Vibe Coding in Custom Software Development: What You Can Expect From It
What founders, CTOs, and product leaders should realistically expect from AI-driven development.
Several years have passed since AI declared itself and started changing the software development market. There has been a lot of hype around it. We got a new phrase — “vibe coding” — that is usually shipped with a variety of emotions.
The emotions about vibe coding in software development can be very different. I have seen hidden pain in posts written by software developers. I have also seen and keep seeing excitement, admiration, and hope in write-ups made by non-technical people.
The hope is that AI and vibe coding will break the wall between ordinary people and software development and make the coding world available to everyone.
Is this a real possibility today or in the future?
Let’s sort out what AI is, how it works, what it can do, and what it cannot do.
Let’s also see what project owners can expect from vibe coding and how they can use it.
Key Takeaways
- Vibe coding lowers the barrier to software creation, but it does not replace engineering thinking or responsibility.
- AI-generated code imitates developer output well, yet lacks understanding of context, purpose, and long-term impact.
- Vibe coding works best for prototyping and experimentation, not for unattended production systems.
- Sustainable software still requires human oversight, verification, and engineering experience.
What Role AI Played in the 1990s
Let me start with my background. I entered university in 1996, graduated in 2001, and got my master’s degree in software engineering in 2002.
The interesting part is that my university was called the University of Artificial Intelligence. My speciality was named “Intellectual Systems of Decision Making.”
That’s right, it was 1996, before the World Wide Web reached every home.
We studied various systems referred to as artificial intelligence. These systems were built on special knowledge bases that included production rules, frameworks, semantic networks, and — you guessed it — neural networks.
At that time, neural networks were considered a structure that had already reached its limits.
Still, we studied them. In 1998, I trained a couple of my own Rosenblatt’s perceptrons in C++ and achieved much better results in character recognition using a Hopfield’s network. Both types of neural networks had been known since 1956.
What Happened Next?
In the mid-2010s, the computational power of workstations increased dramatically. Suddenly, it became possible to increase the amount of neurons and connections among them and teach the neural network to do new things.
Neural networks started recognizing shapes on images. They began producing content that looked far more “human” than ever before.
Eventually, we got specialized language models that could take into consideration the context of a conversation. Backed by massive knowledge bases, these models could participate in conversations and create an impression that they were almost human.
What Is Inside an Artificial Intelligence Model
If you look into the AIs behind Gemini, ChatGPT, and similar tools, you will find the good old neural network, well known since the 1950s. Just much bigger, better trained, and less transparent.
Almost like a human, right? Well…
Companies do not disclose detailed specifics of their language models, but publicly available information suggests that modern language models have on the order of hundreds of millions of neurons dedicated to language processing.
This is huge, but worth comparing to the human brain, which has around 86 billion neurons.
Compared to that, modern AIs are closer to a hamster that does not run, climb, or eat because it uses its entire brain to speak human language.
But it is not just the number of neurons that matters.
Speaking about what makes the human brain powerful, there is also another number — the number of synapses — the connections between neurons. Those 86 billion neurons can connect and talk to one another.
While communicating, they exchange information via electrical signals (similar to artificial neural networks) or complex chemicals that are way more sophisticated than electric signals, albeit slower.
Overall, the complexity of the human brain is too large to replicate using today’s CPUs.
Still, we can build language models and teach them to do certain things very well.
What AI Does and What AI Does Not Do
So what does AI in general, and language models in particular, do well? In simple words, they are good at imitating human behavior.
The key word here is “imitate.” To pretend. To “act as” a human.
Imagine good university students who have to pass a test in a subject they are not familiar with at all. And the test is close. What will they do?
They will spend a few days and nights reading guides and primers, trying to squeeze a massive amount of new information into their heads. They hope to absorb enough material to give answers that sound reasonably close to real ones. To make an impression that they know what they are talking about.
Do they really know the material? No. Do they gain core fundamentals of the subject? No.
Their goal is simply to be convincing enough to pass the test and get the best grade possible in their circumstances.
This is what AI does. This is exactly what it is designed to do: to give answers that are close to what a domain specialist would provide.
When applied to vibe coding in software development, this means the AI should provide a code snippet or a commit that is very close to what a software developer would produce. To generate a commit, AI needs:
- to be trained on a specific platform, and
- to have context, which includes knowledge of the current codebase, the assignment, and general coding guidelines.
If it has those, a diligent AI code generation can produce a very decent commit.
What AI Does Not Do
Limitations of AI сoding:
- It does not challenge requirements (like a real developer would).
- It does not make sense out of requirements — meaning it does not ask “why” and “for what purpose” while trying to improve or even change the requirements.
- It does not run the code, does not test it, does not debug it, and does not assess whether the update makes the project better (in the meantime, even junior developers do all this).
- It certainly does not suggest a different engineering approach to the task.
Also, you cannot expect AI to invent a cure for cancer, find the meaning of life, or convert to Christianity.
