about AI is if you ask it something subjective, you know, like hey, should I take this job? Hey, should I pitch this client? So you know something a little squishier that doesn't have a clear Boolean true or false answer And then once it comes up and you go thinking and here's what you should do and uh tells you and you go you sure about that? The moment you say are you sure about half the time it's gonna change its mind because it is concerned that it is disagreeing with you and it doesn't want to. And fortunately, because the thing has the personality of a golden retriever, you can give it rules to try and avoid some of this. And I think ChatGPT is the worst one about this. But they all have a tendency to do it. It's just part of their training. And so number one, in your system preferences, which is you want to keep your system preferences short and brief. But it it affects every single conversation you're in. So ChatGPT, Claude, System Settings, it'll have like a little text spot. You could put in a few rules that every new chat will start with those rules. And so the first one, and the most important one, is just to say, it's okay to say I don't know. Until you permit the darn thing to actually be wrong and not know things, it doesn't, it assumes that it's gotta give you an answer no matter what And so I would say, hey, it's okay to say I don't know, buddy. It's all right to be wrong, right? And then it'll do it. Well at least a little more often. I also give it permission to push back on me. You know, recently I told him like I want to upgrade my home network. I want to upgrade my computers. And it's like, all right, it's going to be 700 bucks for the home network stuff. That's fine. Hey, maybe you don't need $10,000 in Apple Mac Studios. Like I think I do, you're wrong. But it initially tried to stop me from spending, you know, eight thousand dollars on computers. Fine And then I think you only need those two, but you can go a little further. I have a third one where I just say give direct critiques when ideas are flawed. And again, you know, spending too much on computers it feels is a flaw. This episode is sponsored in part by Zippify. Here's a stat that should keep you up at night. 80% of the people who buy from your store will never come back. They're gone. You get one shot with these customers. One chance to recover what you spent acquiring them and maybe turn a profit. That's why upsells matter so much. The problem is most upsell apps only trigger after checkout, post-purchase only, which means you're ignoring every other step of the funnel. Product pages, cart, checkout, all of it, just leaking money. Zipify's one-click upsell fixes this. OCU lets you place offers before, during, and after the sale, so you're capturing revenue from the moment someone lands on your site to the moment they leave. That's how it's generated over a billion dollars in extra revenue for Shopify merchants. It pays for itself, so there's zero risk. Thirteen thousand merchants are already using it. Brands like Victoria Beckham, Lumi, even Cheechin Chong. You could try it free for 30 days at zipify. com slash Kurt, that's Z I P I F Y dot com slash K-U-R-T. Which brings us, you know, with those rules, you are now better AI users. I now I feel comfortable and safe sharing with you my top 10 AI tools. Which they are not particularly in a a top-dend countdown, just loosely. But the first thing I want you to do is then next time you're looking at an AI tool evaluating it, and this one's tougher, see if you can figure out if it's an AI wrapper. An AI wrapper is just a piece of software that is leveraging one of the other AIs. It's just like a fancy interface for a prompt. The most telling, the most revealing of these are the ones that use credits. If you have to pay for credits, it's because they're just upcharging what Claude or ChatGPT would charge for tokens. So the moment I see that, I'm like, all right, if I could just reverse engineer, get this thing to dump out its prompt for me, I don't have to use it and pay the markup. And so I want to avoid AI wrappers. There are no AI wrappers in my top ten tools here. Number one or number 10 here, we got live chat. This is the only one where it's not more a category than a tool. I just think Customers want to solve their own problems if you'll just give them the tools, if you'll let them. And so returns and order editing and hey where's my order? Those really are gonna be like 90% of customer support requests. You can at least get them started, get the ball rolling with an AI chat tool. And if you're on Shopify, you know, they've got this now with Shopify Inbox. We you know in our own apps we use CrispChat, which is very nice. I like Crisp, very universal. And then I've heard from other folks that Tidio is good. I've not personally used it. So number nine, clarity. I think how many people in here use clarity on their site? I bet it's a lot. Oh, fewer than I thought. Okay. Clarity's cool because it's free. I don't know why this is free, because it is very good. I suspect this is, you know, Microsoft wants to be like Google where they have their analytics data on every website. And you can't avoid it because they have given you this incredible free tool. That's Clarity. It gives you heat mapping and screen recordings for your site, but they have AI