0:02 I found the perfect stack for AI product development. And AI now writes 90% of my
0:08 code. This exact AI stack has enabled me and my team to push out features and
0:13 improvements faster than we have ever done before. The technology is finally
0:19 here. Yet you'll see people saying this third rate fake worst code
0:23 rejecting all AI written code. >> Your customers want you. But I can
0:27 assure you AI product development does work. And if you use it right, it can
0:31 make you more productive than you've ever been before. It can even allow low
0:36 tech and non- tech founders to build incredible SAS solutions all by
0:41 themselves. So in this video, I will walk you through this AI stack and show
0:46 you the exact AI product development workflow. It consists of three tools and
0:51 three approaches you have to use if you want this to work. If you're new to this
0:56 channel, welcome here. I'm Simon Horberg and I run a SAS portfolio of four tools
1:01 to handle social media automation, link tracking, AI, customer support, and
1:05 graphics design. We are a small team of just four people, myself included, and
1:11 we're on track to do $2 million AR in 2025 using the exact AI development
1:16 stack I'm about to show you. The first layer of our stack is vibe coding. This
1:21 is typically going to be the starting point of a new SAS or new major feature.
1:26 Not long ago, this starting point used to look completely different. Typically,
1:31 we would spend a lot of time planning a new feature. We would draft the designs
1:35 in Figma. We would then tweak and polish and turn it into a prototype. Then
1:39 finally, we would hand it off to one of our three developers who would manually
1:44 start implementing it. And if you think this was timeconuming, you're absolutely
1:48 right. And we were everything considered a small team. Earlier when I was working
1:52 as a freelance consultant at huge companies I saw how this step would take
1:57 months sometimes up to a year to complete and it would literally just be
2:01 the starting point of a new feature or product. With AI this whole thing has
2:07 completely changed. We now do all of this using one single tool. That tool is
2:13 called lovable. By simply using natural language and by describing the vibe we want, we're now
2:20 planning, drafting, designing, and prototyping all using one single tool.
2:26 Lovable is built to be non or low tech friendly. So on our team, our UIUX
2:31 designer does this. She will use some UIUX and software terminology, but other
2:35 than that, she lets Lovable go at it just from a vibe. And Lovable does an
2:39 impressive job at putting together an amazing looking, clean, and
2:44 well-designed front end. And if we take a look at the code it produces, we'll
2:47 see that it also puts together a well ststructured project using the latest
2:51 and most popular frameworks and UI libraries. So in just a few hours with Lovable, we
2:57 have a design, a prototype, and a starter in real application code ready
3:02 to go. We'll sync it directly with GitHub to access the project from other
3:06 tools if we need it. The next layer of our AI stack is agentbased coding. Once we have a
3:14 starter, we want AI agents to continue the work asynchronously in the
3:19 background. This is the step where I personally have become the most
3:23 productive. Having a team of AI agents with access to my entire code base, all
3:27 my repositories who will start working on the task I give them, do it all in
3:32 the background, and come back with a PR. This has been a total game changer.
3:37 Here's how it works. GitHub is a platform where you can store code and a
3:41 PR is basically code suggestions that haven't yet been merged into the
3:45 official codebase that goes out in production and becomes available to your
3:50 users. Now, you give an AI agent a task. The AI agent will download your codebase
3:54 from GitHub and start working on the task in the cloud. Once it's finished,
3:59 it will create a PR on GitHub where you can review, request changes, or approve.
4:04 There are multiple tools that can do this already, but one that does this
4:09 really well is GitHub Copilot, which is already fully integrated into GitHub
4:14 itself. In your repository, you simply click issues, which is GitHub's words
4:18 for task. [Music] Here, you describe the task you need You sign GitHub Copilot
4:36 [Music] and a few moments later, depending on the complexity, you will see a new PR
4:43 under the pull request tab. From here, you can review the solution and either
4:47 approve or ask C-pilot to make a few more changes. If you're a developer, you can review
4:54 the actual code it wrote here, which I definitely recommend.