It is not designed to do that. It is designed to be a helper.
And helpers are meant to be controlled by humans who know what they are doing.
To make the difference clearer, here is a high-level comparison of what human developers do versus what AI actually does in vibe coding scenarios.
Human Developers vs AI Vibe Coding
Understanding requirements
Interprets requirements, asks clarifying questions, challenges assumptions
Follows instructions literally, does not ask “why” or question intent
Business context awareness
Understands business goals, trade-offs, and long-term impact
Has no real understanding of business value or strategy
Code generation
Writes code intentionally, based on experience and context
Generates code statistically similar to existing examples
Architecture decisions
Designs systems, evaluates alternatives, anticipates scaling issues
Does not design architecture unless explicitly instructed
Testing & debugging
Runs code, writes tests, debugs issues, validates behavior
Does not test or run code unless guided step by step
Code quality assessment
Evaluates maintainability, readability, and long-term cost
Cannot judge whether the code is “good” or “bad”
Handling edge cases
Identifies edge cases based on experience
Often misses edge cases unless explicitly prompted
Responsibility & ownership
Takes responsibility for outcomes and decisions
Has no ownership or accountability
Adaptability
Learns, adapts, and improves based on real-world feedback
Retrains only indirectly through new data and prompts
Role in a project
Engineer, problem-solver, decision-maker
Assistant, accelerator, and productivity tool
Vibe coding works best when AI accelerates what humans already understand, not when it replaces engineering judgment.
How Software Developers Use AI
AI can be very effective as a helper. Currently, there are three modes how software developers can use AI:
- Asking questions and receiving answers plus indicative code snippets outside the codebase
- Using inline and auto-complete suggestions while writing code
- Giving assignments to AI agents and receiving pre-baked commits
We have been using these tools for some time now, and can confirm that AI-assisted software development has evolved. It works.
It still does not “understand” a bigger meaning or business ideas behind the assignments, does not engineer things, and does not care whether the resulting code actually works.
In copilot mode, it showed a complete misunderstanding of my intentions so many times.
But with good guidance, it can save significant time by generating large chunks of code that are reasonably good for you to iron them out and make them right.
At the same time, I would not recommend relying on AI and thinking that it can write better code, or suggest better engineering than you.
You need to understand the AI better than the AI understands you.
The AI coding tool Cursor provides clear and detailed explanations why you need to have this understanding and why you cannot avoid having it. It also explains why AI can lose its mind and how to apply a developer’s “verification mindset” to fix it.
Without your control and comprehension, code written by AI becomes a brand-new legacy in your project. This often requires a professional code audit to understand its quality, risks, and next steps.
AI just tries to help and keep up with your style. Sometimes I wonder how it would work in an old-fashioned PHP project with five thousands files 600-line long each and without tests.
Would it maintain the style?
Just asking. Never had a chance to test it. But chances are, it would.
How a Non-Technical Person Can Use AI
If you are not a technical person, you are not familiar with all the specifics I have gone through in the last paragraph. You do not have a “verification mindset”.
You just want a tool that takes an assignment provided in human language and gives you a working application. This is what many non-technical people building apps with AI expect.
You are promised that AI can do it.
Vibe coding tools like Loveable or Bolt offer a UI that can build something from scratch — with a frontend, a database, and deployment to their cloud in one go.
Can you use them?
Well… Possibly.
Real Stories Behind AI-Built Apps
MenuGen
MenuGen is an experiment in building a startup almost entirely through vibe coding. Its creator used AI to generate code, iterate on features, and move fast from idea to a working product. The project was developed and released publicly, with the author documenting the process and lessons learned along the way.
Despite the initial promise of rapid progress, the development quickly involved challenges typical for real-world software: payment integration (Stripe), API keys management, configuration issues, and deployment details.
If you read the story the way I have read it, you will feel the author’s pain. He got involved in details he never planned to touch. Not because the idea was flawed, but because production software inevitably demands engineering decisions beyond code generation.
Apart from that, his profile on X says that he was Director of AI at Tesla, so he is clearly not a non-technical person. He understands software development well; he just deliberately chose not to be involved in startup development “in the conventional sense” out of curiosity.
As a result, the experiment was not a “pure” non-technical vibe coding case. But it is still a good example of where AI accelerates progress and where it reaches its limits.
Here is the full story behind MenuGen for those who want to dive deeper.
Ollie: Family Meal Planner
This story is perhaps a better example of how AI tools can be used for app prototyping by a non-technical person. While being on maternity leave, an HR professional taught herself to build an AI-assisted web app. She was able to move from idea to a working prototype and publicly share her experience.
The outcome was successful in the sense that a working application was built. However, when read carefully, the story also shows that the process required more than simply describing an idea to an AI.
The author:
- attended AI coding classes,
- had a husband and a brother who helped her along the way,
- learned new skills after multiple attempts to make Cursor do what she wanted.
Her motivation was clear: she did not want to “fall behind” professionally while on maternity leave. She wanted to strengthen her position in the job market and did not want to lose her job to AI.
