Gemini 3 Pro's Vibe Coding (Fully Tested!)

Okay, so over the last 2 days I've used tens of millions of tokens on Gemini free pro. Basically building my own AI startup from scratch and this video is going to be my own vibe check on how good the model is at various things. So what I was building is that for my application hyposper that has a Black Friday sale going on right now. There will be a coupon code down below if you interested. I basically wanted to make a system where I could easily add a support sales AI agent to the bottom right of all my websites and products to help people like understand what it's about, how it can fit into their workflows and basically to drive more conversions. So, I started building this out from start to finish with Gemini free pro. Currently, it's still on local host, but I will deploy it after this video is up and there will be a link down below in the description. And basically, the idea is that I can build my own custom AI agents for all my websites such as Hyperwisper and then also my Cloud Code Masterass as well.

And then I basically made a bunch of notes during the process that I will be going over in this video. Anyways, I'll quickly go through the application if you're interested. So, you log in and it will be a private beta for now for a select number of people. And then once I have my account, I can press create agent and I'm going to make an agent for my website hypospin in my case. And basically, this is like pretty good and pretty complete. I was very surprised by how well Gemini did over here. I did have to go through a few iterations of versions and then I can add in my own like website content. So like hyposper the help website as well and then it will fetch the websites and do rag across all the contents to basically help train the chatbot and then I can add my own Q& As's and then I can basically go to chatbot which will be integrated on the website and then scroll down and say something like what is cypisper over here press enter and then it will tell me what it's about. So this will be integrated onto a website in like a small chat widget at the bottom and then it has like all the other features where I can see chat logs that have happened between users and the application any analytics as well and then also change the model and do like a bunch of other things that you would expect from a AI support chat agent. And

honestly I was pretty surprised by how complete of a project I could make using only Gemini 3 pro. I only had to switch to composer one occasionally and that was when I wanted something done really fast and for type safety related issues. And basically what I started out with is my own dashboard template that I have on my community. And that basically has a bunch of things like TRPC and Superbase off and stuff set up because I find that for most models they're pretty bad at making like good architectural choices and they would avoid something like TRPC or maybe Superbase and you have to explicitly have some kind of boilerplate code to put them in that right direction. So I came up with a plan for the entire application for me after chatting with what the end products will look like. And yeah, I managed to get a pretty basic working version one in only a few prompts. Unfortunately, during that time, it did seem to loosen it a bit and come up with things that shouldn't have existed like ctx session user ID when it's just user ID and then also using something like AI SDK version 5. It struggled with that as well and really got stuck with the types. So then I had to download the exo code MCP server which has like much more upto-ate knowledge on how AI SDK version 5 works. And honestly by the second version when I got it to fix a few type errors I was surprised by how well some things works out of the box like the embedding system and the rag base system and everything worked well. Of course it's not the most efficient one because it doesn't have the indexes and so forth but it was a good start. But at the beginning stages I noticed that I had to push the model to not adhere to the existing design. So the existing design of the dashboard template is like pretty basic and it kept copying that design over for the application and then I had to say like okay come up with something completely new something fresh and so forth use these colors and then it kind of did something similar and downloaded new components and so forth but one thing I noticed during Gemini 3 pros behavior that it does have some sudden blind spots. So in this case it tried to import something that just didn't exist and didn't check if that file actually existed or download that file to begin with from shaden components. It also struggles with this certain class of errors after discovering it and it just kept burning free tokens. So I had to switch back to cloud code to fix this sudden error and then it just had difficulty integrating something like recharts into codebase as well. So then what I did to push the Gemini 3 pro away from the boring design that the template has to new design is I basically took a screenshot of another website that I like the design of and said okay like here's a good design take some inspiration from it. And then this is what I came up with. And it does seem to adhere to design of images pretty well. And it's pretty good at fixing its own UI bugs. So you can see like this corner intersection does not work well right over here. So I just took a screenshot of this particular corner and it said, "Can you fix this? " And it identified the problem to fix because I guess it recognized the surrounding text. And then it made a simple fix with the breadcrumb and the uh sidebar like aligning it properly. And then this is

