The Case with Vibe Coding: Assembling the Ultimate Team!
What is this so called vibe coding?
In today's society we've entered a new technological age, the era of AI (Artificial Intelligence). As you know in 2024-2025 popular companies have been out there producing high-tech AI models like OpenAI's ChatGPT 4.1 or Anthropic's Claude 3.7 Sonnet, and people now use these models in their day to day lives for basic tasks! For example, you can ask ChatGPT to help you with a homework assignment or help you complete research on a certain topic for your work. However, these models are capable of much, much more! And the biggest newer trend that you might hear thrown around is called "Vibe Coding". This new software development style had been popularized by Andrej Karpathy in February 2025, in which the end user starts talking/chatting to a LLM (Large-Language Model) and asking it to create various different coding tasks/files for you. For example, you could ask Claude to "Make me a website for my burger restaurant that looks modern with good colors in a single html file" and the LLM would provide you with the html code that you could paste into a html file and run it within your browser. This is just the TIP of the LLM iceberg and with the power of the human brain combined with Artificial Intelligence, we are now able to pump out code more fast, effectively, and cheaper than ever! Today we will be exploring how you can harness this power to build the ultimate development team utilizing you and the AI.
The vibe coding icebergs...
As shown in Figure 1, the amount of vibe coding tools out there right now is quite insane. With the large amount of vibe coding tools available and LLMs releasing every single week, vibe coding can spread upon two different icebergs, with the LLM iceberg and the full stack autonomous vibe coding iceberg. In the LLM iceberg, it's really split into two categories, Great and Decent LLMs for Vibe Coding. From my testing of various different LLMs, I can determine larger and more popular models like Claude and their "Sonnet" series is the best all around coding model, which is why it is at the top of this iceberg. However if you want a great all around LLM model for anything, Google's Gemini is your best bet specifically Gemini 2.5 Pro. Then the other major LLMs fall down the order as they are also pretty good for what they are intended for. Then on the second half of the iceberg, we having the decent LLMs. The two models I selected for this part of the iceberg is DeepSeek (R1) and Moonshot AI's Kimi K2. The cool thing about these two models is that they are both OPEN SOURCE and able to be hosted locally on your machine whether you are on or offline. These models typically will spit out good responses for general queries and can handle some basic vibe coding tasks as well and Kimi K2 can do a lot of advanced tool-calling, so if you aren't willing to have larger corporations charge you a lot, with a lot of rate limiting or just want some privacy, look no further than DeepSeek R1 and Kimi K2. All of the above listed LLMs are great if you want to copy and paste the code output into one or more files inside of your code IDE, but what if you could utilize some of these models but inside of your IDE or via a chat interface on the web? This will bring us two the second iceberg of vibe coding, the full stack web application iceberg. In this iceberg we get into Great vibe coding IDEs and decent full stack vibe coding services. At the top of the iceberg we have Cursor & Windsurf which are a vs code fork along with some extras that bring an AI chat sidebar to the IDE with a selection of models to choose from and also offers AI code auto completion. You can choose to build chrome extensions, web apps, or even applications that can run on your computer with these IDES, also You can set up these IDEs to be autonomous once you give it a prompt and/or a set of rules. For example, you can choose to allow a command/code execution to go through once or always allow them meaning you can just sit back, relax, and not touch a thing while your AI developer goes to work for you, which I think is pretty cool. Bytedance's Trae IDE and Microsoft's VS Code also offer an AI chat feature with similar features as well. But what about the other half of the iceberg? The full stack web application vibe coding services. Well what tools like Lovable, v0, Bolt, and Replit do is allow you to use a chat interface over the web like ChatGPT, but can execute, preview and write code right inside your browser and you can login to view this code on any device unlike the IDEs where files are stored LOCALLY. With these tools, they also offer you to connect popular services like Supabase, Stripe, and Auth providers so you can get a fully functional web application with working user data! These tools can even publish your app on the web with a domain on their site or a bring your own domain program. However, I truly am a believer in the saying, "Use the right tool for the right jobs". So you have to be careful and decide which LLM or service you wanted to use to make your vibe coding projects in, some are better than others which we will get into in a bit. Now that we have squared away the two icebergs of vibe coding, let's see how these services: IDEs, and LLMs work in a real world scenario.
