After two years of vibecoding, I'm back to writing by hand [video] | Hacker News Reader

After two years of vibecoding, I'm back to writing by hand [video]

youtube.com

120 pointswritten-beyond134 comments

While I greatly dislike the hype and don't believe most of what people say is real or that whatever they're building is just bullshi, I can definitely see the improvement of productivity, specially when working with agents.

I think the problem is that people:

* see the hype;

* try to replicate the hype;

* it fails miserably;

* they throw everything away;

I'm on call this week on my job, one of the issues was adding a quick validation (verifying the length of a thing was exactly 15). I could have sat and done that but I just spun an agent, told it where it was, told it how to add the change (we always add feature flags to do that), read the code, prompted it to fix a thing and boom, PR is ready. I wrote 3 paragraphs, didn't have to sit and wait for CI or any of the other bullshit to get it done, focused on more important stuff but still got the fix out.

Don't believe the hype but also don't completely discount the tools, they are incredible help and while they will not boost your productivity by 500%, they're amazing.

> I think the problem is that people:

> * see the hype;

> * try to replicate the hype;

> * it fails miserably;

> * they throw everything away;

I'm sure doing two years of vibecoding is is a considerably more sincere attempt than "trying to replicate the hype and failing at it".

The field of "AI assisted programming" is progressing extremely rapidly.

Claude code (arguably the most recent large change, even if it wasn't the first of it's type) was released one year ago.

After watching the video I'd say that it is similar to my own reaction when opening my own code that is 2 years old. (To be clear, code I myself wrote 2 years ago, without AI.) Or even more realistically, code I wrote 6 months ago.

Buy it mostly reads like someone making an exaggerating claim to get s boost from a populist narrative.

Andrej Karpathy posted

>There's a new kind of coding I call "vibe coding", where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.

Just under one year ago.

Language evolves; the term is now used to refer to a range of AI assisted coding. The concept has existed longer than the term.

  > There's a new kind of coding I call "vibe coding", where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.

This is exactly my experience with using Haskell after having couple of beers.

I've started disregarding any AI take that is all-or-nothing. These tools are useful for certain work once you've learned their limitations. Anyone making sweeping claims about vibecoding-style use being viable at scale or making claims that they're completely useless is just picking a side and running with it.

Different outlets tilt different directions. On HN and some other tech websites it's common to find declarations that LLMs are useless from people who tried the free models on ChatGPT (which isn't the coding model) and jumped to conclusions after the first few issues. On LinkedIn it's common to find influencers who used ChatGPT for a couple things at work and are ready to proclaim it's going to handle everything in the future (including writing the text of their LinkedIn post)

The most useful, accurate, and honest LLM information I've gathered comes from spaces where neither extreme prevails. You have to find people who have put in the time and are realistic about what can and cannot be accomplished. That's when you start learning the techniques for using these tools for maximum effect and where to apply them.

>> The most useful, accurate, and honest LLM information I've gathered >> comes from spaces where neither extreme prevails

Do you have any pointers to good (public) spaces like this? Your take sounds reasonable, and so I'm curious to see that middle-ground expression and discussion.

  > That's when you start *learning* the techniques for using these tools for maximum effect and where to apply them.

https://paulgraham.com/popular.html?viewfullsite=1

"They're perfectly justified: the majority of hot new whatevers do turn out to be a waste of time, and eventually go away. By delaying learning VRML, I avoided having to learn it at all."

If you’re counting on LLMs and coding tools to go the way of VRML, I guarantee you’re going to be disappointed.

This one was on HN recently: https://spectrum.ieee.org/ai-coding-degrades

The quality of code generated with LLM decreases over time.

Let us talk again after, say, five years.

You can gamble, but you can't guarantee

> The most useful, accurate, and honest LLM information I've gathered comes from spaces where neither extreme prevails. You have to find people who have put in the time and are realistic about what can and cannot be accomplished. That's when you start learning the techniques for using these tools for maximum effect and where to apply them.

This requires the level of professionalism that 97.56% of SWEs do not have

Agree 100%

> they are incredible help

Yes.

I use LLMs for coding in the exact opposite way as described in the video. The video says that most people start big, then the LLM fails, then they reduce the scope more and more until they're actually doing most of the work while thinking it's all the machine's work.

I use AI in two ways. With Python I ask it to write micro functions and I do all of the general architecture. This saves a lot of time, but I could do without AI if needed be.

