Thursday, June 19, 2025
VibeCoding: How AI Revolutionized the Way I Develop
My confession as a software architect: after 15 years of .NET development, from blockchain to cross-platform apps, I thought I had hit a plateau in my productivity. An entrepreneur at heart, passionate about new technology and a musician on the side, I have always looked for innovation in how I work.
The arrival of AI coding tools changed everything. Between the marketing promises and the reality on the ground, I wanted to test it for myself.
The result? My development velocity was multiplied by 4.
▸ My turning point: when the investment pays off
Let us be honest: a $200/month subscription for an AI tool gives you pause. With roughly 900 messages every 5 hours, or 200 to 800 coding prompts in that same window, it has to be worth it.
First concrete test: I generated in 2 hours what would have taken me a full day. But be careful, it was not magic. I had to iterate, correct, refine. The tool accelerates; it does not do everything on its own.
▸ VibeCoding: discovering a new kind of collaboration
VibeCoding is this new dance between human and AI. You describe your vision, the tool generates code, and you orchestrate the whole thing. But beware the illusion: it takes sharper expertise to validate and correct, not less.
The reality: we spend 80% of our time reading code against 20% writing it. With AI, that proportion flips, but the quality of your critical reading becomes crucial.
A concrete example: refactoring a 300-line class inherited from a legacy project. The AI proposed a base in 15 minutes, but I spent another 45 minutes reviewing, adjusting the architecture and validating the tests. Total: 1 hour instead of 4 to 5, but a long way from a one-click miracle.
▸ Why .NET and strong typing change the game
As a solutions architect specialised in the Microsoft ecosystem, I quickly understood the advantage: strongly typed languages like C# are better suited to AI-assisted coding.
The practical benefit: the tool can read JetBrains Rider's warnings and correct itself. That kills some of the round trips on compilation errors, but not all of them. You still have to supervise.
A recent example: on a complex API with Entity Framework, the tool detected a relational mapping problem and proposed a fix. Handy, but I had to check that it respected my business conventions, which the AI cannot guess.
▸ My VibeCoding workflow: realistic collaboration
The principle: every modified file becomes a validation checkpoint. This is not magic generation, it is collaborative work where the human stays in charge.
My daily routine:
Describe the need: “Create an authentication service with JWT, role management, and integrated logging”
AI generation: the tool analyses my architecture and proposes an implementation
Critical review: I validate, correct, suggest adjustments. A crucial step!
Iteration: we refine together until we get something satisfying
Structured commit: detailed messages, sometimes amended several times
The truth: the tool understands my existing patterns, but you have to teach it your specific conventions. And sometimes its “improvements” are not relevant in my context.
FormCraft: a case study
Want to see the result? FormCraft, my NuGet library developed with this approach. Clean architecture, complete tests, detailed documentation, but also a lot of review hours to make sure the whole thing stayed coherent and maintainable.
▸ The full ecosystem: tests, docs and architecture
My discovery: AI coding works better when you take a holistic approach. Do not generate only the business code!
My protocol:
Tests at every level: the tool knows my frameworks (xUnit, NUnit) and generates baseline scenarios. But careful, you often have to enrich the complex business test cases yourself.
Documentation: README, XML comments. That documentation becomes crucial for keeping coherence between sessions.
Architecture: SOLID, Clean Architecture. The tool follows the patterns, but defining the overall vision is on you.
Field report: on a refactoring project, I generated 200+ unit tests in one afternoon. Those tests revealed 3 hidden bugs. But I also had to delete or correct 30% of the generated tests, which were testing implementation details rather than business behaviour.
▸ The revival of legacy projects
The ultimate test: modernising 10-year-old code. That is where AI shows its limits, and its strengths.
A concrete case: migrating a .NET Framework 4.5 project to .NET 8. The tool analysed the architecture, identified the obsolete dependencies, proposed a migration plan. Impressive on paper.
The reality: it handled the technical aspects well (async/await, new C# patterns) but I had to redo all the specific business logic myself. Work that would have taken me 3 weeks turned into 1 week of collaboration plus 1 week of corrections and validation.
▸ The real stake: becoming a value creator again
My observation as an entrepreneur: AI does not replace our expertise, it repositions it. Less time on implementation, more on design and validation.
Before: “How do I optimise this LINQ query?”
Now: “Does this feature actually answer the user's need?”
The concrete change: I spend roughly 50% of my time thinking about architecture and requirements (against 30% before), 30% validating and correcting generated code, and 20% manually coding the critical parts.
Business impact: on my last project, I delivered 2 weeks early with better quality. But I also spent more time upfront clarifying the specs. AI forces you to be more precise in what you ask for.
▸ The train is leaving: get on board or stay on the platform
My conviction: we are living through a pivotal moment. Developers who master these tools today will get ahead, but this is not an overnight revolution either.
To the sceptics: I understand the reservations. This is neither laziness nor taking the easy road. It is a new tool that demands new skills: knowing how to prompt, critique, validate, correct.
My advice: start small. Try it on a side project. You will find it is both more and less magical than the stories suggest. More powerful than expected on some aspects, more limited on others.
VibeCoding is an evolution of the craft, not its end. Like the move to IDEs, to frameworks, to the cloud, every generation of developers adapts to new tools.
This time, it is collaborative AI. And even though it is not perfect, it is already transformative.
What about you, what is your experience with AI in development? Have you tried it? What is your feedback, positive as well as negative?
Does this resonate with your team?
Let's talk about how Atypical Consulting can help you move forward.
Contact me