Here is the full story if you want to explore the details.
There are things we can learn here.
AI Lowers the Bar, Not the Effort
Yes, it is possible for a non-technical person to learn how AI works and use it for app prototyping or even for something bigger.
But it is work. Effort. A challenge.
Pure vibe coding is not what you think it is. Coding can give you a kind of vibe you may not be ready for.
You may have to break some walls and acquire new skills — which was also an option before AI, by the way. I had students who came to me with no prior experience and became professional software engineers. That all happened before AI became a big deal.
AI will make walls lower, of course. But people who use it still need to learn certain things about how web or mobile software works, otherwise they will not get very far.
Because at the end of the day, software still has a codebase. It still uses libraries and packages and depends on the ecosystem. AI only helps generate code for a selected framework.
But at some point, you will need to support it. And in many real cases, this turns into a software project takeover by an experienced development team.
I have a perfect example from one of my customers. He tried to use Loveable to build a real estate SaaS and wrote a set of prompts. I saw these prompts. He was stuck in the UI and did not even get to the point where he would ask Loveable to generate the main feature.
I am always keen to encourage people to learn new things and acquire new skills. If you are thinking about trying AI for prototyping, go for it.
Just be aware that you may need to learn a few new tricks.
And if you hit a wall and need help, we are around.
Is AI Really Intelligent?
Let’s summarize the above. What is AI?
Is AI a miracle?
Well, yes, but in the sense every new human-engineered technology is. From a practical standpoint, AI became possible because structures invented in the 1950s and called neural networks met high computational speeds and were backed by huge data centers.
Is AI intelligent?
Again, yes, but not in a philosophical sense. Every human-like operation is intelligent. Chess computers were intelligent in the 1980s and 1990s. When we use computers to run certain algorithms that are designed to reproduce human-like thinking, these algorithms are intelligent.
However, neural nets go beyond that. They are intelligent in the sense that they produce unpredictable results. Nobody knows exactly how these results are calculated — only the general principles. We know the number of neurons and we can train the network using a specific data set. But after it is trained, it becomes sort of a black box.
Much like a human brain, isn’t it? Often, when I talk to people, I do not know how their brains function and why they think what they think.
Also, people and AI show a similar reluctant behavior when you try to convince them.
There is a reason for this reluctant behavior, and the reason is already-established connections among neurons. To change them, you need to feed listeners a really big portion of new verifiable information and make sure they are attentive enough.
With real people, you also need to wait while their neurons exchange complex chemicals and grow new synapses. Scientists say, the average time for this activity is three months — be sure to repeat the new information several times to them during that time.
The downside of that approach is that humans will grow new synapses in three months no matter what. And if they find alternative information during this time, you may need to start your teaching session all over again.
Can AI Steal Your Job?
This is an oversimplified question.
It depends on your job and its position in the market. The job market changes constantly, and many factors have a bigger impact on the market than AI.
The biggest one is money.
Ten years ago the market was booming, it was oversaturated with money. The Global Hype common to the Western economy was invested in startups. Everybody wanted a startup and needed outsourced developers.
As a result, there was an insane number of developers entering IT, including many incompetent ones.
Today, when the Global Hype is reinvested in AI in general — and AI in custom software development in particular — outsourced developers struggle due to lack of work. This is not because of AI itself, but because there is less money flowing.
What Will Happen Next?
Disclaimer: personal opinion ahead.
My personal understanding is that in a couple of years everybody will get used to AI, master the tool, and start using it on a daily basis. Some activities will become simpler.
But not everyone will gain the understanding of how AI works and get the ability to maintain complex projects.
Custom projects will still require real people with eighty-six billion neurons — even though not all these eighty six billion are dedicated to software development.
The Global Hype however will move in another direction. It always moves. It cannot stay in one place for more than five years straight for some reason.
Who knows what the next globally praised thing will be?
AI will stay and be with us as a limited yet effective tool which can make many aspects of software development easier.
Professionals will use it, but they will still have to be trained to use their brain without AI.
Because AI does not engineer things.
It only helps people to more effectively do what they themselves are capable of.
FAQs about Vibe Coding in Software Development
What are the main risks of relying on AI-generated code for a startup or business application?
The biggest risks include hidden security vulnerabilities, inconsistent code quality, lack of documentation, and growing technical debt. AI does not validate business logic, challenge requirements, or assess long-term impact, which means issues often surface later, when fixing them becomes more expensive and time-consuming.
What should founders or product teams do when a vibe-coded project stops scaling or becomes hard to maintain?
When an AI-generated project reaches its limits, the next step is usually to review the code quality, identify structural issues, and decide whether refactoring, modernization, or a full handover to experienced engineers is needed. Treating AI-generated code as a starting point, not a finished product, makes this transition much smoother.
Vadim Kondratiev, team lead and customer liaison at Krononsoft, specializes in guiding complex software initiatives and aligning technology with business growth.