How do these tools fair against the "Burger Test"?
In order to see how each of these models performed creating frontend code/design for a webpage, I decide to engineer a prompt that a person might ask a LLM for. The example prompt I gave is "Make me a website for my burger restaurant that looks modern with good colors in a single html file" and here is how these models and services performed using how model output results look in my opinion on a chart. As shown in Figure 2, out of all of the available models that were out at the time of my testing, A Qwen Model, Qwen3 coder actually scored the highest, as it provided a nice and modern website with working images, prices, a menu, and a section to implement online ordering all wrapped into a nice restaurant aesthetic. While the other models did a relatively good job they weren't near as complete in a one shot prompt as Qwens output was. Meanwhile, as shown in Figure 3, the AI code IDE Windsurf using its SWE-1 model scores highest here as it offers a similar and complete experience like the Qwen model gave. However, if you want user orders to be easier Lovable or Bolt would be your best all around choice for easy payment and DB setup via Supabase and Stripe. The Burger Test is just a personal and fun benchmark that I made to measure how these models perform on looks and to see if they are ready for the modern web. For actual raw performance data, math, and scientific benchmark data, I would recommend going to Artificial Analysis as they deeply test each and every new model and put it through their paces, especially since new models keep dropping as I write this such as GPT 5, which did way better than the previous GPT-4o model.
Wanna see the test results and outputs that helped shape these graphs for yourself? Click Here to download a folder containing all of the output .html files I got from these models of the graphs above. Or do you want to run the test yourself using the latest models and get performance graphs? Consider visiting BurgerBench, the AI benchmarking tool that me and Gemini 2.5 Pro created which is using my prompt stated earlier!
Building the ultimate ai development team!
Now for the question you may have been asking and the part you've been waiting for, how do I build the ultimate AI development team? As shown in Figure 4, Through various tests and things I have built which you can view over on my GitHub profile, I have came to the conclusion that my favorite AI models for all around work on those various projects are models from the Google Gemini family and the Sonnet models from the Anthropic family. In order to kick-start the process of building this ultimate team, it all has to start with YOU, the prompter. Your job as the prompter is one of the most important jobs in this team, as you will be acting as the team manager for the LLMs, for example in plain english you would write out and a task for a Claude model in Cursor and it would go out and execute the task for you. You are overseeing and in control over what the models know to do. My best recommendation is to have good english and prompt engineering skills. Like most development teams, managers have to communicate with their Junior and Senior devs. For the Junior dev in this team which will handle the most simple tasks, I like to employee the Google Gemini family of models which is Gemini 2.5 Flash and Gemini 2.5 Pro as of writing this. These models are more than capable of helping you draft out basic and/or good looking frontend UI and help you flesh out your ideas for the product you are building. They are a key essential as the junior dev to help kick start the project, and to give you ideas and critiques. After I see a result I like from Gemini, I then like to pass the work over to the Anthropic Claude family of Sonnet models. I typically will use Claude models for more advanced tasks in a vibe coding project. For example, I will use Claude to build advanced features into apps such as working user authentication via Google OAuth and Supabase, and connecting to and writing my backends such as SQL queries for Supabase, or implementing something like text field that makes an api request and returns a result. However it is essential if you are using an external service over an API like Supabase or setting up Convex, you will need to go over to their respective websites and setup and generate the API keys to give to the LLMs. With that being said, this is typically the cycle I follow for building all of my vibe coding apps recently and it's really been a great experience with these models. If you prefer a different cycle, that's ok! Everyone has their own preferences for what LLMs they like and the product that they want to build. Once again, use the right tool for the right jobs.
Conclusion
We've been over a lot today, from the beginning of "vibe coding" to building the ultimate ai development team, there is a lot to unpack here. If you enjoyed this article, be on the lookout for more like this and feel free to DM me on my Twitter or on my Discord (Corvettefan101 #1738) with any questions, comments, or concerns about it. Your opinions matter to me and I want to publish the best content I can. Thanks for reading, CFNVibez Out!