347: The AI-Powered Solopreneur
Download MP3Hey, I'm Arvid, and this is the Bootstrap founder. As a software entrepreneur, I'm wearing so many hats from the coding, marketing, customer service, all of these things. And I have found AI systems to be game changing tools in my daily work. So today, I wanna share how I use these systems for my work, particularly the large language models like Chat, GPT, and Claude to hopefully inspire you to benefit from these things yourself. This isn't just about the tools themselves though, but about how they have transformed my workflows and expanded my capabilities as a solo founder.
Arvid:And I think you'll find that very interesting. My approach to work today is quite significantly different from just 2 years ago or a year ago, but let's go back 2 years. Let me paint you a picture of the before and after here. So 2 years ago, my work as a software developer and a SaaS founder was pretty much AI free. Sure, I used tools like Grammarly for writing and maybe some transcription software here or there, but most of my daily grind was just that, a grind.
Arvid:Every line of code, every marketing email, every customer support response, it all came directly from my own fingers and my brain. Hopefully, brain first and fingers after, but it was all internal. So as a self proclaimed one X developer, and let's be honest, probably not the greatest at customer service or marketing either, I was stretched fairly thin. I was doing everything myself from deep diving into database issues to writing product copy, all of these things had to happen. And I often felt like I was barely keeping my head above the water.
Arvid:It was quite intense. It was a lot of work and I always had to really prioritize what I would spend my time on because I had to dive into it a 100%. So fast forward to today, and the landscape of my work has changed dramatically. AI permeates every single aspect of my business, PodScan in particular, but also the roots of founder. Everything I do is very much impacted by AI.
Arvid:And not only is the product itself heavily based on AI for data extraction and analysis, when I talk about PodScan here. But I'm also using AI tools throughout my business operations at almost every level. The key shift has been how I view these AI systems. They're not just tools for me, They are actually partners. That's how I look at every AI thing that I use is as a partner.
Arvid:I don't delegate tasks to them like I would to an employee. Instead, I think about it as a collaboration. I collaborate with them, and I use them as consultants or guides or I even think about them as co founder sometimes. Next, I'll explain why. These tools are here to help me do the work better, not just to do the work for me.
Arvid:So let me break down the main categories of AI tools that have become integral to my workflow here, because there are 3 of them that are quite distinct. And I think we should always look at distinct things distinctly. So though here it goes, the first one is large language models in in a chat context like platforms like chat GPT or Claude AI. These are my go to for brainstorming tasks or writing or problem solving and even some coding stuff and I'll get to that. Because coding is also the 2nd big tool, specialized coding assistance tools is what I would call them.
Arvid:These things are integrated directly into my IDE for coding. I use PHP Storm for that and they have like JetBrains has a JetBrains AI thing that integrates several different LLM systems directly into the, the IDE. I've heard obviously there there are way more cursors out there and GitHub Copilot and whatever they might be called. They all exist. They all are integrated into systems and they're always there ready to help me write better code faster.
Arvid:And the third thing that we're not gonna talk as much about today, but they are also important because I already mentioned them are task specific AI tools. These categories can include things like advanced transcription tools where you throw in an audio file and outcomes text. And that for my podcast processing that I personally do, this is quite relevant, but it's only a one off kind of situation. I'm not gonna talk about this because those things I rarely use, but I wanna talk about the things that I always use. So let's dive into how I actually use these tools in my day to day work.
Arvid:The first use case clearly is brainstorming and ideation, and that's one of the most impactful ways for me to use AI. Let me walk you through a typical scenario here. Say that I'm writing a blog post about a new feature in PodScan. Something for marketing, something to reach out to my current customers and my prospective customers for. So instead of staring at a blank page, here's what I do.
Arvid:I open up a conversation with Claude or Chat GPT, it used to be more Chat GPT in the past now at this very point it's Cloud AI because it just feels like a better tool for conversationalism but every model changes that so I go back and forth and check it out. But right now let's say Claude and I give it a clear instruction. I give it a prompt like give me an outline for a blog post about and then I talk about a specific PodScan feature and then I tell it to focus on points that would resonate with podcast creators looking for this kind of solution. And sometimes I even tell it the AI system what it is. You are an experienced writer in a technical field writing articles for people who love podcasts, and then I give it the rest of the prompt, right?
Arvid:Sometimes it's important also to tell it what it is, but that's the prompt. Just say give me an outline, focus on key points and then the AI will provide a bullet point outline and that gives me a baseline skeleton of what most people want to actually know when they read something about this particular topic. And from there, I start filling in my own thoughts and priorities. I start writing a draft because the AI doesn't know the intricacies of PodScan or my unique insights as a founder and a podcaster. So this is where I add the special sauce, so to say, and then I write my little draft around this outline and then I put it back in because it doesn't stop there.