version six and it started to look kind of okay. And what I did notice about the model is that every time it asked it to make improvements on the design, then it actually felt like it was moving forwards, especially with vague instructions to the improvements. Any previous models like Clawude or GPT models, it felt like they reached a point where they just couldn't really make any significant improvements to design no matter how hard I push a model. It also did seem better at making designs consistent across different pages of the application. Like I noticed compared to other models, they would mess up on like the padding or spacing of components across different pages. Whereas Gemini 3 didn't really have a problem with that. And I suspect that its ability to understand design and be consistent and coherent across many different files in the codebase when it comes to design is strongly tied to its ability to basically come up with these examples that you can see here where like someone made a 3D Pokémon on a planet in 3JS. I think someone else made this kind of like video game which is like a roller coaster theme park kind of thing. Someone else made the Lego simulator which is pretty popular online. And I suspect that all this ability is related in some kind of way where because it's a multimodal model and can also generate images as well. Like when it sees a code, it has a better visual of whatever is going on inside its head rather than crunching through the code. And I guess to kind of use humans as an analogy, there are some people that when it comes to math like linear algebra and calculus, they have a pretty good visual of the problem that's happening when they're solving it. like in their head they kind of imagine the equations or the graphs or like the matrices and so forth and the transformations and stuff like that and then there are some people who just kind of memorize the rules to do with calculus and linear algebra and then they just grind through the equations and they get the final result. It feels like the other models like Codex and Claude code are on the second category, the latter category, whereas Gemini is under the former category where it has like a visual understanding of what's going when it's actually going through the code and making design improvements. And I suspect other models like the cloud models or the GPT models will not get as good at Gemini as design unless they become like multimodal models in some way. And in my own experience when using the model for design, it did pay attention to smaller like finer details when it had the freedom to kind of like a mix something from scratch. So you can see in this example with the game, it kind of like shows confetti when the like creature touches one of these objects here or the coins. And when it was making the landing page for the uh AI support agent application, then it added this like tiny detail over here. So I can hover over any of these and it has like a nice jumping animation. This becomes bigger. And if you zoom in, it automatically added the moving dashes. I just did not have to prompt it to add that. And it added it here as well, but for some reason it did not add it over here. And yeah, this component also looks really nice on the landing page that it made. And what I think I'm going to get it to do is generate more SVG diagrams for the landing page because I have noticed it's really strong at SVG generation. So you can see this example over here when it comes to generating Kirby and someone else actually did use it. So they basically asked, can someone come up with SVG designs kind of like this one? And then after Gemini 3 came out, it basically got Gemini 3 to make those designs within Kasa. And yeah, overall it seems really strong and good at SVG designs, especially when it comes to landing pages. And you can see adds a nice animations as well. And then it decides to randomly delete files as

well. And I also mostly use Gemini 3 Pro with Incursor because a new ID anti-gravity is not really stable yet. But as it becomes more stable, I will begin to use that more over time. And you can see that even within curs it's not properly aligned to ID environment because it starts randomly writing to-do lists in markdown and then it just like starts repeating itself over here. Also another time when it came to integrating billing, it started to put its thinking as comments instead of actually having it in the thinking place. And one thing I did notice about Gemini 3 when using

it is it's kind of overreaching in some ways. One thing it likes to do is go on these side quests and that's basically the term that I'm using here. maybe there's a more formal term and when it's integrating something it kind of notices another problem and is like okay I'm going to fix that problem as well. So for example in this case I was asking a question related to something else and then it noticed another problem that could also fix and proposed a fix for it. And then in this other conversation I said does a chat bubble logic work and it says yes what I implemented works and then it says like not for this other thing and then it continues to integrate that. And I think in my experience if I had this particular conversation with Claude or GPT then it would just say yes it works and not be as overreaching. And in other cases, overreachingness and side quests can be pretty annoying because I just asked it to make a change in a landing page and then it starts to do rewrites in the next uh config file.