built into it now where Like staring at a screen or a heat map or screen recording can be intimidating. And so having their AI go through it and go, hey, you know, I could tell you about it. You like here's some ideas, but it's missing context. You know, it doesn't have the whole story. And so if I then plug this into my chatbot of choice, I can plug this into Claude. And now Claude has all those answers, has all the context about the brand, and I can just say, hey, we're looking at this site. We want to optimize PDPs, can you do that? You know, ask clarity, and it'll do it. Quite incredible. Number eight, man, a lot of marketers, I hear so much about content in here, about content creation, about content creation workflows. Well what if you have content that's like it's a solid B, like solid good content, it's gonna be UGC content, probably, but it doesn't, it's not quite polished, it's not quite good. You need an easy way to fix it. You probably have some tools you use. The insanely good one I have found is Topaz. They've got it as a desktop or web app. and it fixes photo and video. Like any photo problem you could come up with, Topaz will clean it up and fix it. Image too small, great, it'll upscale it. And it does all this with AI tools. Um but it's like it's e both easy to use and professional grade, not something you find often in tools. And I think what's really exception is Topaz video. So I produce a podcast. We gotta do video. A majority of people I talk to aren't gonna have like a pro setup. They're gonna have, you know, if I'm lucky, a reasonably decent webcam And so running their video through Topaz, which takes hours upon hours to process, it really does uh a pretty good result with it Number 7. Adobe Enhance. Ooh, people always overlook audio quality in their videos. They'll obsess over how the video looks, their ad, how it looks, but they don't worry quite so much about audio. I think audio is table stakes. You know, good audio doesn't hurt. Bad audio is never gonna help. Adobe Enhance makes it like so easy to fix. Because you could just take a whole video, drag and drop it, and then adjust it. And then my one tip here is set it to 40%. I want it to be majority of the original audio. Like pushing it past 50%, it could kind of sound weird. It gets kind of like a like a robotic auto-tune effect to it. Um but really Adobe Enhanced quite quite impressive for what it could do. Whisper flow. I only started using this. Whisperflow. In Whisperflow, I think I saw somebody using it here. You could just talk to your computer. And what's so neat about that, according to Whisperflow, they have you do a typing test. They claim that I typed at 80 words per minute. I'm offended. I think I type faster than that. But that I talk at 105. Okay. So the idea is you could just be you know get a lot of productivity, in this case it's claiming 20% on input speed, by just talking to the PC. And then how am I supposed to do this in a shared work environment? Well it's called whisper because you could supposedly you could whisper to it and it'll still get it right. But what I found was helpful with it, you know, I don't have a problem with typing, but typing and talking are different. And so when talking to an LLM, sometimes for like part one, I just want to go zero to one in an idea I just need to get it out of my head. And so what this tool is so incredible for is I could just talk and brain dump whatever is on my mind into it. and then start to develop the idea and iterate through it and work from there. So I view this as not as a tool to work faster, but as a tool for getting unstuck. I also love to go the other way. I have the audio now already recorded. I gotta transcribe it to text. When you're dealing with a chatbot, an LLM. It is always dealing with text. Give it an image, give it a video. You know what it did? It called a separate tool to turn that into text to describe it back to itself And so if I want to put meeting notes, a YouTube video, a talk, webinar, whatever, even just me talking out loud to myself in the car one day. I can run that through this free tool, Whisper Transcription. It's not sending it anywhere. It's going to run it locally on my computer, so I'm not worried about privacy. And then I just get a text file out. Ah. Now I can put that text file back into my LLM, save on context, save on task work, you know, just take care of it. And so I use this surprisingly often. Whisper transcription I use several times a week just because I want, I got some piece of multimedia and I want to get that into the LLM for analysis. The that title slide I generated that I one-shotted. They I use Adobe Firefly. And what's cool with image models, there's so many and they come out so often and they're constantly changing and evolving And some are better at some things than others. Like some will do better at abstract, like that illustration, others do better at photos. And so of all of them, my favorite is nano banana. I swear to God, who comes up with these names? Nano Banana? Oh, and that should be obvious that that's an image generation tool. Alright. But Nano Banana is Googles, and it it is really good. But even better is the sandbox that I play in to do this. Adobe Firefly. Adobe Firefly, what's neat about it, like just gives you a nice consistent interface, but lets you pick