Arvid:This initial idea is not everything. This is a back and forth process for me. So after I've written this draft, I might go back to the AI and then ask it specific questions about the draft that I have written. Things like, what might I be missing in this article? Is there anything I skipped?
Arvid:Or is there a part of this post that doesn't quite fit? All of these are one liners that are just throw in and then look at the response from the AI. And because the context windows of these tools are so huge, you can do hundreds of these questions and it will always remember the actual text. More questions, I'm just thinking of a couple more. Sometimes I ask, can you spot any arguments that could be easily defeated?
Arvid:I kinda tell the AI to antagonize me a little bit, right? What is wrong with this article? How can I make this blog post stronger and more convincing? I asked that quite a bit because I wanted to be something that people think, oh, yeah, this is true or this helps me overcome my own inhibitions here or my own misperceptions. And one of the questions that I found particularly useful is if you were a skeptical reader, like something in in the AI allows it to take a completely wrong or a completely opposite perspective.
Arvid:So if you were a skeptical reader what counter arguments would you have to this article? Which I can then defeat by even sometimes having having the AI tell me what I should say to that point or what I should at least address. So this collaborative back and forth process helps me write more robust and well rounded content. I still write it. It's just like there's a partner that helps me making it better.
Arvid:It's like having a writing buddy who's always available, never gets tired, and has access to a vast amount of general knowledge that I might not always have. The key here is that I'm not using the AI to write the post for me. I'm not prompting it and then taking the text adverse verbatim and posting it somewhere. I would never do this. I'm using AI as a sounding board, a brainstorming partner, and the devil's advocate as somebody who can kind of fight with me on the thing that I'm writing.
Arvid:And the final product is still distinctly mine, but it's been refined and improved through this AI assisted and AI guided process. Because let's be honest, the AI pulls me into certain directions that I might not have personally found, but I see this as a benefit because it makes the the articles, the posts better. Let's talk about how AI has revolutionized my coding process because if there is any impact that AI has been having over the last couple months weeks, particularly with my work on PodScan, it's that part. This is where things get really interesting, particularly if you're also a developer. And even if you're not, this is just really cool.
Arvid:This is how technology works today. Let's let's look at the old way in the past. When I needed to implement a new feature, my process looked something like this. I thought about the feature and how to implement it, and then I just started writing code. I would scaffold like the methods and the classes and whatever.
Arvid:I would frequently pause to look up documentation for all the tools that I would need. I remember from my JavaScript days, I would always have a documentation window to lodash or underscore or whatever kinda library I would use for these very useful functions over arrays and objects and things that you always use like mapping, reducing, iterating, you know, wrapping, like zipping, like all things that you need to use all the time that are not part of the core language framework. There are libraries for that. But I always needed to look at the documentation because I just did not remember the function signatures. It was just not not in my brain to just keep these things in line, I always have documentation up.
Arvid:And then I would start writing, I would run into issues, I would debug, rinse and repeat. And in the end, I would get a working version maybe a couple hours, maybe a couple days into it, then I would refactor and optimize it so that I would fit into the system that I would be building it for. And this would often take hours if not days like that for even the simplest features just because the process was unyielding. It was kinda hard to get everything right in the first try and that never works so it was always very iterative and I would write every single line of code myself. Now today my process looks radically different and it looks different to a point where I'm starting to kinda forget how I used to do it in the past, which is not a good thing.
Arvid:And I need to kinda struggle with this a bit more because it's so easy to use AI systems for this. But again, real example here, I needed a function to calculate the average word count of the last 10 or so podcast episodes that I have transcribed on PodScan. Just really to it's it's kind of a monitoring thing, an observation thing where I wanna see, okay, does my transcription stuff still work? So I always check the last 10 episodes that came into the system and get the average count. And if that count is very low, then something is broken.
Arvid:That's the general idea. So here's what I do or what I did for this I guess. I opened a new PHP file in PHPStorm or no, I think I opened the module that this was supposed to be in. I created a new function with a descriptive name like calculate average word count. And then inside that function I wrote a comment that describe what I wanted the function to do.
Arvid:Something like load the latest 10 podcast episodes and summarize the number of words in each of its transcripts and return that number. Just that's what the function is supposed to do. And then I hit enter after this comment, and within a second, my coding assistant which is obviously AI powered generated a solution like under a second. And the AI generated code there looked something like I would pull the latest episodes, it would start a counter and it would go through each episode, count the words and add that to the counter and then divide it by the number of episodes. This guy, I'm not gonna read you the code here obviously because this is a podcast and that would be strange, but it did exactly what I told it to with the exact right module names, with an algorithm that makes perfect sense to me, and that I could quickly check and verify and test and it worked.
