AI Writes Laravel Code That Looks Good… But Fails in Production

Hello guys, today I want to show you a few examples of AI generated code that actually fails, but not in a vocal way. You will not even notice those bugs if you don't double check and if you don't think about potential failures. But before I show you my code and my example which I vip coded intentionally just to see where models fail, I want to get back to Laracon EU recent conference where one of the speakers Yanik had this topic. AI won't fail loudly. It will fail quietly and I agree so strongly while using AI. I notice it so often. So this is one of the screenshot of Yanik talk with example. I will zoom that in. Probably a bad quality screenshot but still a controller which seems okay on the surface but it has thing like ignores query pattern ignores boot method for extra data set like created by it could be in the boot method also uses input instead of request validated also has if statement on some impossible edge case also doesn't use proper policies which are already in place so it overrides the logic and stuff like that but On the surface, the code looks okay. And as a result, one of the points of Yanik was this slide. AI won't break your system. It will kind of work, but it will slowly reshape it if you allow it to. Now, let me show you my example. So, I asked Claude code us 4.6 good model to generate restaurant reservation project with filament admin panel based on a real job from Upwork, just a bit trimmed down in terms of scope, tech stack, overview, front end. This was the original plan of menu pickup ordering. This seems to be a good plan with quite a lot of details and quite a few functionality points described which I then transformed into I like to work in project phases. So then I can tell cloud code to work on phase one, work on phase two and so on. So the first phase is about database structure, then filament and so on. So I think it's a pretty good plan and actually the result was pretty good. So you can for example reserve a table. This is a public form. If we use big filler chrome extension, the time slot can be chosen based on the date. Then you reserve the table. Then it's all working well it seems. Then there's filament admin panel with orders and reservations. So when clicking around I noticed that it kind of all works. But let me show you a few things where AI generated code was let's put it this way not ideal. First, let's take a look at order controller store and this is powering order pickup here. Add to cart and then proceed to checkout and then you fill in the form to pick up the order later. So, you choose pickup date. This is what I want to show you. And select the time and then you place the order without payment. It assumes that the payment will happen in live mode or afterwards. This is outside of scope of this project. And in the store method of that controller, so when placing the order, it's pretty complex. So we check for example that some items may be not available anymore that the quantity of the items may be not available anymore and then we have database transaction which is a good thing to create the order to create the order items and then reserve inventory. It's a method in the order model to increase the reserved quantity on specific menu item in the database. So on the surface it looks good for concurrency to have DB transaction. So order and items and reservations come as one transaction which is good. But actually the things beforehand the check for item availability and the stock happen way before that transaction start. So before we start the transaction of the order we calculate the total we generate the order number we even process the file. So by the time that transaction starts this part may be already obsolete. So a better solution probably would be to move these parts inside of the DB transaction check first and then within the same transaction process the order. So yeah this is one example where AI generated code could be improved. Another example actually in the same controller we can go to store order request and look at the validation for pickup time. So in that form we have time slots available for when that restaurant is opened and it may be based on a date specifically but then there's no validation on the back end at all. So the only actual validation happens on the front end in the UI which time slots are available or not and also similar thing is in reservation controller that I already showed you another page. So this is on the homepage reserve a table. So same here you select a time slot you can do that even without the date actually this is another thing for UI improvement but then again in the reservation controller we have store and in the form request is just required and string and yes some of you may ask what was the prompt for that thing maybe it was not in the prompt and indeed it wasn't for the record I can admit so reservation module is described with this with just form request validation So it's not specified and emphasized that there should be validation for time slots. So this is technically not a failure on the AI agent part. This is a failure in the specification. But as we all know specifications from the client are never detailed enough and even technical specification usually doesn't contain all the potential validations up front. But now in 2026, it's more important than ever to have detail specifications with details like this like validation because AI is just following your instructions mostly. So if you don't specify something, it may be omitted and skipped. The next thing I want to emphasize is filament login form and why it would not work on production if I deploy that project to live because in filament in the official documentation we can see this. So to access filament in non-local environments. So if in your env you have app environment production or something like that you must have this can access panel. So AI generated code with testing login with seed user works locally. But in the generated user model there's nothing about filament at all. No can access panel no filament user. So that system is not ready for production again. Maybe it shouldn't right away and also it was not specified in the plan in the specification phases. But look what happens if in myv I change that to production. Imagine we deploy that project already. And now if I try to sign in on production seemingly I would get 403 forbidden without any explanation. So for maybe not so experienced developer or even VIP coder the project just worked locally but then we deploy it and then suddenly it is broken. So this is another thing that AI may miss something that would work locally but not necessarily in production. And also finally don't expect AI to write code with best practices like don't repeat yourself solid and stuff like that. So for example, I have this approve and cancel links for orders which are also available inside of the view page. Also approve and cancel as buttons. Do you think that AI noticed that the actions are repeatable? So here's the filament code for orders table. We have action make approve with a lot of things and then the actual action is here. Same with cancel with conditions when it should be visible with action as well. LDB transaction. So quite a complex set of operations. And then in the view page we have action make approve with almost the same identical things. Some details here and there are different like how record is defined. So for example this record and stuff like that but most of that thing is totally repeatable and could be offloaded to some service instead with filament layer on top like a button or icon or something like that. So yeah, this is another thing I noticed by AI generated code. Do not expect that to be well structured architecturally in the future because in many cases AI is doing like page by page, function by function and not rechecking with the function that was similar just a few iterations of the prompt ago. Of course, that could be audited and fixed later by another AI agent or the same AI agent. In fact, there was a video on my channel recently with a command called simplify or skill or command whatever you call that by claude code which basically does exactly that review of the code to be reused called quality review and stuff like that. So another kind of piece of advice is if you are creating something like in vibe coding style or something with like less interaction on your behalf let's put it this way then run the check through at least one more another agent. Personally I often use codecs on top of cloud code generated code but in this project I intentionally did not do that in order to shoot this video with what I saw from the first version of the code generated by cloud code. So yeah, these are my observations where AI may not go deep enough, but also I thought kind of overarching thing. AI in this case is similar to a junior developer in a way. So junior developers may totally not think about those edge cases, extra validation stuff in production. So it does remind of a developer with the goal to just deliver working code for now without thinking about what happens when the code goes to production. I think you understand what I mean. We probably all went through that phases of our career where our goal was to deliver something to meet the deadlines to please the client right now. That was our goal. And then as you get more senior, you start thinking deeper about project reliability, code maintainability and all those fancy phrases which will mean that the project and the business is stable for the long run and not just for this version. So from that angle again, AI reminds me of a junior developer. What do you guys think about those specific cases I showed in this video? Or maybe you have other patterns that you noticed where AI is weaker. Let's discuss all of that as usual in the comments below. And if you want to follow my other AI coding experiments, you may also follow my second channel or third channel. I also have filament daily but there's also AI coding daily already pretty popular where I also use Laravel a lot but more focus on AI coding tools how they work and how to optimize the workflow working with AI agents. So yeah subscribe to that channel as well and see you guys in other videos.