How to Avoid the AI Wrapper Trap

0:00What makes you more than a AI rapper? I think one criteria is you have some unique workflow or proven approach that incorporates the LLM. Welcome to the Talking AI Podcast, where we talk AI with both experts in the field and early adopters. I'm your host, Matt Page, and we're here to demystify AI for you so you can get some value from it. Let's talk some AI. Today, we're joined by Tim Hansen, co-founder of PenFriends. And for context, Pinfriend is an AI-native product that generates high-quality SEO articles without having to actually prompt. Plus, Tim's an all-around awesome guy. Has a great British accent, so I know the audience is going to enjoy this one.

0:42Welcome to the show, Tim. Thank you for having me, Matt. That's the one thing I found about living in, now living in Vancouver, is that I'm 100 % more charming because I'm British. Oh, yeah. You've got to use that to your advantage, definitely. Yeah. And recently you relocated over to Canada, which is awesome. I got to hear about some of that later. But so I'm sure the audience is thinking, oh, I can generate content with ChatGPT or Claude or name your LLM. And that's kind of exactly the topic I want to get into today. And that is how do you actually create value with an AI native solution when you're competing against like this insane pace of change we're seeing with LLMs?

1:26and avoid getting that like stamp of being an AI rapper. And I don't mean like, you know, the AI generated version of Snoop Dogg. I mean, like basically having a thin layer of functionality on top of an LLM. So for listeners, if you're not familiar with that term, we'll get into more of that later. Plus, I just want to hear about some of your journey of building TenFriend, which for listeners, it's an awesome solution. I use it, I can vouch for it. But I want to start, what was that moment, that initial opportunity or gap you saw in the market that pushed you to create PenFriend in the first place?

2:05So it came from a combination of everybody getting told that you don't have a job because AI is going to come and just take over marketing and marketers don't have a job anymore. and kind of realizing for myself in the way that I've had to do many many times before because I've worked in SEO is that my job gets effectively rewritten every six months thanks Google the job really really enjoy literally relearning my job every six months and essentially it's the same thing it's like every time that that's happened I get on to whatever educational sources I can learn as much as I can as quickly as possible start testing things and see what works what doesn't work with my own my own sources my own clients projects those kind of things and honestly this was no different this was um just hey I was coming learn as much as you can get things in place and thankfully from my agency days and running my own um client work I had a lot of SOPs in place of how I was doing things anyway so very standard this is how it gets things done it became very easy for me to use AI to replicate things that I was doing already purely because I had done those things and written those processes out and knew exactly what the steps were and knew exactly what the output was supposed to look like it was there was very rarely ever a case with me where I would put a prompt in and get a bad answer and go, oh, the AI is the problem.

3:44It was always like, no, my prompt's just bad. Like I'm not defining the job properly, like well enough. And I think that then very quickly turned into, yes, you can use LLMs to get content, but you don't have a structure on how you want to get content out of those LLMs. You don't know what you want. you haven't made that content or that outline hundreds of thousands or thousands of times and when you put something in you it's very easy for people to do the same thing of like oh it's really bad it can't write posts it can't do this it can't do that and you're entirely right it can't because you're asking it in most cases to do 20 30 40 I mean we'll get into it, but like potentially a hundred jobs in one prompt.

4:35Yeah. And that's where it fails because you're asking it to do way, way, way too much. And you're expecting that from it. You're giving it nothing. Yeah. It's the quintessential, it isn't you, it's me, right? But it is you. It's the person prompting. And I heard, I forget who I was listening to, but they were talking about this in context for coding, like I could ask, you know, again, name your LLM, doesn't really matter, to write code for XYZ. Then I could write, say, another prompt and say, write quality code that's secure in XYZ, and it's going to write better code. It's not just because it's lazy.

5:20It's just you got to give it that extra bit of context, or you could say, write it like some amazing developer, and it's going to be even better. And I think this whole concept of prompting as we know it today will evolve, but it's a critical point there that you kind of hit on. There's another piece, too, you were hitting on. It's this idea of effectively zero shot versus multi shot is another kind of concept. Like, are you just asking it to do something with no examples, one example, two example for additional context? And then how are you breaking it up? Am I asking it to go, you know, do everything in one prompt or breaking it out?