Arvid:And here's the kicker. In about 80% of the cases, this generated code just from a comment is exactly what I need or very close to it. It usually understands the context of my application because, like with this example, it knows what an episode is because it lives in the code base where episodes are defined. And it knows how to query it because it also understands the framework, which in my case is Laravel and the eloquent framework, which allows me to send queries to my database. And then it implemented a straightforward solution to the problem.
Arvid:But this is the thing about AI coding, my job isn't done at this point. I I still need to review this code and I need to test it and potentially refine it. Maybe I wanna add error handling or perhaps I need to account for episode without transcripts. The AI has given me a huge head start here, but I'm still the one steering the ship. That's very important.
Arvid:But this approach has still dramatically changed how I think about coding. Because instead of focusing on writing the code as quickly as possible to get to a point where it kinda works, I can now focus on writing the most precise and clear comment or call it a prompt ahead of time as best I can. The better my prompt, the better the AI's initial output and the less work that I have to do to refine it. It's kind of teaching me to be clearer in my own specifications both to the machine in the comments but also to myself and how I think about the thing that I'm building, this particular function, the particular module, the whole product. And I think this is a valuable skill in itself.
Arvid:Prompting teaches me how to think about things in ways that I can communicate. And the time savings are enormous too obviously because what might have taken me an hour to just figure out often takes just a few minutes because the figuring out part is largely done. Right? I knew obviously I would pull the data from the database, I would iterate over it and then I would kind of condense it into an average. This is how this works.
Arvid:Obviously, this is a very simple example but maybe in between fetching the data and summarizing it I would have forgotten the function signature for for each or I would have not known what the exact, name of the attribute is that I'm trying to summarize. Any of these things particularly as it gets more complicated is always a cognitive effort and the machine has looked through all the other things of all the other pieces of code in my system and pulled in the relevant information. Let's dive maybe into a more complex example here that really showcases the power of AI, in my own software development at least. I have a little story, recently, over the last week actually, I completely reworked what is probably the most high throughput hot code path central system of Potscan, like the beating heart of the system, which you can probably imagine was quite scary to build but it is very important. It's the internal podcast queuing system, right?
Arvid:PodScan is a system that pulls in a lot of podcasts from RSS feeds, 2,500,000 of them every single day. I check these feeds and if there's a new item, I pull it in and I transcribe it, right? And obviously, 2,500,000 podcasts in the world, That means somewhere between 20,041,000 new episodes a day. All of these things trickle in whenever they do, and then they can be worked out whenever I have access to them. And I have a fleet of servers in the background that do nothing but transcribing and they need to get this information from somewhere.
Arvid:So I have an internal queuing system. The problem was that the existing queuing system that I had was a resource train. It was sitting on top of the database. So that would cause a lot of database craze to figure out what is the next item to process, What is the next episode that should now be transcribed? Right?
Arvid:Is it a high priority episode that I should pull in right now? Or do I have some time? Can I pull in an older one because there's no new episodes just yet? It was a lot of stuff. I had to constantly count the items in the database to make sure that I would fetch the high priority ones over the low priority ones and I needed to juggle this with the the actual size of the model that I would use to transcribe it, just a lot of data back and forth from the database resource drain.
Arvid:Because in the database right now, there's over 13,000,000 episodes which all may or may not have a transcript already. So you know that any query that tries to figure stuff out over 30,000,000 items particularly when it comes to text items in that particular database row will take a while. That was a problem. So and I I wanted a solution that uses not the database because that would be a round trip over the over the net. But I wanted to use the redis that is already running on my production server.
Arvid:I wanted to use like a caching system that was already there, already being used for caching super fast, and I wanted to use that as a queue. And instead of starting to read up on Redis, the queuing systems or whatever, I dived not into coding, but into a conversation, an architectural conversation with Claude AI. So here's roughly how that conversation went. I might actually wanna share that at some point, because it's just it's just interesting to see this dynamic. I started with, I wanna build a queuing system for my podcast transcription based on Redis using the Redis commands that are highly performant from inside PHP.
Arvid:Here are the requirements, and then I listed like a list of 6 or 7 requirements. The first one was it needs to have at least 4 different kinds of priority, like high, medium, low and other. It needs to grab new items from the highest priority first and then from medium, then from lowest, then from other. Now if there's none and any left it goes one down. I need to be able to add new items on the fly to any of these lists and the collections in which these queues exist need to be sorted by date, by the date where the podcast is posted at or posted on.
Arvid:So I can always pick the newest episode first for each of these queues. And then I told it I need metrics, queue size, processing time per item, estimated time to process the whole queue, and this needs to be a standalone service in a Laravel app. I think that was that was all of it. And then I told it, can you help me design the system? And then it took a second and it started.
Arvid:It started generating code. And because I gave it such a holistic perspective on what the service was supposed to do, it created an initial version that when I read through it made perfect sense to me and I could directly copy and paste it into my application. I could literally press the copy button, paste this into a new PHP file, call that file from where the requests for the queue come in and get the get next function to return me an item that I could then send over to my servers. It was incredible. It had taken care of all the priorities, it had even made the system adaptable enough for me to add more priorities or remove priorities over time.
