Opinions continue to proliferate:
My experience has been more positive, and I think perhaps the difference might be the architecture, data model, and constraints you give the tools. More on this in the future …
I mostly agree with Alexandru here. While most bug fixes, scripts, unit tests, and smaller projects can be entirely vibe coded into existence with little guilt, larger projects demand a more involved programmer keeping the AI agent on track. If you cannot take full responsibility for what the AI generated (at least read and smell its output), you are vibe coding at your own peril. AI hasn’t made the mistakes good programmers have made. They haven’t felt the pain and regret. They aren’t getting paid to do the work.
Claude Opus 5.5 is a very impressive, capable model, but it misses the mark from time to time and excels even more when guided into sound architectural decisions and the like.
AI agents will continue to get smarter. I hope they become superhuman, but I will be quite sad if the code they write is lacking elegance and sound architectural design. In the future, I want to be able to read an AI generated program and marvel at its beauty and simplicity – not stare at complexity that “magically” works.
This is interesting:
Had this similarly with some prospects regarding dev projects where they’ve vibe-coded a neat-looking but bloated dashboard CRUD app thing with charts and animations but then are asking me to help with the last “few little bits” like architecture, security, database connections, compliance, reliability, general support and to “make us really confident in the application”.
I’ve started calling it as it is - “you have done the cheap and easy bit and ultimately have made a prototype”.
An image from the above article:
