Is there any real high quality SAAS created with vibe coding? Or are they always full of bugs?

[RussianBroski]

What are your experiences regarding to SAAS fully created with AI tools or vibe coding? Are they really stable and trustworthy or eventually they just glitch and shows a lower quality aspect ?

 

[Brigada88]

Base44, Claude Code for Check and ChatGPT for Solution

 

[Panther28]

It depends on what is being built and how many tests they have.



This is like any app. I'd rather have a well tested ai app, than a poorly tested real app, with just as easily some bugs.

 

[Nikolaos]

What are your experiences regarding to SAAS fully created with AI tools or vibe coding? Are they really stable and trustworthy or eventually they just glitch and shows a lower quality aspect ?

It depends on your build. Base44 is usually good, but if you run into technical problems you'd have to know some coding to progress.

 

[flackz]

It's impossible to build an app using vibe coding. The neural network can indeed write code from scratch. But it will be boilerplate code that doesn’t even come close to meeting all your actual requirements. The problems will start when it begins making the changes you need. The code will fall apart immediately because an LLM isn’t a programmer. It doesn’t even know how to think.

All it can do is write statistically continue a given text. As a result, every iteration of changes you make will add some kind of workaround that solves the specific problem you've specified, while simultaneously turning the rest of the code into a minefield. It will inevitably lose some things and rewrite others using methods that are incompatible with the main code it wrote in the very first iteration.

The only thing LLMs are useful for is writing individual functions, small classes, and so on, strictly in accordance with the technical parameters specified by the programmer. If you're not a programmer and aren't able to conduct a detailed review of every change it makes, don’t even try to do this.

 

[RussianBroski]

It's impossible to build an app using vibe coding. The neural network can indeed write code from scratch. But it will be boilerplate code that doesn’t even come close to meeting all your actual requirements. The problems will start when it begins making the changes you need. The code will fall apart immediately because an LLM isn’t a programmer. It doesn’t even know how to think.

All it can do is write statistically continue a given text. As a result, every iteration of changes you make will add some kind of workaround that solves the specific problem you've specified, while simultaneously turning the rest of the code into a minefield. It will inevitably lose some things and rewrite others using methods that are incompatible with the main code it wrote in the very first iteration.

The only thing LLMs are useful for is writing individual functions, small classes, and so on, strictly in accordance with the technical parameters specified by the programmer. If you're not a programmer and aren't able to conduct a detailed review of every change it makes, don’t even try to do this.

Well, it's not impossible, there are many AI apps out there, maybe you are talking about good ones

 

[floqal]

Ours was built with heavy AI assistance rather than by prompting alone, and it's live with paying users, so quality is possible.

Two files each defined a db() helper and PHP refused to load. A stray closing brace broke an analytics cron job and nobody noticed for four months. An unguarded top-level event listener on an element only some users have threw, and that one throw silently killed an eleven thousand line block, so every const below it stayed uninitialized while the hoisted functions kept working.

Defaults bite too. A retry loop kept its five attempts, hit a refused proxy six times a minute and pushed a worker into out-of-memory restarts until we cut it to two tries with a five minute cooldown.

What helped wasn't reviewing harder. One card rendered with no CSS because its class name was invented from the label instead of copied. Now a script reads the live page, finds an element with the same label, copies its classes verbatim, and refuses to run without a match. The stable ones are the ones where every mistake turned into a check.

Best,
Floqal

 

[NaughtyByNature]

To me they are even higher in terms of quality than SaaS built by cheap ass devs that don't give a shit about your project

 

[akwhv]

It depends on the testing. AI-generated code can be solid, but it's not hands-off.

 

[FistFullOfCash]

You have to give it a lot of feedback, if it messes up it can throw back a project a month or more if not backed-up & saved.
I have an eye for things, the troubleshooting has to be done in features by rigorous testing. Most people coding with AI are too lazy to learn to code, so they probably do not want to put in as detailed of work as a programmer who is trained. I have 60+ lifetime software, custom systems setup (some free of cost) and at the core a select bundle of necessary subscriptions (hosting/domains/VPS*/Cnstnt Contact). With a basic understanding of some things and your own logic its possible to mold a nice software suite even. Need to know API/Scrapers or automations and it helps to know different computer things because you will be using PS terminal and entering file paths over and over likely.

Right now, I have scrapers that can pull from public sources aka free info and sell, all within TOS of the source(s). Seems though the AI had a hard time with putting this all together so it sparks up correct. Which is why Im having to update the app soon. Anyways, I explained a process to the AI to pull with scrapers and input into master sheet that adds regularly, outputs to software user. Sounds simple, though this is my only problem feature (keeps giving placement/dummy data on some buttons), will find a solution though!

 

[flackz]

Well, it's not impossible, there are many AI apps out there, maybe you are talking about good ones

I'm talking about software. Specifically, I'm talking about software that works, is secure, and is supported.

 

[alex_zen]

I don't think the problem is AI itself. It's whether someone reviews and improves the code afterwards AI can build a solid MVP, but production- quality SaaS still benefits from human testing and refinement

 

[INCC]

It's impossible to build an app using vibe coding. The neural network can indeed write code from scratch. But it will be boilerplate code that doesn’t even come close to meeting all your actual requirements. The problems will start when it begins making the changes you need. The code will fall apart immediately because an LLM isn’t a programmer. It doesn’t even know how to think.

100% true.

Listen to what @flackz says. Otherwise, you will learn the hard way, as I did.

A very negative experience.

 

[jeezabull]

You will still find bugs in both backend and frontend nomatter what you do. It's good at fixing them. Sometimes switching between AI's helps too. However one shotting isn't really a thing if that's what you are asking about.

 

[FlyingGains]

Yes but use Fable 5

 

[flackz]

I don't think the problem is AI itself. It's whether someone reviews and improves the code afterwards AI can build a solid MVP, but production- quality SaaS still benefits from human testing and refinement

This is precisely the problem with the Transformers architecture. All the existing language models can only generate random text with a high statistical probability based on a training dataset from GitHub, Stack Overflow, etc. This process has nothing to do with programming. The language model simply selects random words that it "thinks" should continue your prompt. For example, when you ask, "What is the capital of Germany?" the most likely answer will be "Berlin." When a language model sees the prompt "Write a loop in C," it will respond with something like "Here's a simple for loop in C that..."

Do you really think that such a system can really create a fully functional application when it doesn't understand what an application is or how it works in the real world?

So, currently, auto-blogging is the only industry where the use of language models feels natural.

 

[KeywordOrbit]

Fable 5 or nothing.

 

[Wraithed]

both sides are right in a way. if you try to build some massive complex software with zero tech skills just by "vibing", you are going to end up with a giant plate of spaghetti code that breaks the first time an api updates. but for micro saas, scrapers, or simple utility tools? you can absolutely get a working product to market. the trick is breaking it down into tiny modules instead of asking the AI to write the whole app at once. i have a couple of small lead gen tools running that were mostly built by claude, but i still had to spend hours debugging the database connections myself when it got confused. if you don't know how to read code at all, you'll get stuck eventually.

 

[rankmaster36]

Hey I think the bigger question isn't whether a SaaS was vibe-coded, but whether the developer actually tested it properly afterward. A quick prototype can become a serious product if the bugs, security, performance, and UX are handled professionally. The development method matters less than the final product.