Before I begin to even start writing about this topic I assume it makes sense to mention that this article has not been created using AI. Not a single word. (with one exception: The german version has been translated automatically using Claude, like all entries on my blog) I also tried to write this for a wider audience than my usual articles. As this isn't a specific problem solving for a weird use case in c# I tried to keep the tech gibberish down and especially do not expect you to have C# knowledge. Any coding experience helps, but isn't required. On the other hand if you are a skilled dev, don't be annoyed if I explain things trivial in your opinion; Just ignore the explanations and feel smart :)

Ruleset

I simply wanted to test how good vibe coding (Letting the AI create your whole application) is today. Does it achieve what I want? Is the code quality good? Would it help someone who has no knowledge at all? So I set up a ruleset:

• Make sure that what I want to achieve hasn't be done before (so the AI can't just copy/paste a complete application)
• DO NOT CODE: Really. I try to not change a single character in a source file even if a mistake is obvious
• Try not to help: Of course if something is wrong, I will tell the AI. But let it find out how to solve this, first only helping if it got really stuck
• Application should not be too ease (no "Hello World")
• Application should not be too complecated either (no "Please create a better Microsoft Teams application")
• No external tools needed; So no Unity or Photoshop requirements, etc. I should not need to leave the IDE (Rider)
• No Security related stuff: No Password hashing, not storing important secrets, simply because I trust AI enough to do this lightly. This might be worth another experiment on its own

Environment and Job description

I used Rider with Claude AI integration using Sonnet 4. Whenever it asked for a permission I granted it. I started by giving a simple task and iterate over time. The task was:

Create a Razor Application that offers a SCRUM Poker functionality with a twist: It should have Joker Mechanics like in Balatro. The PO should be able to create a new game. Then an admin view and a player view should be created. Each of them by generating a random code like play?code=dbddhkp and admin?code=ukule2. No database. Just store everything in memory

I was sure this hasn't been done before because it absolutely makes no sense. SCRUM poker is a tool to determine the "cost" (or complexity) of programming tasks. Each developer decides how complex the task at hand is usually using values from 1 to 21 (as always, there are a lot of variants with other numbers) and at the end these values gets compared and discussed. The main reason is that the number itself isn't communicated openly at the beginning so if a senior dev tells his number other devs might just say "yepp. About the same". The "poker" element makes sure everyone says the number he or she really thinks of and afterwards any discrepancies gets discussed.

Balatro on the other hand is a fun Poker game. There are Jokers that change the value of the cards. E.G a "every heart gets multiplied by 4". Obviously this would destroy any useful results that the SCRUM poker produced. So a nonsens project, a quite complex one nonetheless.

I added additional features lateron like a multi-language feature but did want to give the AI the chance to iterate

Did it work?

The short answer is: Yes. I did not write a single line of code and the application worked.

But the long answer is: Developers wont loose their job very soon. As I will show in the learnings it requires a skilled developer to lead into the right direction when problems need to be solved. In addition some decision lead to bad code. The detailed pitfalls and also positive surprises can be found in Learnings.

(And also I published the code on GitHub so you can see for yourself)

The biggest benefit: I am a backend developer. My programs (usually) work, but I am not really good in UI or UX. That benefit the AI was able to solve. The application wasn't the most beautiful thing ever to be seen in a browser window; Still: It looked way better than anything I would have created

Sources and demonstration

You can find the sources on the FiboNuts Repository on Github. A working example can be found at Poker.Oles.Cloud

Learnings

Code quality

Code quality started good, but got worse, completely wrong assumptions leading to bad code. Forcing Claude to revert these things helped putting the code back in good shape. TBH I expected way worse

Help needed

Quite at the beginning of the project I ran into a problem that could not be solved by the AI itself: There are URLs generated for the players. But when I enter those in a different browser (or even tab) the code does not get recognized. I tried to tell Claude in about 10 Iterations that it does not work. The AI did try different approaches but never came near a solution. After I told him that there is a problem in the Dependency Injection implementation and the class holding the state should not be added Scoped but added as a Singleton it was able to continue. But without my help this would be the end of the project.

It was surprising to see that such quite simple cause could not be recognized while much more complex issues got resolved.

Also at one point the communication between the server and client did not work and it seems like Claude simply could not believe when I said that. I had to offer assistance telling him to create logs that I can then paste back so Claude can analyze the issue.

Adding random stuff

Claude also decided at one point to add https. (It was http before) when I was asking to add a multi language feature. I assume he learned implementing that feature from a site where https was used. The https-part completly broke the application because ports had been assigned multiple times in the config.