But recently I also started making small C utilities that each do exactly one thing and for those, the LLMs write most if not all of the code. I start very small with a tiny proof of concept and iterate over it, adding functionalities here and there until I'm satisfied. I still inspect the code and suggest refactorizations, or putting things into independent, reusable modules for static linking, etc.

But I'm not a C coder and I couldn't make any of these apps without AI.

Since the beginning of the year, I made four of them. The code is probably subpar but they all work great! and never crash, and I use them every day.

I wonder what sort of training data the AI was fed with. It's possible that such if whatever was utilized most was put together into a reference cookbook a human could do most of the work almost as fast based on more normal searches of that data in an overall more efficient way.

What about stack/buffer overflows, use after free and all of the nasty memory alloc/dealloc security pitfalls? These are what I would worry about with C programs.

With respect for your priorities that may be different from mine, it would sadden me a little to always have to work like this, because it would rob me of the chance to think deeply about the problem, perhaps notice analogies with other processes, or more general or more efficient ways to solve the problem, perhaps even why it was necessary at all. Those are the tasty bits of the work, a little something just for me, not for my employer.

Instead my productivity would be optimised in service of my employer, while I still had to work on other things, the more important work you cite. It's not like I get to finish work early and have more leisure time.

And that's not to mention, as discussed in the video, what happens if the code turns out to be buggy later. The AI gets the credit, I get the blame.

"perhaps even why it was necessary at all" not being asked anymore is what I fear as well. Stumbling over problems repeatedly gets attention to architectural defects. Papering over the faults with a non-complaining AI buries the defects even deeper, as the pain of the defects isn't felt anymore.

>the chance to think deeply about the problem, perhaps notice analogies with other processes, or more general or more efficient ways to solve the problem, perhaps even why it was necessary at all. Those are the tasty bits of the work, a little something just for me, not for my employer.

You should be aiming to use AI in a way that the work it does gives you more time to work on these things.

I can see how people could end up in an environment where management expects AI use is expected to simply increase the speed of exactly what you do right now. That's when people expect the automobile to behave like a faster horse. I do not envy people placed in that position. I don't think that is how AI should be used though.

I have been working on test projects using AI. These are projects where there is essentially no penalty for failure, and I can explore the bounds of what they offer. They are no panacea, people will be writing code for a long while yet, but the bounds of their capability are certainly growing. Working on ideas with them I have been able to think more deeply about what the code was doing and what it was should do. Quite often a lot of the deep thinking in programming is gaining a greater understanding of what the problem really is. You can gain a benefit from using AI to ask for a quick solution simply to get a better understanding of why a naive implementation will not work. You don't need to use any of that code at all, but it can easily show you why something is not as simple as it seems at first glance.

I might post a show HN in a bit of a test project I started over the Christmas break. It's a good example of what I mean. I did it in Claude Artifacts instead of using Claude Code just to see how well I can develop something non-trivial in this manner. There have been certainly been periods of frustration trying to get Claude to understand particular points, but some of those areas of confusion came from my presumptions of what the problem was and how it differed to what the problem actually was. That is exactly the insight that you refer to as the tasty bits.

I think there is some adaptation needed to how you feel about the process of working on a solution. When you are stuck on a problem and are trying things that should make it work, the work can absorb you in the process. AI can diminish this, but I think some of that is precisely because it is giving you more time to think about the hard stuff, and that hard stuff is, well, hard.

> The AI gets the credit, I get the blame.

Now that is some stark truth right there.

I don't think that's really true. You still need to think to make what you want to make. You still got to design the program. You just do less typing.

In a sense, AI coding is like using a 3D printer. The machine outputs the final object but you absolutely decides how it will look like, how it will work.

Using AI for coding (or anything) is like being a manager. Successes get passed on to the team (with some to the manager) but the blame stops at the manager

I don’t type code into the editor anymore and I still think as deeply about the problem as I ever did, if not more.

Wait, your job was len(x) == 15?

The hardest part is figuring out how to use leftpad.

The hardest part is checking to see if the version of left pad you just pulled or one of it’s 300 dependencies has a supply chain vuln in it.

quick validation for which you wrote a 3 paragraphs long prompt? sounds like you've wasted time

For plenty of people, writing three paragraphs of prose they're already aware of can easily be <2min of work.

This comment took 15s, typing can be very fast.

