Why Developers Should Be Worried About Vibe Coding
AI product development is the loudest conversation in software right now, and one of the least understood. Everyone says they are building an AI product. Very few can explain what that actually means.
In this blog of The Seezoned Crew, Lucas Luengo, Director of Software Architecture at Seez, cuts through it. He explains what LLM development involves day to day, why AI coding tools like ChatGPT, Claude, GitHub Copilot, Grok, Llama and TabNine still need a human reviewing every line, how frameworks like LangChain connect models to real systems, and what actually breaks when an app scales overnight. He also makes the case that vibe coding is quietly damaging junior developer careers, and explains why prompt engineering is a smaller skill than most people think.
When people say they're building an AI product, what does that actually mean? How is building AI different from building regular software?
First you have to know whether they mean it, or whether they are selling smoke. When it is actually AI, most of the time we are talking about parsing some fuzzy text, some raw input, and getting valuable information out of it. That is the short answer. There is more complexity underneath, working with RAG, doing base matching, categorization, that sort of thing.
Right now, about 99% of the AI we use in business is LLM based. There is another type of AI, but it is mostly used in research, things like the serious neural networks that are genuinely useful. Image, video and generative AI are a whole other world again. And machine learning is the step we do not see.
So the way to think about it is this. An LLM is a base tool, and you can build amazing things on top of it. But that depends. It is up to you, and up to the company doing the work.
Do you think AI products require a completely different mindset for developers, or can normal engineering skills carry over?
Most of it carries over. But you need to know that you are working with a fuzzy technology, and you are not going to get the same result every time. It is not like accessing a database where you always get the same rows back. There is a randomness involved in AI. That is where the typical error comes from. Someone complains that the bot answered X or Y and that is not what they were expecting. No, you are not getting exactly the same thing every time, and anything you build on top has to be designed with that in mind.
There are also a lot of new tools to learn, and LangChain is one of them. It lets you plug into outside services and connectors, officially called “tools”, that work together with the LLM. So in the middle of answering a question, it can go and fetch information from an external service and drop that result straight into the response.
What's one nightmare you've faced that taught you a lesson you'll never forget?
There is one problem that is very particular to AI. It usually starts when someone comes to you and says, I want this bot to answer in this way.
Maybe you cannot force it, because you do not have full control of the way the LLM is going to assemble the answer. You try to force it with prompting, tweaking here and there, changing values. But there is no way to completely get the result that you expect. Not always. That is a nightmare.
It is a back and forth, dealing with people who do not understand how this works and convincing them that it is fuzzy. So the lesson is that there needs to be a certain amount of education, not just for us but for everyone around the product, so people know what to expect. Call it management of expectations.
If a product suddenly went viral overnight, what would break first, the servers, the UI, or something else?
That depends a lot on your architecture, and on your weakest block in the chain. If you did things right, it is going to be your database. It will not support the load. That is the most likely scenario, and it is easy to fix. You increase the database size and you are back on track. If you did not use serverless infrastructure, maybe your API breaks down, gets throttled, and you cannot manage the requests. Usually the UI and UX is not a big problem, it is just reflecting what is going on in the backend.
And it is not always about going viral. Sometimes you have a spike because of holidays or during a promotion and you are hitting a limit you did not expect to. Then there are weirder things, like deadlocks, where two people try to access the same asset at the same time and block each other. You do not see that until very deep into the process, and it is alarming.
How do you personally define vibe coding, and where do you think it fits in the current software landscape?
Vibe coding is describing what you want to an AI tool and taking the code it gives you. And I am very concerned about it. I do not know how to say it, but people are not using it responsibly. That is the feeling I get. In most cases it is a very powerful tool and it is going to save you a lot of time, but you need to know what you are doing.
It is not only killing products, in the sense that you are pushing code that is breaking stuff. It is also killing careers, because it is getting people complacent. If you use it properly, if you review what you are getting from your vibe code, you are going to get there faster and it is a super useful tool.
This is where it gets risky for juniors. A junior is never going to stop being a junior if he relies on vibe coding. They need to push themselves to get past that stage. But a lot of this comes down to the manager too. Give juniors simple tasks they can handle, the kind that might be long or repetitive but are not technically hard. That way they build real skills and learn step by step, instead of leaning on vibe coding to get through the work.
One last thing about these tools. They tend to come up with the same response every time. Even following your prompt, the knowledge they have is kind of the same. It is like asking the same developer to do a million apps. The core concepts are going to be the same, so differentiation becomes much harder.
If AI can generate the code, is the real skill now knowing what to ask instead of what to write?
Prompting is a new skill, definitely, but an easy one in comparison. It is almost like Googling. Knowing how to search for information is knowing how to ask for information, and it is very similar in that sense. You learn it the hard way, trying again and again until you fine tune your skills.
But it does not replace knowing how to code. Prompting is more forgiving, because if you are vague the model just fills in the rest and still gives you something back. Code does not work that way. Nothing is assumed for you, so you have to know exactly what you want and account for every detail yourself. That is why knowing how to code still matters. It helps you write sharper prompts, and it lets you actually understand the code that comes back. The newer tools remember your earlier chats and carry some context along, but the moment you start fresh, you are explaining everything again.
Do not trust the code just because it works. You have to go in, read it, understand what it is doing, and check that it is really what you want. That is the difference between AI making you faster and AI making you lazy. When you can read the result and confirm the solution makes sense, it genuinely adds to what you can do. When you go in blindly, it just makes you lazy.
We're seeing people build full apps through vibe coding in days, do you think someone could actually sell one for millions?
That may happen. But that does not mean the product they sell for millions is going to be worth millions. It is going to be a product that is sellable.
Maybe it is a prototype, something that shows a good idea in action, and as a way to prototype an idea it is very valuable. But I do not think we are at the point where you can go into production, into a service that is massive and worth millions.
On scalability, you can prompt your way through some of it, but not all. There is a whole layer around the code itself, things like managing servers, testing and deployment, that vibe coding does not really handle yet, though that is probably coming soon. Security and authentication are actually fine. The gap is more about how you deploy your code and what kind of servers you run, and today that sits mostly outside of vibe coding.
There is also a speed trap here. AI becomes a crutch if all you care about is moving fast. If you actually want to innovate, you have to take your time, even with AI, otherwise you just end up building the same thing as everyone else.