So I have a couple concluding thoughts here. First of all, it is very fast most of the time. And whilst being fast, it is also creative as well. So what I did is when I was making the landing page, I got it to come up with 10 different designs. Some of the designs using Tailwind and some of them not using Tailwind. So this is the first design I came up with and that's like pretty basic using Tailwind. This was a second design I came up with. This is significantly better. This is a third design I came up with and I think this was not using Tailwind but of course it looks more like an agency website rather than a SAS website. This was the fourth design I came up with and this was also using Tailwind. This is also using Tailwind, the fifth design I came up with. This is also another Tailwind design. And I liked this the most because it has these widgets over here. And then I told it to expand on the widgets and make something like this. But honestly, this was also pretty good as well. This is another design that I came up with without using Tailwind. And this is more like an agency website, so I didn't go with that. And then in this design, it ran into a small type error, which I'll quickly fix now. And yeah, this is a design. It looks more like a kind of hacking website almost or like a terminal website. So honestly, I'm not going to go for this either. This is another design not using Tailwind. And for these last four designs, I kind of reminded it that hey, you're like capable of being extremely creative. Use that to your advantage. Um, I think I may go for this if I was building some other kind of application as well, but honestly, I was pretty surprised. So, yeah, it does seem very promising when it comes to design. And honestly, you could have it come up with 10 to 15 different landing pages, maybe based on different images, all inspired by your copy or your brand, and then just cycle through them and pick the one you like the most and iterate on that. And one thing I did find out about it is that it's really good at iterating on the designs that it did make itself. I find that in cases of Claude and GPT models, they kind of reach a stagnation point when it comes to iterating on a design that I came up with and it doesn't really make any more changes. But I did notice when I was trying to get it to come up with an onboarding for Hyposper and there's no real on boarding for the application right now. So I got Claude Code to come up with an onboarding for me and basically this is the one that I came up with. It's like super basic. Doesn't look nice at all right now. And then I tried to get Gemini to make improvements to it. So this is the first improvement that I made for one theme that I proposed. Honestly, this is better. Like for example in this case, but it's not good as well. So I was like, okay, can you come up with a different design? And then it came up with this. And honestly, this isn't really good either. So, I think I might get it to scrap the onboarding completely that Claude code came up with and then just say like come up with a brand new on boarding uh with these steps because I find that when it comes to iterating on a design that perhaps another model has written, it just does a worse job than if you tell it to come up with a fresh brand new design by itself. But some of this behavior may also be because I'm not prompting it right because people are still learning how to prompt the model effectively. But when coding with Hyper Whisper, I did notice a few more things. Firstly, it's not really good at writing Swift UI related code. For example, I had to go back to cloud code several times to make these corrections because whatever it wrote just did not compile. But it's surprisingly good at fixing bugs that something like Codex CLI could not identify when given an image of it. And I think in some cases overreachingness of the model can be good because for example in this case it identified some redundant code when it was working on a slightly different task and then just cleaned up that code and like simplified it. And honestly, I think this can be pretty good in my own workflow because I noticed when using cloud code or codeci, it likes to write a lot of redundant code in the sense that it writes something new and then does not remove any existing or previous code that's no longer required despite me telling it to do so. And maybe because Gemini 3 Pro does have a bigger context window, meaning that's better able to piece different things together, it can identify redundant code more easily. or maybe it just fixs in a slightly different way and can identify redundant code that other models have created but that itself has not created. So I'll have to see over time that if it can identify and then remove its own redundant code unlike claude and GBT related models can. Anyways, continuing on I do find in some cases that overreachingness and like the side quests that it does do lead me to start a new session because it feels like that session has kind of been compromised by whatever side quest it decided to do. uh and then it's just like easy to have a brand new session. It does write really clean and structured code which is good as well. It is also really great at getting a project started for the first time and coming up with a pretty complete implementation of that project and I think you do notice this in some of the designs that it does make when it comes to these like voxal landscapes and so forth. They are pretty complete on the first attempt. I find that this is better than other models like Claude and GBT models because Claude and GPT do like to add stubes and placeholders instead of actually finishing off the entire implementation when given a pretty big task. And maybe that's because they're trying to squeeze it all in into its context window and Gemini doesn't really worry about that because a context window is so big, it just does a complete implementation first of all. Or maybe that's related to its ability to have like a big picture