I bet writing the code directly could have been even faster, llms aren't magically fast

> llms aren't magically fast

They literally are.

I had to make a small CSS change yesterday. I asked the LLM to do it, which took about 2 min. I also did it myself at the same time just to check and it took me 23 seconds.

Curious what was the validation/diff size?

Was it 3 paragraphs to change a line of code?

I agree the tools are amazing, if you sit back and think about it it’s insane that you can generate code from conversation.

But so far I haven’t found the tools to be amazingly productive for me for small things like that. It’s usually faster for me to just write the thing directly than to have a conversation, and maybe iterate once or twice. Something that takes 5 minutes just turns into 15 minutes and I still need to be in the loop. If I still need to read the code anyway it’s not effective use of my time.

Now what I have found incredibly productive is LLM assisted code completions in read-eval-print loop. That and helping to generate inline documentation.

Final diff is 235 lines added. I could definitely write the code faster but then the write/ci/run tests loop would make it take longer anyway, having the agent write the code and do the cycle based on what i asked it to do ended up being more productive. It obviously wasn't if len(thing) != 15 exception.

This isn't the usecase that's being criticized. Here you are responsible for the fix. You will validate and ship it and if something goes wrong, you will answer for it. The AI won't answer for it.

My stance is that it's similar to hand tools vs power tools.

You need to pick and choose when to use them, it largely depends on the task.

I use AI a lot now for programming and spend more time code reviewing, sometimes though, I use my hands.

My nimber one rule is, don't try to do too much in one prompt.

You have a pragmatic approach, but you don't have a PR unless you have inspected and tested the code. Just submitting code generated by a tool without personally vouching for it's correctness and sanity is not engineering.

Have you watched the video? One of the points made in the video is to come back in two years and read the code base resulted from vibecoding.

Same, it's so good for these little things. And once you start adding rules and context for the "how to add the change/feature flags", etc you get that 3 paragraphs down. Now our head of product is able to fire off small changes instead of asking a dev, making them context switch, etc. Devs still review but the loop is so much shorter.

> didn't have to sit and wait for CI or any of the other bullshit to get it done

You run CI on human-generated PRs, but not AI-generated PRs? Why would there be a difference in policy there?

The agent sat through the CI process, not me.

So you watch the agent progress bar instead of the CI progress bar?

This is just sad to read. AI makes all of us stupider by the day, no matter what the tech bros and wanna-be billionaires of HN tells us...

Multiline autocomplete is still the biggest productivity boost for me. This works well in a familiar codebase with reasonably consistent patterns.

After that it’s the “ask” capability when I need to get oriented in unfamiliar and/or poorly documented code. I can often use the autocomplete pretty effectively once I understand the patterns and naming conventions.

Similarly, agents are good for a first pass triage and plan when troubleshooting tricky bugs.

Still haven’t had a good candidate for going full vibe code. Maybe that’s because I don’t do a lot of greenfield coding outside of work, which seems to be where it shines.

Just my experience. It’s new set of tools in the toolbox, but not always the right one for a given task.

Multiline autocomplete is very disruptive. It's overly verbose, and it constantly pulls me out of my flow, because it's not what I want.

I know what I want before I type it. Having to parse the auto-completion disrupts the thought process of what I _wanted_ to write.

I'm working on a greenfield project right now and my experience has been 100% in line with the video

I think it might be even worse for greenfield work, as that's when you're establishing a lot of patterns. You don't want AI to have any role in that

Yes. And AI is bad at design.

But that's why you tell the AI to refactor.

I've started a greenfield project and went 100% AI for learning purposes (of course it's more like 95%) and my takeaway is:

- it's fully possible

-- but the AI is of no great help with figuring out what the architecture or interfaces should be

- Keep a refactoring backlog

-- Spend 30%-40% of your time on refactoring, aligning patterns, improving architecture

-- depending on your codebase, this can happen in parallel

-- sometimes you need to get your hands dirty and do the cleanup yourself

-- ... but usually, you only need to establish the pattern once

- once the patterns are established, it becomes easy to talk to the AI in the context of your codebase

-- you can reference patterns by name or location

re: your last bullet.

This has been very effective in my experience. “See class foo for example implementation “

I should clarify. I do very little greenfield development, even outside of work. So my understanding of vibe coding being good for this use case is largely rooted in the relayed experience of others.