Arvid:It had started, a metrics analytics function that I could explore and maybe do more with, and it worked from day 1. From second one, I guess. I quickly put it into my my servers, current system. I implemented it very quickly, and I tested it for a couple minutes and it worked. The load on the database immediately vanished because all of this was now in cache.
Arvid:But obviously, just a starting point, right? I said earlier coding is a process of code review and constant iteration, so we continued the conversation. And by we, I mean myself and whatever Claude is, that magical being on that platform there. And as we continue that chat I realized I needed a couple more things, right? Because now my database wasn't created at all anymore but there was still some kind of list of items in there.
Arvid:My queue was previously in the database so it was just ignored right now. So So I needed a way to preload existing items from the database into my new caching queue and I needed a method to estimate processing time based on historical data because I wanted to know is the queue going faster or slower because that kind of decides on what level of transcription system I use for certain items. Right? If the queue is going fast then I can maybe use slower transcription and if the queue is going really slowly I need to use faster transcription for it to catch up. And Claude quickly added those features.
Arvid:It got a preload from database function that automatically would load all the items that were in the database that were potential candidates for the queue and it would add my estimate processing time and all of that. It was really nice. The result of all of this was that within minutes, I built a high performance system that is capable of handling millions of podcast episodes. It is mind blowing to me that this critical part of my application that is now couple days after I did this running successfully in production was written entirely by an AI system. With me primarily doing code review and providing guidance.
Arvid:And what's even more impressive at this point is the scalability of all this. Claude helped me to do some quick calculations because you can't just keep talking to it, right? It creates code for you sure, but it still like a person you can talk to who just created code. And I I asked it, well, how much resource drain would that be at scale? And we found that even if I were to queue every single podcast episode that ever existed, we're talking about a 170 +1000000 episodes, the system could handle it without breaking a sweat.
Arvid:The memory footprint that was calculated it was based on the number of items 170,000,000 and the size of the data, right? Like there's an integer for the ID, there might be a string that is like 5 or 6 long for the priority, and then some other thing and it added them all up, multiplied them by 170,000,000 and told me this is the amount of RAM that I would use. The memory footprint that this would have at max at all the podcast episodes out there would still be manageable on my existing application server. And it would use less than 40% of the available RAM on that machine, like 30 gigabytes at max, which is a lot of RAM, but for a server that can handle it. And that's that's the biggest the queue will ever be because over time it will obviously shrink, right?
Arvid:Because I will transcribe these things into the database. And this is a massive improvement over the previous system that I have. Because that one would slow down to a crawl with just 20,000 items. Now I can finally import every single podcast episode out there that I can find in the feed that I read every day. We are talking about millions of episodes that I could ingest every day.
Arvid:It's really cool. The impact on PodScan's capabilities here, as you can probably tell both from my excitement and the fact that these numbers just jumped from 20 ks to 170,000,000, that impact is enormous. And it all started with a conversation with an AI. That's the crazy part. I could just open up a chat window, a cloud dotai is where I go and then a new window pops up, I start talking and code happens.
Arvid:That is mind blowing. Like this did not exist at the last decade. And even if there may have been a tool kinda like it, it did not cause these results. Let's switch gears a little bit and talk about AI not just as a code generator, but how it has become my go to assistant for troubleshooting my server issues. Because solo founder, right?
Arvid:I wear all these hats and sometimes those hats don't quite fit right particularly when it comes to system administration. It's something that I've always kinda understood somewhat like how to administer a server, how to install software, how to keep it running but I wanna code, I wanna build things and I wanna maintain them. It's not the necessary the thing I'm best at. And recently I ran into an issue where I just could not connect to one of my back end servers. The others were fine, but that one did not work.
Arvid:Like the SSH connects to to that servers sometimes worked, sometimes didn't. And that was preventing me from deploying updates on that server. And as you can imagine, that's a pretty big problem if you wanna deploy new data on new code on your configuration. So here's how I tackled it with the help of AI. I took the error message that I found in my deployment system, right?
Arvid:I'm using Laravel Invoyer that just takes GitHub repository, pushes it onto the server and takes a couple steps to deploy and came up with a message. I copied this message and I also tried to manually SSH into the server and I copied that error message and I opened the conversation with Claude again and I said something like this. I have a Ubuntu server that I'm having trouble connecting to. Here's the error message I'm seeing and I posted the error message. Here is what happens when I manually connect to it, post that message.
Arvid:What could be causing this? How can I diagnose the problem further and what are potential solutions? Then, you know, it takes a second to respond, and then Claude analyzed the error message and suggested a series of steps to diagnose the problem. It explained each step along the way and why it might be relevant. It was also very cool so I read through that and then following its advice I checked several log files on the servers.