understanding of the entire codebase. But I found that even when starting a brand new project like the agent stack project that I'm making right now, I never had to tell it like hey this feature is incomplete or like you put a stop or placeholder here. Another thing that I did notice about Gemini 3 Pro is that it kept getting lost in the types of the project and then I had to switch over to composer one for example to fix all the type issues that it was causing or just not being able to get over and I basically suspect that it has some kind of contradictory behavior in some way. So, like Gemini 3 is very creative, and that same creative behavior is like really good when it comes to designing because it does a lot of things design-wise that you don't have to ask it to do, like the smaller details, but that also means that it can be overreaching in some ways, and it ends up doing tasks that you did not intend for it to do, or just like does random fixes, which can be good or annoying sometimes. And maybe you can prompt it better to like prevent it doing that thing. And I'm sure a lot of us over the coming weeks will be learning how to prompt this model better over time because it does take a couple weeks after release for everyone to know how to prompt a model better. And I suspect that it's ability to have like a big picture understanding of the codebase and the big context window also does make it worse at like really small details like fixing a certain type in the codebase. I found that even when I told it to avoid using any types in Typescript, it still continued to use those any types. So ultimately, I think that many large language models these days do have some kind of contradictory behavior or like a double-edged sword. For example, with GT5 and 5. 1 and CEX, it's pretty bad in design, but it's really good at paying attention to smaller details and fixing types. And I think that behavior may be somewhat mutually exclusive. Like it may be hard to have a model that pays attention to types and can resolve those really fine grain detailed like type related problems and also be really creative as well. and also pay attention to bigger picture of the entire project for example at least not for now. So, I think that I will still be continuing to use Gemini 3 Pro, especially for brand new projects because I found that out of all the models, it does the best job at actually getting the project off the ground. And then when it comes to other things like fixing up types, for example, I'll switch back to Composer or GPT 5. 1 or maybe even seeing when cloud code fits into my workflow in this case because now I think that Gemini 3 is finally usable for coding. Because Gemini 2. 5 Pro, which was released about 9 months ago, I don't actually think that anyone was using it in the last couple months for many things coding related because other models were just significantly better. But now, since Gemini 3 Pro is really good at actually making a fullyfledged project and getting you through the entire process, like I did with agent stack, for example, I think that many more people will be using it more for fullyfledged projects. But I also saw this earlier today where someone from the Google deeper mind team shared some of their best practices for using Gemini free or just prompting it well. And one of the interesting things about this, it will be linked down below is a multimodal coherence. So text, images, audio or video should all be treated as equal inputs. Instructions should reference specific modalities clearly to ensure the model synthesizes across them rather than analyzing them in isolation. So I think what this means practically is that if you share another modality then you have to reference something from the other modality clearly otherwise a model may just like implement it as it is and I kind of have seen this behavior myself because what I did is for hypersper there is basically a support model. So the support page kind of looks like this. And I wanted a similar support page implemented for agent stack for this. And I basically took a screenshot of this page. And then I gave it to uh Gemini 3. And it implemented the exact same thing with the exact same colors and icons and everything like that as well. So yeah, I thought that was pretty interesting. Like you may want to be careful when it comes to actually giving Gemini free images, for example. And when you do, you want to reference what exactly you want for an image. Otherwise, it may just implement it as it is. I thought my prompting was pretty clear. I think I got into a habit where I write prompts a way that Claude GPT understands, but I have to change my prompting style a little bit when it comes to Gemini free. Anyways, that is basically my vibe check of the model. If you do want to buy my application, Hyprosper, there's a Black Friday sale going on right now and there will be a coupon code down below. And of course, if you do want to get access to private beta of agent stack, then the website will be linked down below once I have deployed it. And then you can just contact me by pressing the like button over here. I will continue to work on this particular project with Gemini free pro more to understand how good the model is and learn some best practices for using it and then also integrate some harder features as well by integrating live kit and voice agents so you have like AI agents on your website that can help you sell your products better and provide support over audio instead. But yeah, I will be sharing more my progress with this application in future videos as well. And of course, if you do want to learn more about vibe coding and how to make applications like the one that I'm making, then I share a lot to do with that in my AI startup school, including how to monetize your applications across all these classes over here. A bunch of people have already joined and have seen pretty good success with their own applications as well. There will be a link down below if you are interested in joining.