I think the points about code ownership and responsibility are spot on. Management wants you to increase velocity with these agents so inevitably there is pressure to ship crappy code. You then will be responsible for the code, not the AI. It’s the idea of being a reverse centaur.

I also like the comments on how developers should be frequently reading the entire code and not just the diffs. But again there is probably pressure to speed up and then that practice gets sacrificed.

The rush it out the door mentality of this industry is stupid.

If your Management thinks it's acceptable to increase velocity by shipping crappy code, I'm surprised they never thought to do this before AI.

From what I've seen, companies using AI well are shipping better code because of all the artifacts (supporting context like Architectural Design Records, project-specific skills and agents, etc.) and tests needed to support that. I understand that many are not using AI well.

Of course there’s always been that pressure but now with AI tooling the sheer volume of code that can be generated is massive.

> there is pressure to ship crappy code

this is true with or without AI

Yes but now there is so much more volume of code created in the same amount of time.

"If you're a software developer and you're worried about your job, you haven't spent enough time actually using these AI agents. Anyone who spent eight hours plus a day over the last year using these agents is not at all scared of these agents taking their jobs. They're not... Your job is not going anywhere."

I agree with this take... for now. I wouldn't be surprised if the AI agents improved exponentially (in the next few years) to the point where his statement is no longer true.

Software developers should be worried about their jobs, not because these tools are capable of replacing them or reducing a company’s need for human developers, but rather because the _perception_ that they can/will replace developers is causing a major disruption in hiring practices.

I truly don’t know how this is going to play out. Will the software industry just be a total mess until agents can actually replace developers? Or will companies come to their senses and learn that they still need to hire humans - just humans that know how to use agents to augment their work?

Software development hiring is terrible right now, but hiring has been pretty slow in general. We gained 2 million jobs in 2024 and only 500,000 in 2025.

That can't possibly be a long term disruption. If it doesn't work it doesn't work

If AI can't replace developers, companies can't replace developers with it. They can try — and then they'll be met with the reality. Good or bad

You may be right if AI truly can never replace devs. But history shows many examples of inadequate technologies getting hammered on by industry until they work due to distorted perceptions of economic gain. I’ve heard stories like this regarding mechanical looms, CNC, and cloud services. My understanding is those all work (decently) now not because they were obviously better, but because economic pressure pushed innovation to make them work, for better or worse.

> the _perception_ that they can/will replace developers is causing a major disruption in hiring practices.

Bingo.

And it’s causing the careers of a majority of juniors to experience fatal delays. Juniors need to leap into their careers and build up a good head of steam by demonstrating acquired experience, or they will wander off into other industries and fail to acquire said experience.

And when others who haven’t even gone through training yet see how juniors have an abysmally hard time finding a job, this will discourage them from even considering the industry before they ever begin to learn how to code.

But when no-one is hiring such that even students reconsider their career choice, this “failure to launch” will cause a massive developer shortage in the next 5-15 years, to the point where I believe entire governments will have this as a policy pain point.

After all, when companies are loathe to actually conduct any kind of on-the-job training, and demand 2-5 years of experience in an whole IT department’s worth of skills for “entry level” jobs, an entire generation of potential applicants with a fraction of that (or none at all) will cause the industry to have figurative kittens.

I mean, it will be the industry’s own footgun that has hurt them so badly. I would posit it may even become a leggun. The schadenfreude will be copious and well-deserved. But it’s going to produce massive amounts of economic pain.

> Juniors need to leap into their careers and build up a good head of steam by demonstrating acquired experience,

Junior devs at least have the option of building a portfolio of usefully software on their own machine at home, while eating ramens.

They can build websites for mom'n'pop stores. They can participate into open source projects. Etc, etc...

I dread the people who won't get jobs into other fields because managers have been told by corporate that "we don't need people, chatgpt can do everything".

> Junior devs at least have the option of building a portfolio of usefully software on their own machine at home, while eating ramens.

Bold of you to assume today’s young adults can live without a steady paycheque that doesn’t suck up 40-80hrs of their time a week. I mean, how else will that roof end up over their heads?

And most parents have been brainwashed to believe that any child not living on their own once they become adults are failures, and still need to be kicked out of the house such that they are forced to learn self-sufficiency.

AFAICT, most parents of adult offspring have zero clue about how bad things actually are out there, with most still telling their children to go from business to business with printed-off Résumés. Outside of blue-collar jobs, I haven’t seen this work for a good twenty years, now.