Arvid:Once I managed to get in using you know another SSH attempt took me a while but it worked And then I checked those logs files, I copied relevant lines from these logs, and I pasted them back into Claude for further analysis. So through this back and forth, we, and yes, this felt like a we, it was very collaborative, like there was a systems administrator sitting on the other side of it. We identified the issue that was related to a configuration in the s s h d config file. Also related to the fact that there was kind of an avalanche of bots trying to connect to SSH, so some kind of limit was raised which then made the SSH server drop certain connections in random intervals. And Claude suggested a specific change to make.
Arvid:Before I made the change, obviously, I asked Claude to explain the implications and to kind of second guess itself. And I Googled for what this is and what there might be, what issues might be around that Stack Overflow is still very helpful with that. So once I was satisfied with that the explanation I made the change and voila, I could connect again. But Cloud did not stop there again, it went a bit further, it suggested how I could kind of prevent this in the future. I should look into activating or reactivating fail to ban which is a tool that helps protect against these kind of attacks.
Arvid:And then it walked me through the process of setting that up as well and making sure that the configuration works. And again, mind blowing, I didn't know what fail to ban was in the beginning. I didn't really know what the s s h d config file looked like, but Claude just gave me the right commands to inspect it and explain it in a way that I felt safe to actually execute the commands. I still checked, I always do whenever I run a command on a server particularly if it is done as a super user, I check what is this and what does this configuration do. Sometimes I Google the full command before I run it just to see if Google finds anything wrong with it, but I was being exposed to the solution of a problem immediately.
Arvid:It's like somebody had seen this problem before and that somebody was Claude and then it kinda adjusted its description to what I would understand. It's extremely helpful. And a couple weeks ago, I had a different issue. I was having trouble connecting to my Redis instance on my main server, which is problematic. Just for a minute or 2 or 2 in the morning, my server had an error message that it didn't connect.
Arvid:So I shared that error message with Claude and asked for potential causes. Claude gave me 3 or 4 potential causes and the commands to check, checked the first two, didn't find anything, checked the third one, and all of a sudden it was like, okay, there seems to be a problem with the recent system update. And then it advised me to check specific log files which I didn't even know existed. I looked into these logs and sure enough there had been an automatic software update right before the problem started. It had restarted the Redis system and for those 20 seconds or so no connections would go through because the system was going down and going back up and loading everything into RAM.
Arvid:So Claude helped me interpret those log entries, suggested roll back of the update or how it could prevent us in the future, and then it guided me through the process of adjusting my automatic update settings to prevent similar issues in the future. And what amazes me is it's not just that Claude could help me solve the problem, that's already pretty amazing. But how it walked me through the process and it explained each step. It was like having a very patient, knowledgeable, white bearded systems administrator available 247. Willing to explain things in detail and never getting frustrated with my lack of expertise.
Arvid:They actually I think Claude enjoys my lack of expertise because then then it can explain even more. One of the challenges of being a solo founder is obviously then keeping track of all the solutions that you find to issues like this, right? You have to keep track of all the processes and procedures that you just develop over time. So after resolving issues like the ones I just described, I've made it a habit to create documentation or an SOP, a standard operating procedure for future reference. And guess what?
Arvid:AI helps me with this too. Process for this, I draft the document, I layout the steps that I took to resolve the issue. I actually shared this on Twitter the other day, just a document as I had it, and then I asked Claude to review it. I might say something like, I've written an SOP for handling SSH connection issues. Can you review it and let me know what I may be missing, What a future employee or future me might misunderstand?
Arvid:Is there anything wrong or any sensitive information here that I should not include? And how can I make this more clear or comprehensive? And then Claude provides feedback and it often catches things that I've missed or suggest clearer ways to explain certain steps. Maybe pull them apart, maybe put little explainers in there for things that I assume, but they should not be assumed. I revised the document based on this feedback and then I might ask call to generate a quick summary or checklist based on the SOP, which I can then use for quick reference.
Arvid:And this process again helps me create more robust, clear and comprehensive documentation. It's a third kind of writing, right? The first writing that I had was for blogs, the second writing was kind of code for the system and now there's one for people to execute documentation. And it's always like having a fresh set of eyes look over my work. Something that never tires of finding errors and catching potential issues before they become problems down the line.
Arvid:That's the 247 here again. A lot of my day involves communicating with customers and that might be responding to support tickets or writing marketing emails, crafting product updates, building public tweets anything. Clear and effective communication is very important to me because I need it. I need to be able to communicate very clearly very succinctly what I want to say and AI has become indispensable here as well. As a communication coach of sorts, Right?
Arvid:When I receive a customer inquiry or support ticket, my process typically looks like this. I draft an initial response, and all of these are kind of the same. Right? I draft the initial thing. I bay just based on my understanding of the issue and the solution that I have in mind, I take this draft and I paste it into a conversation with Claude or chat GPT and I ask like, hey, I've drafted this response to a customer who's having trouble with this issue, can you review it and suggest improvements?