4:57 [Music] And if not, I suggest configuring your hosting service like Verscell or Amplify
5:04 to create a preview link where you can see the changes in action.
5:09 And the awesome part about this workflow is that the agents are working on your
5:14 solution completely asynchronously. So you don't have to be in front of your
5:18 screen or keep a code editor open while it's doing its thing. Instead, you can
5:22 split your work up in three, five, 10 smaller tasks, fill them out on GitHub,
5:27 and let your agents work on them all while you do other things. Now, it's
5:31 worth pointing out that the quality of the results highly depends on the
5:35 quality of the task description you give it. And if it doesn't get it right, it
5:38 does take a bit of time for it to go back and work every time you ask it to
5:41 do something new. If you can include some technical description and
5:44 engineering terminology here when you write these task descriptions, it does
5:49 help the AI tremendously in doing a proper job. If you're a non-tech or low
5:53 tech, GitHub does have a basic chat interface too where you can include your
5:57 codebase and talk to the AI about it. And it might make sense to spend a bit
6:01 of time with the AI here agreeing on what the task is and then have GitHub
6:06 Copilot write its own task description. As I mentioned before, there are other
6:09 tools doing the same thing. One of them is claw code, but personally I really
6:14 dislike using a terminal for this and it's not really friendly for low or non-
6:19 tech users. Also, GitHub C-Pilot is really generous with their limits right
6:24 now. I'm on a $39 a month plan and I haven't managed to max this one out yet.
6:30 The final layer is AI assisted coding. With this approach, we're pair
6:34 programming with AI, but we're actually participating in the process using a
6:38 code editor with built-in AI capabilities. For this part, you need to
6:42 have at least some basic understanding of programming fundamentals and know
6:46 your way around a code base. I would say low tech people can work at this level,
6:50 especially if you're patient and naturally interested in trying to learn
6:54 to code. Though, you will get most out of this step if you are a more routine
6:58 software engineer. The most powerful and most popular tool for this layer of the
7:03 stack is cursor. However, there are native extensions you can install in VS
7:08 Code such as Klein or again GitHub Copilot which are picking up and getting
7:12 really good too. Personally, I bounce between cursor and GitHub copilot and VS
7:17 Code. They're both really powerful and strong options when it comes to AI
7:21 assisted coding. Although cursor is generally ahead when it comes to
7:24 delivering a smooth developer experience, but honestly I haven't
7:29 really spent a whole lot of time coding like this recently. Both I and my team
7:33 are now spending most of our time in layer 1 and layer two of this stack.
7:38 Just this last month, we've built and launched an end toend Vibe marketing
7:42 experience for one of our products, FeedHive. We've updated and tested a bunch of new
7:48 features and improvements, and we've finished a completely new SAS that I'll
7:53 be launching soon. If you go to foundersstack.pro, you can see all the products we're
7:58 building, and you can buy lifetime access to all of them for a single
8:03 one-time purchase. 90% of our efforts were spent in layer 1 and layer 2 when
8:08 building this stack. A lot of people say vibe coding produces extremely
8:13 lowquality results. And as I see it, it all comes down to two reasons. Most
8:17 often, it's simply a matter of half-assed, rushed, lowquality prompts
8:22 and task descriptions. If you went to Upwork and hired a human developer and
8:27 gave them this same description, I can guarantee you the results will be
8:32 extremely low quality, too. Like, what did you expect? Secondly, there is a
8:36 limitation if you know absolutely nothing about coding. So, if you're
8:41 completely non- tech and want to build a SAS this way, I highly encourage you to
8:46 become at least low tech or maybe even a lightweight software developer. You
8:50 don't have to pass an exam. You don't have to pass a job interview or get to a
8:53 level where anyone would formally hire you, but just spend a bit of time
8:57 learning some basic fundamentals of software engineering. And I promise you,
9:01 the final output of VIP coding will become exponentially better. So, now you
9:06 know how to build, but maybe you still don't know what to build. I got you. Cuz
9:11 there's a way you can access thousands of ideas that people need but haven't
9:15 yet been turned into a SAS. And in this video, I'll show you how. So, jump over