I'm probably picturing "junior" as "people still in college preparing to get their first job".

My point is that a luxury of the software engineering craft is that you can practice "at pro level" very cheaply. Even as a teenager, learning and using vscode/python/react/etc... on your own is a possibility.

Learning Salesforce and SAP and the internal support tool of BigCo is not.

That being said, I completely agree that we're going to put a generation of "wannabee white collar" in a dire situation. Cynically, this might be an overdue correction from the years of "college degree for everyone", and maybe (just maybe) some people will actually thrive in the "hard-to-llm-ize" profession if they can retrain.

(If the market laws apply, someone will build "turn advertisement-copy writers into electricians", and it would not necessarily be for the worst ? I know, easy to tell for me, who got the opposite deal by being a software engineer on demand at the right time at the beginning of my career.)

> while eating ramens

For many, even cutting their budget isn’t enough to pursue what you’re describing. Modern careers in software are very hard to reach for people who can’t afford to wait for a real paycheck, and it drives away a massive group of potential talent.

I expect it'll be a sigmoid curve — we're in the exponential growth phase, but it'll flatten out. Then we'll need to wait for the next Big Idea to give us another the next sigmoid.

That’s what I said about self driving cars nearly a decade ago!

The 80/20 rule is a painful lesson to internalize but it’s damn near a universal constant now. That last exponential improvement that takes LLMs over the finish line will take a lot longer than we think.

I think self driving cars is a good analog. We got lane centering and adaptive cruise control pretty much universally, and some systems are more advanced, but you cannot buy a fully autonomous car. Sure there’s Waymo and others pushing at the edge in very very limited contexts, but most people are still driving their own cars, just with some additional support. I suspect the same will be true for software engineering.

You can buy a Tesla with FSD today. It works. Sure there are some corner cases, but it's good enough in enough cases to get in, punch in a destination, and get you there. But you have to buy a Tesla, and they only come in electric. Waymo's are fine too, you just have to live in a supported area. The real gem for self driving, is Comma.ai.

It's an aftermarket system, but can be installed in most newish cars with a built in lane guidance system. Even a Corolla. And it's really want we want. It's not self-driving, it'll just take you down the freeway and you don't have to do anything except wait for your GPS to tell you to exit (and then you have to exit). Freeway-grade gurves in the road? Fine. Stop and go traffic? Handled. It's better than Tesla's FSD in two specific ways. One is that because it's not self-driving, it's way easier to trust the system because it's not going to change lanes or anything surprising on you. It's not going to do something fancy, just brake or accelerate or follow the lane or car in front of you left or right. I highly recommend it to anyone who does any amount of freeway driving. If not for the coolness of it then simply the safety aspect. Now, I'm sure everyone here is a better driver than most (though that has a problem, mathematically), but this thing is better than a tired/angry/drunk version of someone else driving.

But as you point out, most people are still driving their own cars.

Which, I think is where we're going to see the software development industry going. There's gonna be the AI maximalists, who, like Waymo and FSD, will have AI basically do everything. And then there's the pragmatists, for whom AI doesn't do everything, just enough to be useful.

Then there's everyone else, still writing their own code. Thing is, an app on your computer isn't a car. It's $20/month if not free if you're a total cheapskate, to get codex or claude or another assistant, vs many thousands of dollars for a self-driving or partially self-driving car. The other difference is in time. The value of a partially self-driving car (FSD or Comma) is in the mental fatigue of driving, and in improved safety, but a 7 hour road trip is still going to take 7 hours even if you're not driving. The only time a self-driving car helps is if you're going cross-city in a Waymo, and you're in the back seat working on your laptop. AI assisted coding though is different. If I take on projects I wouldn't do before with AI, that's a win for me. If I'm able to write software faster and with fewer bugs with AI, that's also a win for me, but also a win for my employer. If, however, it goes the other way and I write more bugs, then that's a loss for me and my employer.

I'm not sure what "exponential improvement" would mean in this context, but large models have been a massively hyped and invested thing for what, three-four years or so, right?

And what do they run on? Information. The production of which is throttled by the technology itself, in part because the salespeople claim it can (and should) "replace" workers and thinkers, in part because many people have really low standards for entertainment and accept so called slop instead of cheap tropes manually stitched together.

So it would seem unlikely that they'll get the required information fed into them that would be needed for them to outpace the public internet och and widely pirated books and so on.