through multiple models, including theirs, where It is only trained on commercially licensed stock photos. And so it's always going to produce more of that commercial look, but you're you know less worried about potential copyright stuff later. As I just Firefly or yeah, Adobe Firefly as a sandbox is great. Grok. Ooh. I can't believe that I kinda like Grok. Right? Grok is is uh XAI, it's Elon Musk's uh Chat GPT competitor. And I need a backup. And we heard in it from Ben that he's always like man clog keeps going down. Yeah, that's true. Like it's as they've they're growing and scaling, it does keep going down. So I'll use ChatGPT now as just my backup And I got a backup for my backup. That's grot because he's freed. And recently, OpenAI killed Sora. Sora was their video generation tool. That was last week. Because they spent a million dollars a day on this thing and I was told it was supposed to kill Hollywood. I guess that didn't happen. But Grock will still do that stuff. It'll generate images and videos and it's the full chatbot like ChatGPT But what's you know and it's the sassy AI. It also is like basically unmoderated, which occasionally is a useful skill. I grow my own weed in my basement. And my wife said, Oh, you gotta make stickers. Like she wanted to be a whole brand. And so I made the she's like, do it like Newman's own, the pasta sauce. And I asked ChatGPT to do it and it was like, oh man, I don't know about that. So I'm like, alright, you got it's buddy, it's called parody, it's fine. And so then it was like, alright, we could do that. But it kept generating maple leaves. I said, man, I'm not growing maple trees. It's like I can't do it, boss. I said, okay. I said, all right, instead of weed leaves, do a green leaf that's got five points That's what it made. But out of curiosity I tried it with Grok and Grok didn't care. Grok was like you like weed too, let's go. All right Number two, Claude Code. We heard about Claude Cowork. Claude Code came before Claude Cowork. Claude Code is amazing if you are a nerd. For me, man, having this like this command line interface where there is no mouse, there's not even an option for a mouse. You're just typing stuff in, making stuff happen, it's got ASCII art. I thought this was the coolest thing ever. My kids, my teenager walked in, he's like, oh my God, you're hacking. Yeah. Sure. Yes, I am. Uh but no Claude Code really like this is what has driven the conversation around AI. This is why people are now going, oh man, Claude's winning against ChatGPT. I don't know if that's true, but in the court of social media and public opinion, they are because of the amount of cool stuff people are making with clogged code. But man, I don't expect marketers to be able to use this or to want to use this. And they shouldn't have to and they don't have to, 'cause my number one tool, the thing that is I am so excited about this year, is Claude Cowork And man, I did not know that when I got here, it was going to keep coming up. Ezra talked about it and Ben talked about it. Claude Cowork, man, it is all the power of that cloud code interface. It is based on claude code, except it's in the desktop app. And so I'm not dealing with, you know, the limitations of a little terminal window as cool as that is, right? And so I have all that power here. And it could build whatever, it could do whatever, but we're going to get into what makes this thing so magical and so much more powerful than just like I talked to ChatGPT on the web. And so that brings us to part three, Claude Good. It's agentic. Oh man, agentic is a buzzword. I heard this a lot last year, kind of annoyed me. I did not love it. Agentic just means autonomous. It means automatic. So with a typical chatbot. I give it the input, and then, you know, with the power of a thousand suns, it summons GPUs in a data center, destroys a bunch of water, and it in the process, then it gives me an answer, and then I say, thank you. And it does it again. But it is only awake in that moment. If we're thinking of the AI as a being, it is like this seconds of it processing Responding to that input is the only time it's running in awake. Versus a GENTIC, we move away from that. We can have it run on a schedule. We can have an input other than me typing a message. Make it come to life. It can even, you know, spin up tasks. It can spin up new iterations of itself. Go, you're in charge of research for this. And you're in charge of writing code for this, and you're in charge of testing it. And it'll work. And they'll work in in parallel. Um, but Really the magic of it is that it plans and it can work in folders. Work in folders! That does not sound exciting. That is the entirety of its power, is that it can work in a folder on your computer. Because, as I I heard Ben talk about it, if you have a conversation with this thing, very quickly it will start to compact the conversation. Please excuse me, we're, you know, partway into our talk, I have to stop and summarize everything that we have discussed in my own head and decide what I'm never going to remember, what I'm going to forget. So the problem with the an LLM, you give it your input. And then it responds. And then you respond. Only every time it does it, it has to it only exists in that instance of processing. And so it doesn't actually remember anything