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.
Here be lions, since these things tend to lie to you or themselves when skipping the enforcement part.
If the software was capable of creating high complexity software I, or anyone, should be able to replace and build anything I want with just some prompts and guidance.
Anyone who has worked with LLM heavily knows this is far from the truth.
> I should be able to […] build anything I want with just some prompts and guidance.
…Yes? You can. That’s why some people are excited. The fact that agents can do a thing does not mean that you don’t (currently) need skill and taste to guide it in the right direction. No one said it was a magic button. Just a new form of cognition. And god knows I don’t expect all the human cognition I encounter in my life to be the image of perfection…Same goes for AI slop comments. I can just prompt AI into more or less writing like me.
> 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.
The people building those specifications and the tight validation loops you describe are the technical software workers of tomorrow.
Best, friend.
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.
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".
For sure, the guys who lost out when automobiles displaced horses were the ones who were certain they only knew how to make buggy whips.
Adapt and thrive.
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.
It's not that I don't think those skills are important, but they betray a fundamental misunderstanding of what SWE is, which is about managing systems, juggling complexity, and having some wisdom about how those things should fit together.
Hopefully we start to better formalize some more of that now that the grunt work can be taken out of the equation a bit. And teach it in schools, find ways to measure it in interviews. etc
currently working on an initial product for a startup and while I'm substantially further along that I would be without LLMs, I'm at the point where the LLMs are now randomly breaking things.
There's typically a combination of issues that are occurring when the codebase goes off the rails.
One is that cruft builds up. Large vibecoded codebases need regular "review cleaning". Where you review the whole thing top down for hygiene, contradictions, bad comments, style, redundancy, design. It's a standard thing for me now. A full review and cleanup pass once every couple months.
The second thing that tends to go wrong is that after a certain complexity threshold, the models cannot keep track of the relevant details for a task very well. They lose sense of project directionality. Every task gets disconnected from other tasks and they start pulling the system apart at the seams, exacerbating the cruft issue above and accelerating collapse.
I've found that separating my documentation into three distinct lobes helps: goaling docs, guardrail docs, and implementation docs. Goaling docs describe in prose the goals of the project which includes scope, certain design targets, and other things that you are aiming towards. Guardrails are for design and implementation choices you have explicitly ruled out, the "what not to do" part. Impl docs describe the state of the system as it is.
The top-level agent documents refer to these and make their purpose explicit.
You'd think a ticketing system might capture this, but it doesn't really provide the richness of expression a collection of documents in a directory does.
These approaches aren't perfect. The process still has a tendency to wander off the rails, but it happens less often now and I catch it more quickly. Every session misses little bits of the goaling alignment, but they also find and address small issues missed by previous sessions. Overall it does a better job of keeping the codebase hovering around some threshold of coherency.
I'm having a lot of fun to be honest. Yes it's bizarre and nothing like what development used to be, but it finally feels like there's some new things to learn now.
These days they know how to handle their own context in review. So I will use that to find improvements too in retrospectivea and a small usage script hook it runs each push that alerts if high token usage. Hint the subagents are getting excessive context, which is also good hint of drift. So then I can measure when it starts happening typically via that script that Claude checks and readjusts when one goes over budget.
However it will be similar to cybersecurity in that there will be no 'entry level' product engineer. It will probably have requirements having worked in customer service or finance or some other adjacent field to get a larger picture of what you are delivering. You will also not need 3000 of these individuals anymore (did we ever really??)
I can see a world where you go from "AI Shepard Customer Success rep" -> some middle role with increasing 'product' responsibilities -> Product Engineer.
The new roles will also probably be more generic - instead of backend and frontend you have 'Backend' (Infra + API) and 'Frontend' (Frontend + Design) and both of these roles will require genuine product thinking
What we the people need to do is gain control and hold all the power of AI in our hands - not the filthy rich's hands! Devise systems that flip the power into our hands and create economies and jobs around it (i have ideas and surely 1000s to more do to around this).
No, and it's the same with leads, they can be given a system they understand but need to reference it.
I'm very certain these are people brushing up against their lack of processes, workflows, or engineering practices.
The context are large enough. You do not need to understand the world to build a chair.
I see the very opposite in many different domains happening now, and some codebases are monsterous. They are figuring out ways that do indeed work well.
The engineers that are figuring out these problems when working with a new tool, will have jobs.
Those who continue to doubt, and distance themselves are the ones that will be behind.
I think it is mastake to lay this completely at the feet of ICs. What you are describing is a form of engineering, and a lot of management feels that "engineering is solved" (not just "coding") and have no patience for these things. ICs are being forced into extremely naive and unsophisticated AI usage by managers with incorrect mental models. That's why so many ICs are cranky.
Its out there, it can do things, and it has its limitations. What it can do gets lot of light, and what it can't do well becomes a ICs 'skill issue'. Say with AI your velocity can go up by factor or X, but expectation is Y where Y >> X. This leads to lot AI-usage which is nothing close to "engineering". The whole idea that engineer should be able to defend what they build, while showing productivity gains that are factor Y can't be done without cutting corners, and it looks like leadership is mostly ok with it as long as it not immediately causing problems.
This is again not problem with AI essentially, but the overly optimistic expectations from it.
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.
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.
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.
Then after i understand it i can make a targeted investigations for that issue and after that it will be obvious for the AI, and all of those we still didn't change even one line in the code.
After that, we fix that issue (which is the root of the cause), release and congratulation the bug was fixed this way works even in lowest powerful models.
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.
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.)
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.