The counterpoint to this is that information is sealed up in bottles that previously haven't been worth it to be unsealed. How much can you charge for the source code to a program writing in Zig to calculate the fibonacci sequence? That's worth approximately zero. But generating a million tested programs, with source code, that have been run through a compiler and test suite, to be used as "information" suddenly becomes worth it to the AI labs to buy up, if not generate for themselves. So imo there's still ways to go, even if the human Internet isn't growing as much post-AI as in all the years before it.

As I understand it synthetic data and information, and likely also synthetic code, is already used in training of large models.

Counterpoint…

I am not worried about losing my programming role to AI.

I am worried about hiring employees and contractors. I haven’t had to hire anyone i office since, but I have specially avoided Upwork and new contractors. It’s too hard to tell if anyone knows anything anymore.

Everyone has the right or right enough answers for an interview or test.

The bar to detect bullshit has been moved deeper into the non-detectable range. It’s like everyone has open-book testing for interviews.

Even if I can sus out who is full of shit in a video or phone interview… the number of people I need to sort through is too large to be effective.

For Upwork specifically, this was an issue for years already. With people buying US accounts and lying about their location or subcontracting to cheaper foreign labor.

So, is vibe coding something I want to hire? Absolutely not. But, I don’t see being able to avoid it or at least not suffering from someone cutting corners.

Having used Claude Code in anger for a while now, I agree that given the state of these agents, we can't stop writing code by hand. They're just not good enough.

But that also doesn't mean they're useless. Giving comparatively tedious background tasks to the agents that I check in on once or twice an hour does feel genuinely useful to me today.

There's a balance to be found that's probably going to shift slowly over time.

To me the biggest benefit has been getting AI to write scripts that automate some things for me that are tedious but not needed to be deployed. Those scripts don’t have to be production-grade and just have to work.

Similar experience. I just tried Claude for the first time last week, and I gave it very small tasks. "Create a data class myClass with these fields<•••> and set it up to generate a database table using micronaut data" was one example. I still have to know exactly what to do, but I find it very nice that I didn't have to remember how to configure micronaut data, (which tbf is really easy) I just had to know that that's what I wanted to use. It's not as revolutionary as the hype, but it does increase productivity quite a bit, and also makes programming more fun I think. I get to focus on what I want to build, instead of trying to remember jdbc minutiae. Then I just sanity check the generated code, and trust that I will spot mistakes in that jdbc connection. It felt like the world's most intuitive abstraction layer between me and the keyboard, a pretty cool feeling.

Just for fun, once I had played a bit with it like that, I just told it to finish the application with some vague Jira-epic level instructions on what I wanted in it and then fed it the errors it got.

It eventually managed to get something working but... Let's just say it's a good thing this was a toy project I did specifically to try out Claude, and not something anyone is going to use, much less maintain!

> Just for fun, once I had played a bit with it like that, I just told it to finish the application with some vague Jira-epic level instructions on what I wanted in it and then fed it the errors it got.

Would you finish the application with "some vague Jira-epic level instructions"? Or, even if you don't formally make tickets in Jira for them, do you go from vague Jira-epic-sized notions to ticket-sized items? If I had a mind-control helmet that forced you to just write code and not let you break down that jira-epic in your thoughts, do you think the code would be any good? I don't think mine would be.

So then, why does it seem reasonable that Claude would be any good, given such a mental straight jacket? Use planning mode, the keyword "ultrathink" and the phrase "do not write code", and have it break down the vage Jira epic into ticket-sized items, and then have it break it into sub tickets that are byte-sized tasks, and then have it get to work.

I mean, I didn't really expect it to work, I just wanted to see what would happen. I'd had pretty good results thus far and wondered how far I could push it. Jira-epic-style prompts was, not surprisingly, pushing it too far.

It did manage to get the application working though, with only a couple of "this thing broke, plz fix" style prompts, and it did better than I had thought it would to fulfill my intention, give how vague I was.

My point was that if you're going to build an actual product, you should probably not use Claude in that way. Break down the epics to smaller more manageable chunks however, and Claude can do an amazing job! I'll definitely keep experimenting with it this way, it's way better than full-manual coding, or at least that's my initial impression of about a week of experimentation!

I will give my perspective as an academic who writes R and Python code for data analysis (including a lot of data cleaning).