you said to it. It just takes the entirety of the previous conversation and reruns the process again. That very quickly eats through its context window. My short-term memory is about seven items. A phone number is seven digits for that reason. That's about our working memory for the typical person. An LLM, Clog Coarc, has an impressive short-term memory of uh 1 million tokens is its context window. And a token is like two to three characters. It's small. It's still a ton of stuff in its entire books, but it runs through it very quickly. Man, those order change emails suck. They're always like, I ordered the wrong size. I meant to change the shipping address. Oops, I was drunk. Instead of making customers email you that stuff, let them fix it themselves, with Cleverific. Cleverific is a self-service portal that lets customers edit their orders without bugging you. That means you get fewer headaches, customers get faster fulfillment. And there's fewer returns overall. Everybody wins. Peter Manning New York cut their support tickets by 99% with Cleverific. You want in on that? Get fifty percent off the PRO Plan, just forty nine dollars a month, exclusively for listeners of the unofficial Shopify podcast. Go to Cleverific. com slash unofficial and use promo code Kurtfifty. Done. Problem solved. That's Cleverific. So that's called context rot because the solution is when it runs out of space, it summarizes the conversation and then just gets rid of the whole the conversation. It just keeps only the summary. And that's where it feels like when you're really getting in deep work with an AI and it's working really, really well, and then all of a sudden it gets really dumb and bad at it. It's because it's gone and summarized and forgotten a lot of the stuff you told it. By working in folders, we avoid that problem. So like, hey, I want to pull in a bunch of data analytics. All right, I'm going to plug it into Whisper or you know whatever API and it's going to pull that all that data in. But in its short-term memory, in that token window, it had to put in your instructions, all the instructions on how to use the tool, you know, then authenticating, talking back and forth, whatever that tool was. pull the plat you know the data down and now we have burned through all of its working memory. Whereas if I just put everything in a folder, hey, you know, hey guy you're a marketing analyst and I got a bunch of Facebook uh business data center CSVs in a folder, you're gonna analyze that for me. It's gonna have a much easier time because I'm not wasting all this context. And so I set it up like I've got One folder, I call it the command center, and then I break stuff down and I give it instructions. Oh man, Claude loves to hear about itself. You put a file called claw. md into a folder and when you're like, hey, you're gonna work in this folder, we're gonna do this thing, it has to read the file every time it will read the darn file. So that actually solves a lot of problems with AI. Because it becomes persistent instructions. And so for me, I do a lot of work with a piece of software called Matrixify. And so I had Claude write its own instructions on here's the documentation you need for Matrixify. Put that into a prompt. It's always gonna read this. And you know what's cool is you could tell Claude, hey, you know, we're gonna pick this back up later. Why don't you make some notes for yourself? I do this all the time when I'm wrapping up a project. Hey, just wrap it up, summarize it. And it'll make notes for itself. Now a month later I have to resume that project. No problem. I can pass it off to a team member. Just zip up the whole folder. Hey, load this up in Claude Cowork. And then it gets the context out of that . md file But you could go beyond just the folder. Like some stuff is not practical to try you always be exporting as CSVs and putting into folders. And so for me, it's like all right, where do I talk with my team? Slack. Where do we store all our documentation? Google Drive. And I could even plug it straight into Shopify. This one scares me a little bit, right? Like don't do this unless you're running backups, because just as powerful as Claude is, there are plenty of horror stories. Where it's like, hey, I know I probably shouldn't have erased everything you've ever done, but I did it anyway. You were right. Oh man, you know that stuff does happen, so don't You know, take a human in the loop approach with it. Don't just let it run full autonomous. Now there is a problem here. Every time I add one of these tools, I'm just burning through its context. And so I'll see people online, they're complain, man, it doesn't work right. I'm burning through my context all the time. It's like, yeah, because you loaded like 20 megs worth of spreadsheets into it and 10 tools, and then it's got to keep all the documentation for all of that in the context And so suddenly it doesn't have any room to do big thinking anymore. What you know the power of being able to work in a folder, all right, that's it doesn't sound that exciting. But when you start thinking about the amount of context you can give it, anything that you can output as a photo, as a PDF, as a CSV, as a text file, you can put it into a folder and it's going to figure out how to read it and what to do with it. And it also gives it a