Arvid:Tone, clarity, completeness, is there anything I might be missing or any follow-up questions that I should ask? Then the AI provides suggestions for improvement, might point out an area where I could be clearer or maybe suggest a more empathetic tone which is really helpful when you're particularly frustrated with a particular customer or it reminds me of additional information that I should probably include. I revise my response, send it, and I make sure to maintain my own voice and style throughout the conversation. And as I use tools like CrispChat that has a magic reply feature, it's also powered by AI. And I look at these things immediately, new customer message comes in, I look at the AI suggested responses.
Arvid:And that gives me a quick starting point because these things are always based on frequency. Like here is what has helped people in the past. So I can look okay this should probably be the the area in which I respond to this person. I never use these things verbatim. Again never ever verbatim but I use them as a base and I modify them to match the specifics about this conversation about POT scan.
Arvid:And if I'm unsure about the modified response I might even put it back into cloth for final check. And again, more quick and consistent than just randomly typing out stuff and wondering if it's okay or not. Putting it into an AI system and getting a quick response back always really really useful and ensures that each response is personalized and genuinely helpful. Marketing emails is another thing, I think I'm almost done. I think this is the last thing that I do with AI.
Arvid:I mean, obviously, everything is AI. But for marketing emails in particular, again, I outline key points, maybe I even brainstorm it with the AI, then I ask AI to generate a first draft based on the tone. And once I have that I very heavily edit it just to infuse it with my voice and add details about parts scan that AI would never know. And then I asked CI to review the edited version and look for ways to make it more engaging and persuasive. Obviously, particularly with marketing emails AI systems come up with really boring bland marketing you know catchy headline kind of stuff tone.
Arvid:If you don't tell it to have your own tone. So one thing that I found is if you like the copy of any given author any given person or even yourself, if the AI knows how you write, you can tell it to sound like that person. Right? Do I wanna sound like, I don't know, Justin Jackson writing his emails? Do you wanna sound like a marketer I know?
Arvid:Like all of these things I can tell the AI to use and then it uses that to enhance my own abilities. And it's not a replacement for my own judgment and my personal touch, but it's like having a really smart friend who's always available to brainstorm and provide feedback. But doesn't know the intricacies of my business or my relationships with my customers. That I need to bring to the table myself. And you might be wondering, Arvind, this sounds like a lot of back and forth with a a I.
Arvid:Is it really worth it? Is it worth the time? So let me break down why I find it so valuable. First off, it's cost effectiveness. At 20 to $30 a month for most AI services, it's incredibly cheap compared to hiring human assistance.
Arvid:I'm a bootstra founder, this is a game changer. I get to pay $20 and I have experts in every single field available to me. And I don't need to pay much more than that. And unlike human employees or contractors, these tools are there 247. Whenever I'm working like late or early, I can always have a capable assistant at my fingertips and I don't need to know what time zone they are in because they're in every time zone.
Arvid:And these tools help me become better at everything I do that they help me with. They make me a better marketer, a more efficient coder, a more responsive customer service rep. For a solo founder with all my hats, this versatility is invaluable. And while it might seem like I'm spending a lot of time interacting with AI, it's saving me time in the long run. AI handles about 80% of the initial work and that leaves me to more time for strategic thinking and refinement.
Arvid:Refinement is what I'm getting better at. The initial work is outsourced to the AI systems. And there's so much I don't know in all these areas outside my core expertise, right? AI provides immediate answers to questions that I might not know. Sometimes I don't even know how to ask the question and it guides me to the answer like server configurations, log file locations, who am I to know these things?
Arvid:But the AI points me at it. It's like having a team of experts on call. Using AI helps me maintain a consistent level of quality across all aspects of my work even in areas where I'm less skilled because it fills those knowledge gaps and interacting with AI almost always teaches me new things. Like it taught me about Redis, it taught me about how you Ubuntu update software, it teaches me how to architect a queuing system, it teaches me Redis commands that I've never seen before. Like whether it's coding technique I hadn't considered or marketing strategy I was not aware of, I'm constantly learning.
Arvid:That alone is worth $20 I don't even need to use it professionally, just asking questions. So AI has become an integral part of my workflow but it's not without challenges. So limitations exist and I've encountered a lot of them. So here's how I deal with them. The biggest one I think is accuracy and reliability because while it's rare, I can provide incorrect information or suggest problematic solutions, right?
Arvid:And this is particularly critical for tasks like server management, security related issues or just talking to people. You can really run into issues there. So to mitigate this I always verify critical information especially for tech stuff and I treat AI suggestions as the starting point not gospel. It's never the end. It's never the final result.