6:04And obviously, you know, best practices to do kind of what you were hitting on, break it up, right? Yeah. Quick break in the pod. If you're listening to this podcast, chances are you've been thinking about how to actually use AI inside your business. And that's exactly why we built the AI Opportunity Finder. It's a free tool that helps you uncover high impact, tailored AI use cases based on your business, your goals, your pain points, and your industry. No fluff, no generic use cases, just real ideas that fit your business and the ranked by ROI potential. It takes about three minutes to run and it's like having your own personal AI strategist for free.

6:40If you want to try it for free, check out the link in the show notes or go to hatchworks.com backslash AI dash opportunity dash finder. I mean, for Pen Friend, initially went like the first version that we released back in november last year my co-founder and i we sat down and we broke we broke down like how many decisions do you as a human need to make to go from i want to write a blog to i have a published blog yeah how many decisions as a human do i have to make something to do that process properly and we worked out that it was a minimum of 28 decisions to do the whole thing. And for each one of those decisions, you're wearing a different hat.

7:27You're taking different variables into consideration. You have a slightly different goal for you to pass on to the next, like the next version of you that makes another decision. Some of those decisions, you might go and make five in one direction for one of these decisions in another direction to be made. You need that information. You need research. You need to go and read a load of other blogs or have a look at titles and see patterns before you can make this decision over here. And yet you have people out there that are trying to write an entire blog in one prompt. It's like, we reckon you need at least 28.

8:07So the first version of Pen Friend that came out was about 25 to 30 different prompts for it to write it. Now, the version of Pen Friend that we have at the minute, whenever it writes any blog, it's about 80. it's almost four times as many prompts in some some of the different ways that we do it and it's purely because we realize that there's even more nuance to a lot of those decisions it's a lot more nuance and there's a lot more like we weren't placing anywhere near enough importance on some of the tiny tiny little decisions that actually turn out to be really big things that we need to know but also we've got like five different more sources of information that we use now to write blogs versus the way that we were doing it previously like now we're pulling information from like people also ask and doing fact checking through perplexity and like doing a lot of the like numbers side of things uh analytic side of things are all being done through like chat GPT, OpenAI, and then the writing side is through Anthropic.

9:18So it's not really to your point of like, is it just a wrapper? No, we're effectively using each of the individual LLMs for their strengths. But the way that we have it as well as we have an agent that chooses what is the right LLM for the specific job for the specific blog that it's writing. And all of those are built off of templates. And those templates are built off of Like we build prompts as it needs them and then asks the right LLM that prompt based on what is the blog that you'll currently write. All right. There's so many angles I want to go down. You're starting to get into that point of what makes you more than an AI rapper.

10:01I got to go back there real quick. Was there actually a moment in time that you remember that kind of triggered this whole idea of, was it using ChatGPT for the first time and you were just like, oh, shit? And was it an oh, shit moment or was it like, oh, this is really cool. I could use this in some unique way. Was it the negative scary or was it the opportunistic? Honestly, it was a little bit of probably unfounded bravado. In that I had a client and they were paying a lot of money for a lot of content to be written. And I genuinely was like, I reckon I can write thoughts that make you better content.

10:46And it was just, I have to prove that I'm better than this. I can do this. And this is the direction that things are going. And maybe at the time it was entirely unfounded. I don't know, but it was like, I don't see anybody doing this properly. And I think I can. I'm pretty sure I can do this. Yeah. And then realizing like just how hard of a process that was, but really enjoying the process. And it was very evident of like, this is the way that things are going to go. I have to learn how to do this just to keep my job. for me to still be useful to clients for me to still um be able to provide the the service and the advising and the consulting that i love doing i have to learn this yeah i have to know that like this is the way to because i'm going to get questions about of course i'm going to get questions about it but i have to i have to know yeah no yeah no that's it's interesting so ai rapper let's let's get into this uh you know there's a lot of people out there saying if you're an ai rapper you know good luck you're dead the next version of chat gbt or whatever is gonna kill your product right uh and i remember it was i think sam altman was saying if the future stuff we have coming excites you you're good to go if it scares the shit out of you you're in trouble So I'd be curious to get your take there.

12:19But there's one piece you got in the PenFriend website, and I think it gets back to what you've been hitting on. You have a proven content strategy. It's delivered millions in revenue to your clients. You spent 1 ,800-plus hours refining prompts and guiding you through the article creation step-by-step. And I was thinking about this before the call. I'd be curious your take. If we can break this down for the audience of how they should start to think about this at the atomic level, what makes you more than an AI rapper? I think one criteria is you have some unique workflow or proven approach that incorporates the LLM.

