I've been building for the past few weeks a fancy text editor for myself to work with AI in a more sustainable and sane way. I don't have much time for this, so I've entirely vibecoded it: I don't look at the code and I've provided only some initial design and engineering decisions, debug hints, etc. on top of natural language feature requests.
For about a week using the strongest Claude models (Fable 5.2 and Opus 5.5) this went great, and I had a minimally functional tool that worked very nicely. It seems that I hit some sort of complexity ceiling for those models though. The second week, every feature I try adding creates a new type of jank that can (at least by me) be only vaguely characterized. Most recently inserting into a line (even in the middle of the viewport) causes the viewport to jitter about.
Our work will change, we will spend less time on tedium, we will become more ambitious. Barring an architectural paradigm shift we will still need people with strong engineering and design instincts, and even more importantly the ability to write reference code and formal specifications codifying desired behaviors.
When systems get complex, someone needs to know them and know where to burn the tokens. You need to know what events need to be produced because you have specific consumers. You need to know your business invariants and why.
If you’re working on something that isn’t complex, yeah I can see why you’d think you can be replaced. If it’s just a front end? Yeah I can see that. A basic CRUD service? Sure. A business critical domain service with events, data migrations, complex business logic? I don’t see it.
LLMs won’t always guide you to do this correctly so you need to understand how important it is from the start and always enforce it.
It is hard to say how things will develop the coming years, but this sums it up for me today. Almost constantly, LLMs (e.g. Opus 5.5/Sonnet 5.5) make choices that a junior programmer would think is ok, but I know that will create technical debt, subpar designs, etc. Note that I'm not saying that LLMs program like a junior programmer, in many ways they program like senior engineers who have happened to memoize all API docs as well. But they are often missing the larger context or don't understand the time dimension (what will be hard to maintain, what will introduce an API that we can not easily backtrack from, etc). So, I am not surprised that vibecoding breaks down as a project becomes larger.
At the same time, current LLMs make me a lot more productive (which was not true for me ~one year ago). I can do in 30-60 minutes what would take me a day 1-2 years ago. In many ways, programming also became more interesting, because I can focus on design, the really challenging parts, while the LLM can quickly implement it. LLMs really feel like a different, in many ways complementary intelligence for development.
FWIW, I'd suggest that every major feature or change still needs review at the planning stage, at least. If you keep just requesting features without considering about whether the model's solution is the best (or at least, a good) solution, something will eventually break. That's not really surprising.
I think that's the conclusion you were drawing, also?
(In your particular case I'd just roll back to the last unaffected commit and try again, with a mind to how any solution might affect both performance and the viewport. Trying to fix what a model's already broken is a fool's errand.)
My other firsthand observation is seeing the difference between what quality of software that I, as someone with many years of experience, can produce versus my coworker, a talented designer without SW experience who is very AI pilled and excited to be making things, is able to produce.
Through a combo of me and it over about a month we got it working well enough to roll out though. The entire time I had to hand hold the model through the changes though, and it loved to try and throw it away for quick fixes that would break something else.
For example I don't see any mention that you ran debt reduction loops, code simplifications loops, and in general asking the model to clean up the code after it's initial phase of "organic growth".
After about a dozen failed iterations I reverted to the last known working commit (albeit with the new module only partially implemented), isolated the issue as best I could, then asked Gemini again and this time it correctly identified a latent timing bug that was the cause of all the problems.
In short, I had to manually reduce the complexity of the problem for this magic infinity-dimensional-processor to deduce what the problem was. Maybe it would have gotten there eventually, but without AI I would have just reverted the commit a lot sooner and likely figured it out on my own with some basic debug logging.
It's still a net gain, as it's successfully one-shotted coding errors and bugs that might have taken me hours to find before, but it can miss basic problems as well.
LLMs have massively improved the quality of our company's main app. However, someone less skilled at our company tried to vibe code an internal tool and it was pure crap.
But, as @sama said, these models you see today are the worst they'll ever be.
It's only a matter of time before someone figures out how to maintain focus and drive a product to a polished end state over many iterations.