Arvid:It's always the first initial step. And I never fully delegate tasks to AI as well. I remain involved in the process always. I review and I verify outputs. It's always a step.
Arvid:Right? Verification is never omitted. And this means I still need to understand the underlying principles of what I am doing. It might be coding, marketing, system administration. I still need to get what I am doing.
Arvid:I can never give it away completely. And because there's a risk, there's a risk of becoming too dependent on AI tools. I'm very cautious about this. I need to ensure that I still understand how the work is done so that I can do it manually if needed. Now we've seen these outages, claw has been down, chat GPT has been down.
Arvid:It's important to use AI as a tool to enhance your abilities, but not as a crutch and not as a complete replacement. It's still somebody else's platform. Platform risk really still exists here. So I'm very careful not to over rely on it. But also probably because AI tools are just evolving so quickly because what was cutting edge a few months ago is probably outdated already.
Arvid:There are people that get into AI now that have never heard of GPT 3.5 or anything like this, the systems before it that that only know what mid journey is but don't really know what stable diffusion is. Like all of these things are so quickly evolving. I need to stay aware of new capabilities and new limitations and that requires ongoing learning and adaptation which I try to spend time on but I cannot guarantee. So it's a it's a risky field to over rely even on this tech because they might just roll it back one day. Now you always have to keep up with changes.
Arvid:Even while it's incredibly knowledgeable, AI does not know the specifics of my business or my customers. So that context I always need to bring myself and verify that the AI suggestions make sense for my specific situation. And as it becomes more integrated into my work, I kind of have to be mindful of ethical considerations as well. Particularly around data privacy and disclosure. So getting customer information and copying pasting it back into Claude, I try to avoid it and I try to anonymize things as much as I can.
Arvid:I'm always transparent with my customers about how and where AI is used in my process. I think I have it in my terms and conditions as well that I have and use these tools. But that's just just the reality of today, right? You have to make sure that you don't feed data into systems you don't want the data to be in. And through my journey of integrating AI into my personal workflow, I've developed a couple best practices that I might wanna share with you because those things help me get the most out of these tools and they probably will help you even just conceptually to to get.
Arvid:The first one, prompting. Understand prompting, that's the most important thing. The quality of AI output is extremely dependent on the quality of your prompt input. For creative tasks, I've learned to ask for variety in results when I have a prompt about like find a headline for a blog post or come up with a couple ideas for something. Now, I always tell it to be extremely diverse in the length of the sentence and the topics of the theme that the way you talk about it to perspective you take be diverse, have a variety.
Arvid:I always put those words in. And for specific tasks, I aim to be as detailed as possible in my prompt to be almost exclusionary to tell it need to you need to do exactly this and nothing else. Learning how to prompt is a skill that improves with practice, but you need to experiment with it to see what different results come out of it. And I treat AI like a coworker. I find it helpful to approach AI interactions as I would conversations with a very knowledgeable colleague.
Arvid:I learn how to communicate effectively with the system like Claude or ChatGPT, and I refine my approach based on the results that I get. Just like I would with a human being that I have a conversation with. Number 3 of my best practices is to never use results verbatim, except sometimes with code if it's small. So whether it's a written response or a suggested solution, I always review, test, and verify AI outputs always, and even with code. And I might use AI generated code directly more often, but I always review and test it thoroughly.
Arvid:And I have a general workflow for this. I reflect before I prompt, then I prompt, then I verify the results. Think about what you say, say it right and then check if what comes back is actually what you asked. And this helps me using AI effectively and not just throwing questions at it haphazardly which is easy if you can do it all day long. And I iterate, I use AI in multiple rounds for a single task always back and forth, initial draft, review, modify, feedback, and then modify again.
Arvid:And I found the best results come from combining AI capabilities just with my own expertise and standing behind what I say. Understanding my business and telling stories that I know is better than having AI come up with some general stuff. And I experiment. I experiment with new ways to use AI all the time. Sometimes that works, sometimes it doesn't, but often it leads to new and more efficient processes as well.
Arvid:So where is this going? Because as I use these AI tools day in and day out, can't help but get excited about the future here. So what what I what I think is gonna happen is that we're gonna see very much specialized models. I'm excited about the potential for more specialized models trained on specific code bases. That's what I'm hopeful for the most.
Arvid:I imagine an AI that understands not just the code in the file you're writing but the full project, like the full code base of the project that you're working on and the history of your project including all these little experiments that didn't work out, all these little refactorings that you did. I think that would be incredibly powerful for maintaining and evolving complex systems. To have models that are trained on currently and always keeping up with the code base that you're working on. That would be an extremely interesting avenue to go down because I believe that those models could then work for you. They could start writing code for you behind the scenes, optimizing code for you behind the scenes and test it behind the scenes because they know the whole code base and all the context in which it exists.