13:01I think that's a critical piece that can be differentiating. are you leveraging proprietary data in a unique way? I think that can be differentiating. And then are you serving a very specific niche, creating this like super tailored experience for their job to be done, how they operate? Like those are the three I was thinking of. And I think the thing, the biggest point against the whole AI wrapper, you're screwed, is humans are very lazy. Yes, I can go to chat QBT and do all kinds of things. But guess what? I can go to Excel and do all kinds of things too, but I don't use Excel to run my CRM.

13:38I use HubSpot or Salesforce or whatever. But I had the three that came to mind. Thoughts on those? Are there others that you think differentiate you, kind of build a mode, make you more than just a wrapper? I mean, the process of how we build the blogs is very much, it's not a linear process. So if you were to put a keyword into PenFriend, the first thing that we do with that keyword is we take it to Google and we see who is ranking for that search term. So immediately, that's not something any of the LLMs currently do. Yeah. You can't currently see. So we go and have a look, see what ranks. read all of those articles and try and work out what are the patterns, what are the things that Google are rewarding for why those things rank.

14:30And then the other thing that we do with that keyword is we take it through an entire flow of who is the person behind this? What are they looking for? What are they expecting to find? What actually answers the reason why they searched in the first place? because I think the best way this was ever explained to me was somebody searching for an ergonomic chair isn't looking for an ergonomic chair. They're looking for a solution to back pain or potential back pain or because they're building a new desk set up and they just have those things in mind. So you can potentially put all of the chairs you want in front of them.

15:11But if you don't address the core reason why they're searching for in the first place, they'll never continue on. It's just a really difficult thing for them to do. So we try and work out as much of that as we possibly can so that before we even put a title in front of you, we've tried to work out what is the title that makes the most sense to the reason that this person is doing this search, so the human side of things. But then also, what are the patterns that Google is rewarding? And what is the gap that nobody else seems to be talking about? and then we craft titles from there. So you put your keyword in, we've done so much work before you ever get titles.

15:52And all of that comes from, we've got our own proprietary system and how we like figure those things out. But also it's just a lot of human understanding. There's a lot of prompts in the backend of just, okay, what are the reasons? Do they want a product? Do they want information? Do they want guides? So they're looking for comparisons, like all of these different types of things. You get the title. That's one click. You like the title? Cool. Now we'll show you what the outline is. So that outline, the way that we write that outline as well, is whenever we create the title, we also try and work out.

16:30So if this person is looking for a guide, I don't want to accidentally start trying to write a blog for you that is A product versus B product. I want to make sure that the titles that I put in front of you actually create content that that person who's searching is looking for. So if they're looking for a guide, I want to make sure that the titles are conducive to how-tos or step-by-steps or things like that. So you get to see the entire outline. You can edit the outline if you want. I would prefer people edit the outline so they can make it more their own article at this this point we found that like 10-15 minutes of editing at this point can equate to up to two to three hours of editing on the other side if you push it in the direction that you actually want to push it in in the outline stage right yeah if you just go and like add just the bits that you want to it it helps helps a ton And then during the writing stage, what we then go and do is for each section, we'll then go and pull facts and quotes and statistics from different libraries that exist.

17:46And they're live libraries as well. They're like proven places that exist. They're archival places on the internet that we have API access to that we pull everything into. And then for that individual section, this is the bit that I really like, actually. We put a lot of work into this earlier this year. So the entire blog is made up of different blocks. And each of those blocks have got different prompts that write those blocks. But those blocks, those prompts change depending on where the block is in the blog. So if you have like the topic overview, that could, depending on the right blog, be at the beginning of a blog or be at the end of a blog.

18:33Now, I don't potentially want to write the same thing if it's at the beginning versus if it's at the end. So we have a thing that knows where the block is in the entire blog and knows what came before it and after it, and we'll write the section that fits depending on where it fits in the entire blog. So it's kind of dynamic in nature there. It's this crazy comparison of human old way of doing it, Chi-GPT or LLM way of doing it, and then pin front everything you've built into it way of doing it. I think I heard you mention this earlier. A big thing we're leaning into at Atrix AI is around the agentic approach to things, and you're leveraging these effectively agents for different tasks within a workflow.

19:21They're very specific and tailored to X thing, and they're working with other agents to execute on this thing. Are you actually taking that approach from an architecture standpoint? To your point, is there an agent specifically tasked with going and checking the competitive side on Google and whatnot and gaps? and then one for the titles and then outlines or their different agents that have kind of their specific role and prompts tied to them that you've architected in a way? Yeah, yeah. So we've got like mini versions of those at the moment. The main one that we're working on like right now is fact checking.