1. I find AI written code verbose and inelegant. It makes it difficult to troubleshoot. I take pride in my own code and share it confidently with my doctoral students and coauthors. I can't share AI written/assisted code with the same confidence let alone pride.

2. Often AI takes shortcuts and writes terrible code. This is especially true for Bayesian models. My first check with any Bayesian model is to recover parameters using a simulated dataset. If the code fails to do that, there is no point in going forward. I used Opus 4.5, Gemini 3.0, and GPT 5.2 recently to write rather simple code for random parameters dynamic panel model. There are already papers that have done it. All three failed numerous times. I got it to work after a lot of handholding.

3. AI helps tremendously while creating web apps that make my analysis more actionable. A lot of reviewers now want to see something in action, so using Streamlit or Shiny is the way to go.

I've noticed a similar pattern. AI assistants are incredible at kinetic coding i.e. generating boilerplate, refactoring, writing tests. But they are detrimental to potential coding especially at the architectural thinking that happens before you touch the keyboard.

Writing by hand (or whiteboarding) forces you to load the entire context into your working memory. AI allows you to be lazy with your working memory. The code gets written faster, but the mental model of the system in my head is significantly weaker.

Definitely, it's great at making a template but, it's hard to see or continue from what it has generated.

I agree that writing by hand sets up that mental model, and because it's written by you, it's easier to branch out from there.

Recently I've seen coworkers frequently turn what should be a <10 line bugfix into a 500+ line refactor across multiple files. I suspect it's due to AI.

There's a time and place for refactoring, but just fixing an isolated bug isn't it. But I've seen that often AI can't help itself from making changes you didn't ask for.

This is a major annoyance for me. In one PR I reviewed recently, it introduced concurrency where it was not needed and then cascaded to a bunch of synchronization logic everywhere as a result. What could have been 30 lines or so turned into a couple hundred line diff and the submitter did not think to question it. It was harder to review and introduced a potential for data races with zero benefit. These are all competent people I work with, who do good work normally too.

I'll admit that I've done that type of bugfix in the past. Though, usually it's because the bug is caused by an underlying design flaw, and the 10-line fix would have only papered over a single instance of the bug, and wouldn't actually solve the underlying cause.

On the other hand, how do you know what they're prompting? For all you know, the AI proposed the hacky 10 line fix and tour coworker was all hey let's fix this problem better.

Before AI, "now" was seen as the wrong place for a refactor because of the time and effort involved, but if the hacky patch of a fix is roughly the same amount of wall clock time as a 500 line proper fix that makes the problem go away forever, what's the argument that the 10 line hack is actually better?

Certainly refactoring is better than a hack fix, but what I'm talking about is where the 10 line fix cleanly and fully addresses the bug, and the refactor is just thrown in anyways.

> if the hacky patch of a fix is roughly the same amount of wall clock time as a 500 line proper fix

That only makes sense if the coworker isn't validating the fix and no one is reviewing the code. 500 lines of slop is going to take much longer if you have a human in the loop.

I wonder when I will feel compelled to go back. Right now it just feels too productive for me to let AI write the code.

I wonder if I will feel compelled to go back. Right now it just feels too productive for me to not let AI write some of the code.

This is the best summary of everything that’s wrong with:

1. One-shot app development using Lovable with English requirements

2. Spec driven development tools like SpecKit, Kiro that promise English as being sufficient for LLM generated code

3. Using agent swarms with Claude Code and thinking you are making progress and you never have to look at the generated code, when the overall coherence or elegance of the codebase is lost.

Loved the analogies in this video:

“It looks like an airplane. It flies like an airplane. But when you inspect the code in its full context, it’s a paper plane.”

and

“It looks like an airplane, walks like an airplane and talks like an airplane. It’s probably an airplane. But it’s not an airplane, it’s a psychedelic airplane …filled with hallucinations”

But, still the tools are incredibly helpful when you have the right guardrails and take the time to ensure (somehow) the output matches your taste and code quality expectations.

Elegance or "taste" for the codebase is wrong. It's not just a stylistic preference. It's objectively bad code that's going to be slow to iterate on and very easy to introduce bugs to

I'm not saying AI has no place, but letting it take the lead in any way is just building your app on sand. Fine if you're building a sandcastle, bad if you're building a bridge. It's not "ohh well I just don't like the way sandcastles look"

"It pases test completely consistently".