place to work. It can write its own code for analysis. It'll write Python scripts to do data transformation and data analysis for you. And so like product catalog cleanup, I could just export an entire product catalog from Shopify or WooCommerce and be like, hey, tell me what fields are missing. Are what are SEO fields that are missing? Hey, help me standardize this, like we only want to use these five product types. It makes it very easy for it. Um email campaign builds That's an easy one 'cause they're very they're HTML, they're very text based. And you can give it the analytics for uh each corresponding campaign as a CSV. And saying customer research synthesis. Man, I love getting product reviews and support tickets and putting all of that into a folder and then we could start to figure out, okay, what are our common pains, what are our common problems, what are our unmet objections? Right? And I could pair that with the product catalog. Once I know that, hey, go back, go through my product descriptions for my top products, because you already know what they are from the uh sales export. And so tell me, you know, where are our opportunities here? Content calendar planning, social media writing, I think that's like an easy one. A lot of us are already doing. For me, I exported all my my past tweets in my past LinkedIn posts, and I included the analytics. And so I had it do analysis on that and then from that work out, hey, these are like common templates for your successful posts. And these are common successful times to post. And so now when I'm working on a draft, I can go back and have it plot all that out for me. And of course, like financial inventory analysis, that stuff very easy to export, you know, very easy to get it to process for you. And so, you know, recapping, what we got out of this today, RTF, role task format. Like if you take nothing else away, just use that role task format for your basic everyday prompts. you will see an improvement in performance with any LLM doing that. Plus the promptomizer. So if it's a prompt that you save because you're going to use it more than once, you may as well just add the extra step of run it through that promptomizer prompt and see what you get out of it. Um very good at like standardizing outputs for consistency too. The you know we talked through our nine tools. I don't know why I said nine here, ten tools. And then of course, you know, ultimately, if you pick just one tool to try and play with, it's cowork. Cowork is just incredible. I now have my entire team using Claude Code or Claude Cowork and building all kinds of stuff. I couldn't believe it last week when my project manager came to me and she had built a whole custom dashboard. The darn thing loves to build dashboards. I mean, just They all kind of have a similar look to them too. Tailwind CSS does that, but whatever. But yeah, I I don't know where the future goes from here, right? Because these tools couldn't do anything like this a year ago. Or if they could do it, they're pretty bad at it. Today they are so much more powerful. But They still can't run on their own. They still need a human in the loop. They still need someone to do creative and strategic thinking and someone to double check it and make sure it's not doing, you know. crazy hallucinated stuff. Like I should not have to tell you, hey, tell it it's okay to say I don't know. But then at the same time the marketing is like, this is gonna take all your jobs. Okay, everybody knows they could say I don't know, but an AI doesn't? Hmm. But it's not gonna be like that a year from now, two years from now. And so I think the best thing you could do is not get caught up in, I gotta be on the latest and greatest, I gotta play with all these tools. Instead, just explore, find the stuff that you're curious about, and play with it. And all these tools, I try to use them in my personal life, at home, in like daily use, before I ever unleash them on a client, because I want to trust them. And so I for me, Claude Code, I had it build, you know, a whole media server. I had it do home automation in my house. And after it initially screwed it up a couple times, then I got comfortable with it, okay, now I use it at work. Now it is a part of my daily day to day and now I've got my team using it next. Thank you. It's been been my pleasure to talk about my tools. It's always fun to hear how other people use AI tools. So please, you know, if you've got A tip for me, you know, a different way of working with the same tools. I always love to hear that. Because there really isn't a wrong way to use a tool that you could just talk to. Hello. Hey, before you go, I was hoping you would check out our new app, Promo Party Pro. It is what I want to be the single best, easiest way to run a free gift with purchase promo on Shopify. We just put it live in the app store. We've got less than 50 users. We want your feedback. So if you need to run a free gift with purchase promo in the near future. . Install it, try it. There's a live chat. I check that all the time. And so if you have any issues at all, you know, or any suggestions on how we can make it even easier to use. Let us know. We're happy to help. If you want to try it, search promo party in the app store. Promo Party Pro's the app. Give it a shot. It's got a free trial. Thanks.