20:05Yeah. because it's we um we did we had in development at one point we were like oh we'll go and write all of these different things and then we put it through a fact checking um agent and it would do this really interesting thing where it would if it didn't know if a fact was true or not it would just leave it and sometimes it would just be completely false yeah and then there would be times when it would go and grab like new fat like good real facts fact facts not fake facts and put them in and still leave the other ones and what we realized was the better way to do it is the other way around is this is what we're going to write about before we write that section we then give that to like the fact agent the fact agent goes and finds facts and stats and quotes and things like that that fulfill like that blurb.

21:06And then we give the blurb and the facts to the bit that then writes. And that gave us a much, much better output because we're now telling the bit that writes, like don't make shit up. Yeah. Figure out the story of this blurb, fit these jigsaw pieces in where it makes sense, and then it's good. and then we've got another one that then goes and checks that section and makes sure it didn't add anything in that it doesn't make sense to. But it's crazy when you change the order of things because it became, I think it almost doubled the amount of prompts per section. So it then doubles the cost of the section and all of this kind of stuff that comes with it.

21:56but it makes something that is ultimately considerably better. Yeah. That quality versus cost kind of trade-off, right? Yeah. Yeah. But then the problem also with that is like it's slower. Yeah. It's like now I need to go and run all of this different information and pull from this and this and this. And I think fact checking of anything of these kind of things is ultimately going to be such a differentiator. I think so many people at the minute don't trust AI content because it's very easy to say something that is ultimately not true, but sounds very true. Very believable, right? It's like you almost want in that user experience standpoint, if I'm sitting there going through the fact checking stage, I almost wanted to take it a second just for my own like peace of mind.

22:52Okay. It's working really hard on this. But the interesting thing with this agentic approach is these agents can be working together in what feels like, you know, near real time to us versus like our human mind thinks, okay, I got to go talk to Tim about this. I got to work with Judy on that. And it's just a very interesting way of architecting a solution. So for any listeners curious about that, Reuven Cohen is a great resource. You can find him on LinkedIn, talks a lot about agentic stuff. We do a lot of talking about it with Hatchworks AI as well. I think that's a big area you kind of need to understand as you're thinking about AI native solutions in the future.

23:38uh yeah have you have you played around much with i think one of the next big stuff for us is all of like the like vector storage and any rag rag things that come next is oh yeah the yeah i want to do is get it so people can give us their content or we can understand it better and like write a new article that can link to those but also just understands it and we can write in that tone, it's very much the thing that makes the most sense for us going forward. Yeah. So, I mean, we could nerd out on this a minute. I need to connect you with our new CTO, Omar, on this. So, we're actually working on a rag as a service solution that just makes it super easy to do exactly what you're talking about.

24:21But that's actually a really cool idea, just thinking in context of your product, because you could connect that to, you know, to your point, all of your past content, anything brand related, tone of voice, like all of these different components that make it uniquely you. And that's the beauty of it is you can connect to those data sources and you don't necessarily need to go like train some new model on it. It's just an LLM is using that as a resource. And to your point, using kind of vector databases and whatnot, making it easy to go through and search that stuff. So yeah, I think that's a really cool kind of product evolution right there.

25:04I mean, it's where my head was kind of at that point, like this time last year. I was like, this is the way I want this to go because it makes so much sense. But it's, I mean, we're a very small team. I think what we've done so far has been incredibly impressive with how little there are of us. it's I think the main thing that I've learned with working with our dev team and like what we've done so far is take so much longer than I think it's going to yeah yeah that's that's that's how it goes when when you're building a product I'm all too familiar uh with that it's funny I was just chatting with uh Matt Lewis doing another talking AI episode and he's the chief AI officer at NECO Medical.

25:55So they're huge in the life sciences space, but they're looking at RAG as a solution as well for that reason. Because, you know, obviously in content, you don't want that to be wrong. When it comes to like medical, you really don't want it to be wrong there. And you mentioned human in the loop earlier. They're thinking very intentionally because you do want a human in the loop for some of those like medical related or life science use cases there as well. I'm curious, Tim, like the SEO front in general, like there's also the angle that SEO could be completely disrupted and creating content in general just gets completely flipped on its head.

