like most developers these days I use AI agents as part of my workflow, but I've never tried hands-off vibecoding an app from scratch. I have been learning to sing for half a decade, and take a braille display to lessons since it's easier to read lyrics than it is to have them spoken through an earpiece for example. It is pretty much my only use for braille these days. I've found though that lyrics on the web are often not very well formatted for line-by-line scrolling on a display, and navigating an app interface as someone who is not as proficient in braille as I once was can be a challenge.
A perfect excuse, then, to try vibecoding an app. Lyrx searches for lyrics using any of a few free public APIs and formats the output to suit line-by-line reading, with configurable character limits. It's built for MacOS and iOS with a native interface for each platform. It was built with GPT6 Astra in Codex, on a pro plan, set to the extra high effort level.
I did cheat a bit by specifying Swift 6 with strict concurrency, strict adherance to Apple's API and human interface guidelines, a few keyboard shortcuts, and an interface overview in the initial prompt. Other than that, I gave the model no architectural guidance, I did not specify the APIs to integrate for searches, and every other decision was left to the model. The full app took 3 prompts to build; 1 initial prompt did most of the work, and I refined the interface a bit in 2 prompts. The code is on Github, here.
Why am I publishing this? Firstly it might be useful to someone, though there is no pre-compiled build. If you want to use it, you'll need a Mac with XCode to build it yourself from the repo, which any agent will help with.
But I also think it's interesting since the vast majority of the vibecoded projects we see here are built with Claude. I have never liked Claude. I find it generates over-engineered code, and I don't like the way it interacts. It's too agreeable, too keen to please, and it doesn't check itself very well. In my experience GPT is better at following instructions, writes cleaner code without as much baggage, doesn't constantly tell me I'm right, and is better at staying focused in longer tasks or projects.
Code wise, is it perfect? no. But it's honestly not as bad as I thought it would be. It has demonstrated to me that having a strong set of agent rules and a good initial prompt can get you quite far. The repo is open for anyone to take a look at, so if you are a dev and are so inclined feel free to critique. To me it was an interesting experiment and I got a small personal use app out of it that I was going to build anyway.
Comments
gpt
I thought about buying claud but the more and more slop I'm hearing about it, the more and more I think gpt is way better with this kind of stuff. I'd like to try claud to compare but that's not my top priority right now and codex can get ya pretty much what ya want anyways.
GPT for me also.
It just works, what else can I say?
Another thumbs up
I use GPT with the plus subscription. As a retired software developer, corporate trainer, and tech writer, I have tons of software experience. And I am very impressed with what ChatGPT can produce.
Three prompts? That's amazing. I typically have an extended chat with ChatGPT to develop a firm grasp of what I want to develop. Then I have ChatGPT produce a prompt that I paste into Codex running in Terminal. Codex produces a first implementation from that prompt, but there are always revisions, so I end up with dozens of prompts before I get exactly what I want.
When the code has bugs, they're often the same types of bugs that humans produce, such as the familiar off-by-one error. So ChatGPT doesn't feel like a step backwards. Rather, it feels like an efficient and accomplished developer that responds well to guidance and instructions. I feel more like a manager, with ChatGPT and Codex as my development team.
In the future, blind persons will look back on this era as the greatest revolution in accessibility since the invention of the cane.
bug fixing
Yep it just works. I haven't really tried with complex stuff but what I'm building right now I can tell ya that when I found bugs, I'd usually mostly described them in simple programming terms it'll find and fix and usually 99 percent of the time it's right on the money but only three prompts? Shit.
gpt here as well
I have not tried claud, however, I wouldn’t say that it’s claud that is slop, or a particular AI. That is just a tool, nothing more, nothing less.The what, or I should say Who, makes it produce slop? is the person guiding it. If they don’t direct it properly, if they don’t check it’s work? in other words proofread it to the best of one's ability, insure a consistent project road map like ok first this than x then y with clear objectives / goals etc? In other words, verify its work weather its code, writing etc? yeah, no doy it’s will likely produce what people call slop. But then, again, people already by themselves have produced slop long before AI, this is nothing new. But as said, I wouldn't say its, oh claud slop! Its people who use it that can make it produce slop because they don't direct it as they should, of course that experience will come the more familiar one gets with AI but yeah.
Agree on the bug thing, if you do your best to describe bugs well, What’s happening, what should be the expected behavior, etc.? The fixes for said bugs can be pretty quick, yes.
With that being said, yes, AI can produce bad code, but so can a person. However, how you direct it, your project roadmap, to have everything structured, matters just as much as the code itself, if not more so. But yeah AI can mess up on its own yes, but you can minimize it, and if, when, it messes up, direct it to fix it or to the correct expected behavior, but its a complex topic for sure.
Re: gpt here as well
I think Slop is probably too harsh a term, and I don't doubt that Claude can produce workable code. Some (perhaps most) of the best known vibecoded projects, such as the recent open source recreation of Fusion 360, are written using Claude. It is all in the direction. I just find that as a working partner, claude requires a lot more babysitting and frustrates me with its written output more than GPT does. I just thought it was an interesting comparison, since we don't see as much real code from GPT whereas most open source repos have Claude as a contributor now.
+1 for gpt
my website, https://rlreserves.com, bar the data, was completely made with codex. for the first week, it was made with clord, however i didn't like the direction it was taking the site
Re: bug fixing
4 prompts now, as i had it fix a couple of small architectural things today that I just couldn't leave alone. But otherwise yes, 3 prompts to finish the whole app. To be honest ,if I wasn't curious to see what the model would do on its own, I would have started with a better prompt from the start with more guidance, and would likely have built the app in 2 prompts, maybe even 1. The importance of a good set of rules in agents.md and a good initial prompt cannot be overstated. With those things in place, vibecoders would probably cut most of their development time in half, especially when developping apps in languages other than Python, C or Javascript.
i,
with codex, i also ported over svox piko to the mac, and was even able to make it be able to be a voiceover voice.
I did get a bit over ambitious though, thinking I could port over a 2021 build of google TTS. I had to reinstall my Mac-OS after that it broke speach everywhere. No I don't regret it even if it almost gave me a heart attack.
Another GPT github project
Here's a relatively simple utility for MacOS to set mouse pointer to Ignores. It works around a MacOS accessibility issue that was reported ages ago and never addressed.
Because the utility doesn't use a public API, I used ChatGPT for extensive research and testing before implementation. After ChatGPT and I had working code, ChatGPT walked me through publishing the code through github.
No more waiting for giant corporations to address accessibility issues. With vibe-coded utilities like this,, accessibility is now in our hands.