Also until the very end there always was an ugly "About" link on the upper part of the page linking to https://learn.microsoft.com/en-us/aspnet/core/?view=aspnetcore-9.0. I assume this is part of any "how to create a Blazor Application" Tutorial and so Claude thought this is required

Wrong assumptions

At one point I asked the AI to add a multi language feature using best practices. While it did that correctly it forgot to add a crucial Package from NuGet. The Build errors have been quite obvious stating 'Are you missing a reference' but Claude did not seem to understand that. Instead it assumed that the code generated was too complex for the compiler (it absolutely wasn't: Just a few lines adding language-strings to an array) and simply removed its code over and over again until the program compiled again but also had no language code left. Again I had to help and tell him: "Hey. That wasn't the issue. You forgot some NuGet Packages". After that the issue was resolved.

At some other point Claude managed to break the application completely. Right at the start it simply would not load anymore after the progress bar was finished. Claude did an assumption no dev would do: It was sure the cause was the logging. Too many log entries causing the application to slow down. So any log in any exception handling or anywhere else got removed. Not solving the issue, obviously but creating a way worse variant than before.

As a rule of thumb: Whenever the Bots writes something like "this is a known issue" it probably found some very weird edge case that has nothing to do with a problem that should be simple. In my case multi language not working caused really weird actions: Creating dlls for each language, trying to use reflection to fix some (non-existing) issue of the localizer when at the end the text in the razor page was simply static. After I told that there should be variables or method calls instead of fixed strings Claude was able to revert all the nonsense changes, but without that help this was unsolvable.

What is considered easy?

Tasks that I see as quite complicated have been solved sometimes without any issues while others, that shouldn't be too hard seemed to be unsolvable. For example I wanted a '/metrics' endpoint to be used with Prometheus (A tool to collect metrics and display them in a Dashboard). While it looked right, it was a html page that rendered the content. Claude did never manage to follow the simple task: "please output plain text. No html tags." Even when providing the correct Mimetype it still did not understand correctly. An additional Server application was created where the clients sent metrics, suddenly a IIS Express configuration appeared in the launchsettings and other things that made the program way, way worse but did not even move an inch into the right direction.

After that, Claude was so confused that not even rolling back worked. The AI wasn't able to find its files anymore and even tried to find it in regions of my computer where it isn't supposed to look:

Needless to say that these operation have been costly:

After a few retries I had to surrender. Claude was never able to create a simple output in text/plain. Even if I gave example snippets. It always rendered HTML that looked like Text but still had html tags.

Interesting tactics

After breaking the application when adding https claude did some interesting things. It did not simply analyze the code but was sure it wasn't its fault :) Instead it assumed there must be some other application causing the issue and so started nmap to analyze what other services on my machine might be to blame. It did not expect it to "break out" of my IDE. A reminder to not just allow everything but have a close look on all actions. I would not be surprised if it had decided to simply kill that other process if it really had been in external issue. (But the implementation explicitly asks for your permission whenever it does something it hasn't before)

Conclusion/Opinion

Useful, but needs to be monitored closely. And you must be aware that the code isn't really "good", just working. Just a minor issue in this application but when customer data is processed and stored bad design decisions can lead to big problems. There are already a lot of vibe coded applications, used by millions that are just really badly written and leak data all the time. Just one example of many: https://cybernews.com/security/ai-girlfriend-app-leak-exposes-400k-users/

No developer wants to be responsible for a catastrophe like this even before thinking of GDPR fines.

Also: Even more like in human developing: Use Source Control. Claude can create git commits without any problems. This helps rolling back to previous state if the AI created a mess after an error. But I would suggest to only use a local repository and push manually to be able to make sure no secrets end up in the code. Also don't just check the current code but any commit if there was some secret at any time.

Nonetheless we also can't simply ignore AI when creating programs. Even if the AI bubble at some day bursts the tools will remain. LLMs will still be there and be just any other tool on the Cloud or maybe even on local machines. No matter if someone makes big cash selling AI tools or not. So in my opinion a senior dev has to decide where it works and where it does not. Also because of the nature of AI : I will get better, but there will always be the same kind of errors you can't get rid of. Hallucinations are a thing and always will be.

I also think the next generation(s) will be the ones that benefit the most but also suffer the most because of AI. Benefit because it never was so easy to start. Need something done in your OS? Ask the AI to write a python script. Having trouble about Dependy Injection? The AI will provide a document with good examples. You could find that without AI, but that is just simpler and faster. On the negative side: Your problem gets solved, but maybe not in the best way. And this isn't just an academic argument: You have to understand why we do not store passwords but use hashes instead to understand the underlying risk.

And in addition from now on not much more content will be created. You get every information you need through AI. But when no new information is created on the internet in a few years AI will mostly scrape content it created itself leading to a perfect endless confirmation bias loop.

And on a personal note: On the one hand: AI is a great tool, on the other hand it nearly already destroyed any use of searching the web. Have you ever tried to find information on "Is it ok if my dog eats stuff"? The first two search pages mostly contain AI generated content. And these range from "No problem. stuff is even very healthy" to "Never ever! One bite and your dog is dead!". I got around this by using DuckDuckGo and the "Block all results from this site" - feature for a while but am at a point where sites with random domains get created faster than I cen add them to the list.

Maybe I should use AI to keep these lists up to date.... 😀