26:39Any perspectives there? I know we've started seeing like people finding us via perplexity and things like that, which is kind of interesting. Is there an evolution you think is going to happen just on the base principles of SEO past, like generating the content, but people finding what they're looking for? And maybe you don't need a blog. Then you're like SEO optimizing for, you know, tools like perplexity and things like that, just getting directly to the answer and being a source, right? I mean it's interesting that you say things like that because I we've seen other clients as well come up with like oh my god I would appear in like Microsoft Copilot and and things like this it's like it's it is interesting to see it's really cool to see things like that just starting to appear I think people should always own their own content and and the media in one way or another and I think yeah the way that people find those things will probably change it probably will go from being like specifically in Google to whatever the if you ever thought that in like our lifetime we would see Google actually have a massive like massive competitor and not just one like four or five like all at the same time it's it's kind of nuts but um yeah for a i think the medium of which people get found will change but the medium in which people buy i don't think will change much and there's always going to be a point of convincing the trust yeah um really showing that you know how to solve people's problems.

28:32And just getting people from wherever it is that they're searching for something to actually needing a human or a service or a product to help them and then paying for that. I don't know how much of the bottom part of that funnel is going to change. I think there will always need to be some level of content that it just makes sense for you to own. in that funnel in one way or another. I think the thing that made the biggest impact to me on just how important content is was, it was a workshop that I was on like years ago. And it was a study done by Deloitte, PwC and Google. And if you were to take between zero and 10, zero is never heard of you ever before 10 is like the second before i give you any money or whatever your service is what number do you think is the average point where a person gets in touch with another person on that like purchase journey uh yeah it's it's 8.6 to really far down that entire funnel, which means you have so much space at the beginning of that that you're still selling.

29:59Yeah, you are still selling to someone without you ever knowing that you're selling to them. So you need something in front of those people. You need something that makes sense for those people to make their own way down that funnel. Yeah. And if it's not on a blog, If it's not, maybe it's on YouTube, maybe it's on social, like these kind of places. But if it doesn't exist, then you're just not going to help those people at all. Yeah. Here's one take. I'm just like riffing off the top of my head, like alternate reality. So we talked about agentic and agents and things like that. Say I have an agent in the future and I'm saying, hey, go find me the best product for X or have this problem, do that.

30:44that will the agents still got to go and do some amount of research and figure out the problem space. I wonder if in the future, like we're like SEO optimizing for agents and not humans, which could be interesting. And then are we just writing blogs that are actually agents that can just speak better to other agents? Yeah. Why don't we just connect them and just let them do their thing. LinkedIn just becomes like an agent marketplace. Yeah. Oh God. That's like, when, when agents start having, AI agents start having their own LinkedIn profiles where it's all going down now. I'm moving to the mountains at that point.

31:24I'm done.

31:28A couple of other just questions to wrap it up. I'm curious on the Ben Friend journey. What, like, because a lot of our listeners are probably interested in thinking about how do I starting adopting AI, build it into my products, build completely new and novel products. What was like the biggest challenge you or your team faced in this journey of creating PenFriend? So, I mean, the three of us are not technical. I am the most technical. And that is like, I'm really good at writing prompts. Yeah. like actually putting that into software code like there's these kind of things like really wouldn't know where to start let alone build a product the thing that i i would recommend if you're in the same boat that we are is yeah like our developer matthew is incredible he's always asking questions about content about seo about like the bits that we know and the problems that we have and and how how he can make the product better and i think the main the main bit of advice that i would have is like just find a developer that obsesses about the product because I've worked with many developers before that kind of just are like, the code is really nice and they're kind of obsessed about that side of things, but then just don't give a shit about what the product is and what it does and what it solves and why we're trying to make it and those kinds of things.

33:15I think we got insanely lucky in that we found that person very, very quickly. It's interesting. You're almost defining, I think, what will be the future prototype, prototypical engineer that's successful in the future. Because a lot of the tasks they do today will be made easier with AI. But if you get one that's truly product-oriented, product-minded, and curious, that's what's going to make you stand out as an engineer. So I think that's really interesting. What surprised you the most in the journey? Any just like big surprises, good or bad?

33:56It's a hard one. It's like I've gone from being head of SEO in an agency to being a consultant and an advisor for startups and then deciding, okay, cool, I'm going to run my own. Like it seems like quite a typical like, oh, those things sort of make sense. And it's like, it is just so so much more work like hard work of just no one else is going to do those things for you you you really need to just yeah like it's it's late nights it's working on things constantly it's no breaks i think the not necessarily like the biggest surprise but the biggest lesson i've