I find this baffling tbh as I regularly ask Claude for basic components and they come out completely broken, wrong and buggy.

The last: I asked for a quick TCP server in C++ that handled just a single client (disconnecting the existing client when a new client connected), with a send() that I could call from another thread. It was holding mutexes over read(), and trying to set the SO_REUSEPORT port socket option on a socket that had already been bound. Subtley broken garbage.

It would literally be better to copy and paste a solution off Stack Overflow, because at least there's a chance it'd have been reviewed by someone who knows what they're doing.

> a large financial application

Those could mean anything. Some people think 5k likes is large. Others think 100k is small.

Oh well, at least they didn't say "complex".

Also, they didn't say how accurate or secure it is.

Good one! :)

Since its proprietary, it runs in a private cloud environment and is processing only data of one "user per instance"; there is not public interface, only a VPN you have to dial in etc., so no frontend/frontpage facing the public.

Though, there are some design flaws from this perspective, because of convenience: E.g. it lets you persist the account number in the DB, if wanted.

"large" in a sense of functionality for the given/required usecase.

LOC is currently around 200k, so for sure: Its not Microsoft-scale :-D

i think its all about caring and knowing what you want to make and willing to iterate on the result until it is actually good. If you want the ai to do your job for you its probably not going to work, but if youre really good at using its advantages you almost certainly will be winning

You are simply doing it wrong.

I could do this TCP server in another time at all and it would be perfect. I have done stuff that complicated and more many times.

You need to rethink how you are using the tool because you absolutely could get excellent results like I do.

The biggest things I suggest are... Treat it as collaboration or pair programming. Make sure to work through a design before programming and have it written to a file for your review before execution.

You can do this.

this is an abysmal level of condescension, the kind that makes me wish mods actively moderated it, as it's that insulting

That doesn't sound like it would actually save me any time, or produce better results than just doing it myself.

What is this app and what does it do? Can we see it?

I find it very hard to believe anyone could code anything complicated with Claude that 5-6 competent developers could do.

I am currently working on a relatively complicated UI on an internal tool and Claude constantly just breaks it. I tried asking it to build it step by step, adding each functionality I need piece by piece. But the code it eventually got was complete garbage. Each new feature it added would break an existing one. It was averse to refactoring the code to make it easier to add future features. I tried to point it in the right direction and it still failed.

It got to the point where I took a copy of the code, cut it back to basics and just wrote it myself. I basically halved the amount of code it wrote, added a couple of extra features and it was human readable. And if I started with this, it would have took less time!

Its a financial asset management system, and its for proprietary use only. Maybe Im doing some YT insights in the future.

> I find it very hard to believe anyone could code anything complicated with Claude that 5-6 competent developers could do. <

I should have put a disclaimer - Im not layman, instead 25y+ IT experience. Without my prior experience, I think this project wouldnt have come into existence.

I had trouble in my early days with the quality of things I made.

One of the things I found helped a lot is building on top of a well-structured stack. Make yourself a scaffold. Make sure it is exactly how you like your code structured, etc. Work with Claude to document the things you like about it (I call mine polyArch2.md).

The scaffold will serve as a seed crystal. The document will serve as a contract. You will get much better results.

Are you using this large financial application just for yourself?

I think the difficult task is/will be to sell vibe coded software from the lone developer to anyone.

Hyper-individualized software is what LLMs are best for IMHO. They lower the bar so much that it's becoming perfectly feasible and reasonable to amase a large amount of software which is fit to your exact personal needs and preferences.

Yeah, I have a dozen random tools that do specific things I need that wouldn't be useful to anyone else, and that I wouldn't share in their current state anyway. But they're fine for me, and without LLMs, I wouldn't have spent the time to build them.

Exactly: The app can do only one thing, and does this one thing very well. Other approaches are not implemented or planned. Its like a specific dentist-tool that you need for one specific task.

Yes, its for proprietary use among friends; its not for sale, instead I get a cut of their returns for providing support & maintenance.

It is not 100% vibe code, by far not! I use cloud for method-by-method or simple class instructions and integrate in the app manually, I do not use any of the API integrations, I just use the standrad WebUI for discussing, planing & implementation.

Would you have needed 6 people? I find that Claude, Codex etc. are able to output so much because they do a lot of reinventing the wheel whereas a human, given constraints, would make much more pragmatic choices around which technology to use. That’s not necessarily a bad thing, and regardless, you’ve been able to achieve something you’re happy with, which is what matters. But, I’d still like to hear more about what it has done that you think you couldn’t have done in a year yourself by choosing existing technologies. E.g: what is novel in your application? What background do you have?

These models making bad / tasteless decisions about what dependencies to pull in is one of the main reasons they work best in the hands of experienced developers. You've got to know what you want it to use and tell it, and anticipate what shortcuts it will want to take and tell it not to do those things. For these reasons we're not yet at the point where inexperienced non-programmers can get high quality software out of these tools. I do think this will improve with time though.

You need to tell the model what to do and what not to do: The dependency thingy is an issue, yes, but you can tell the model not to do so - and you should always know which result the prompt should create: For sure you must be able to read/understand/judge the code - completely fire and forget is not possible (to my experience), though I see many people saying "I had one mega prompt and after 2 days the app was ready", I take those always with a grain of salt.

Absolutely. These models still need a lot of this sort of hand holding, so they work best in experienced hands. I'm also skeptical of those very long runs, letting it go so long without active oversight must surely produce at least some objectionable design or implementation details, right? So I guess the people claiming those sort of results have less care for these sort of qualities.

Yes, even Claude Opus 4.6 is still running into accidents on longer chats which lasts for 3 - 4 days. But its getting better and better.

The explanation is simple:

a) Speed - it included a lot of boring stuff, esp. in the beginning on when I was in the discovery phase and had to figure out some basics relevant for the context

b) I think I would have given up very early on, esp. of all these boring things, which are required but take long headache time to develop (e.g. The app has a somewhat complex data rendering component, containing hundres of GDI+ calls; the file is currently around 5000 lines, writing this by hand woul have taken very long and would have been very frustrating)

c) Debugging - Sometimes bugs are so deep down in some components and after 1h you stop seeing the forrest because of all single trees: The LLM can greatly help here

d) Fresh ideas: If there is a pyramid of know how in this niche, then Im currently working in "the first floor", basicly; discussions with the model about enhancing and more complex things helps to see the next island where you could swim

Yes, I could have done it without the models - but it would have taken sooo much much more time, that I wouldnt have taken the route.

Novelity: The app does one specific thing and is designed only for that specific usecase - I do not know how novel it is but since its a niche, maybe you could achieve the same thing with with existing solutions and their plugins (but then I would have had to learn how to edit/change those)

Background: 25y+ IT experience, Master degree and some other certs

Can you provide a link to this app? Or alternately, share a few of the prompts by which you built it? I only ask because, if it's really that easy/simple, I'd like to do the same thing!

Its for proprietary use only.

Regarding prompts: a) In general I clean up the work space on a regular base, so I do not store prompts

b) Overall, Id say so far above 200 - 300 initial prompts for the code developed with the LLM (and then 2 - 50[?] follow up prompts to change & update things)

c) The initial prompts are always long and very elaborative, like 60-70% of screensize

d) The model is always aware of the source files used for a given prompts (in Claude you can create project workspaces and put your stuff in)

e) I always tell the model the current state, where want to go and which steps are necessary according to my opinion and I specify the result as detailled as possible

f) I give contraints in the prompts, telling what not to do etc.