Arvid:I also would love to see AI systems that can analyze code and pull in customer conversations and social media mentions to start just suggesting the next steps for your product development. Could be great for people like me to make more informed decisions about where to focus my limited resources. AI driven product planning shouldn't be too hard like connecting these couple systems and have an AI go over what needs to be done, what is intended to be done, what your roadmap already has and what people cannot stop talking about. Feel like that would be a sass in the making that I have no time to build so you might wanna have to build it. Hey, and and just generally I think we'll see the job of AI operations like DevOps, there will be AIOps, a new role.
Arvid:Something like an AIOps specialist that orchestrates AI systems within an organization. That make making sure that everybody has the full context and things are probably integrated like all the all the the the task system, all the the road map systems, all the communication systems, custom emails, all of this could flow into a central database that AI systems can then take information from. This could be a crucial role in maximizing the benefits of AI in an organization. It's just to have someone who is at this very moment probably more jury rigging than anything else but just pulling things together so that AI systems can benefit from data in all these systems. And software education will have AI as a primary component, I guess.
Arvid:Like if you start learning how to code today, you need to learn how to code with AI. It makes no sense to not learn AI coding because it will be and already is an integral part of software development in the future. GitHub Copilot, everybody's gonna use stuff like this. We might even see programming languages shipping with built in large language models. I'm looking forward to seeing the first programming language that deploys an LLM that can help you write code in that language.
Arvid:It's like an extension of the documentation of a language is having a a large language model that can already write code in this particular version of that language. Like imagine, PHP PHP9 coming out with an LLM attached for 9.0 that can automatically help you write code in 9.0, help you migrate code from 8 or 7 or 6 to 9 and help have a conversation with you about how to build it right. I could see Laravel do this. I could see Laravel deploying an LLM that helps you write Laravel code. There are tools that integrate into IDE systems for Laravel in particular that already exist.
Arvid:So they pull in AI data maybe in the background but maybe the AI itself is something that the language could come with. I would find that very interesting. And we already see chat interfaces and integrated tools. I think integrated AI assistance will become even more comprehensive, and they will understand your entire deployment and development ecosystem much better than they do today. Code review also gonna be AI assisted like obviously, catching bugs suggesting optimizations.
Arvid:I think Sentry is already doing this like if they see an error, they try to find where the error happens which usually they can do because of their tracing capabilities and then they have an AI system that analyzes the code around this error and suggests a new code review for you to look at like a new, pull request, I guess. So that you can review the pull request to see if that fixes the actual underlying issue automatically. It's really cool. And I guess prompting being the thing that it is as AI gets better at understanding code and generating code, we might move towards a future where programming is done through natural language instructions. More prompt than anything else.
Arvid:What right now is a comment that then expands it to code might just be something that is defined as code but it is spoken word. Don't really know how that's gonna go but I I see these things kind of converging and into a place where we type way less and we speak way more or we think way more. Speaking is just a way of thinking strongly I guess. So let's get back to the actual solopreneur part of all this. For solo founders, founders, entrepreneurs in general, I think I cannot recommend integrating AI tools into your workflow strongly enough.
Arvid:Like this is probably one of the most impactful things you can do. They will significantly impact and shape your productivity and capabilities, and they will allow you to punch above your weight in the competitive business landscape. Like you will be much better with AI than all the people without it. Much more efficient, just more polished and that makes a big difference when people look at your work. And look for opportunities to use AI in any task involve writing or problem solving.
Arvid:Whenever you think, an AI can help you think better. Whenever you write, it will take your output and make it better. When you code, that code can be improved. Doesn't matter, there's likely a way AI can help you do it better. And I think using AI assistance is a new form of early hiring.
Arvid:It's like getting consultants that cost you $20 a month. In aggregate, for every business aspect that you run into problems with. It's crazy. It's incredibly empowering. And I'm excited to see how these new tools and the system to businesses built around them continue to evolve and support founders like us.
Arvid:Remember, the goal is not to replace your own skills and your judgment, but to enhance them. So use AI as a collaborator, a brainstorming partner, a knowledge extender with this approach you can achieve things that might have been impossible, seemed or been impossible for a solo founder just a few years ago. And as we move towards an AI enabled world, stay curious and keep experimenting and don't be afraid to let those powerful tools amplify your capabilities. There is no purism here that's gonna help you. AI tools are here to stay and the future of entrepreneurship is here and it is augmented by AI.
Arvid:And that's it for today. Thank you for listening to the good sir founder. You can find me on Twitter at avidkahl, a r v I d k a h l. You find my books and my Twitter course set too. If you wanna support me in this show, please tell everyone you know about PodScan and leave a rating and a review by going to rate this podcast.com/founder.
Arvid:Makes a massive difference if you show up there because then the podcast will show up in other people's feeds and I would like that. Any of this will help the show. Thank you so much for listening. Have a wonderful day, and bye bye.
Creators and Guests