There are some managers and developers out there who would charge and ship whatever code AI provides no matter what multidimensional horror there is inside. As long as these people point to test and say "look, the checkbox is green" they will sign on it without taking a single look at the codebase.

So what if tests don't actually cover important parts of the functionality. Checkbox is green and AI review agent said it's fine - so it must be fine.

Artisanal engineering is the only true way to code. No autocomplete. No linting. No prettifying. No colored brackets, braces, or nesting lines. If you're not instinctually converting your Rust to assembler in your head as you insert it into vi, can you even really call yourself a real developer?

I posted this comment using only curl by the way. Don't bother replying because I only engage in bidirectional conversation via SSH tunnel and netcat. I doubt that you could figure it out.

"You like code that's easy to work with, easy to understand, change and evolve? HAH! Bet you hate colored brackets and prettifying too!"

The mistake here was the expectation that LLM code would not be junk.

It's the same junk produced by fast-growing startups. It was always here, now there's more of it.

I'm convinced that people will continue to use it. Therefore, I need to be able to deal with it, maybe even leverage it somehow (there should be coding skills that I possess that people in that realm of fast-growing startups do not bother to learn).

There is a balance to be struck. Not everyone is going to be comfortable with ralph loops. Some are going to be OK with running a single agent, some with advanced code completion or code generation for specific functionality and so on.

The tooling is going to change how we do development no doubt, but people are going to find their comfortable spot, and be productive.

The key word from the video is "responsibility".

Who is going to be responsible for the code? AI is definitely not responsible.

it's usually the balance and middle that is most beneficial. you can't deny the value LLM code generation and research provides. But the extreme of using only LLMs or mostly LLMs. or not using LLMs at all is self-harming.

So far, LLM generated code hasn't lived up to my standards. I'll use it for things that aren't critical as-is, but mostly I use it as a reference, an example, a starting point. Essentially, where in the past I'd find a code base that does things and I'd try to do something similar, now I let the LLM generate the code base. There are to questions it helps me answer:

1) What are the possible ways of solving problems?

2) What are the pros and cons of each approach?

That said, there are people successfully deploying apps that are entirely vibe coded. How many fail or succeed, that I don't know. But there are enough, and you can't deny the evidence.

Why is not using LLMs self harming? We've been writing code without help of LLMs for many decades just fine.

This is what I find amusing, we're in tech, how can you ask that? We've been rubbing stones to start fire too, but if you stick to that and refuse to use electricity, that's self-harming because you lose out on all the benefits.

You can't keep doing things the same way, that's now how technology as an industry works, the whole point is to come up with newer and better things. If you're a user of technology, then you can think like that and keep using old tech until the natural balance of things forces you.

I fully expect LLMs to be obsolete in a few decades, and I'm now wondering if people then will say how LLMs have been serving them fine for a few decades.

"just fine" isn't good enough in tech, "better" is always the goal.

> that's self-harming because you lose out on all the benefits.

This is some big stretch in reasoning. Four years ago we were not harming ourselves by not having AI and writing code ourselves.

At most you're hurting your ability to compete with "vibe coders" whose metric so far has been lines of code and $ spend per day running agents, not successful products.

Have you considered how relying on AI affects your own programming skills?

I think you have to keep in mind that what you write is more important than how you write it. Can you write better programs with this tool or not? All the time I used to spend on stack over flow, asking people questions, scouring through bad documentation and source code, now AI can do all that for me. Now, if you're a 10x engineer or whatever the term is these days, and you're telling me you have no need for research, for PoC code, for critical analysis of your code, then power to you. But most people aren't in that boat.

> Have you considered how relying on AI affects your own programming skills?

Yes, about as much as it did before but now I have more time to think more deeply about solutions instead of cosmetic things, syntactic nonsense,tedious b.s.,etc... It's like asking how using an IDE with auto-complete affects your programming skills when you could be writing in vim or notepad. Doing tedious things, knowing how to beg right on the correct forum, or hunting down the best library, the best doc,etc.. these are not programming skills. Like I said in my original comment, I'm not using it to write the code for me, I'm using it to do things that have nothing to do with solving the problem at hand.

With AI's assistance, you can become a better problem solver. heck, you might even be good at spotting mistakes other people make, just by force of habit learned from correcting mistakes AI makes. It's just a utility, take all the benefits from it and discard things you don't like about it. Nothing forces to use the code it generates. think of it as a junior programmer that's doing PRs for menial tasks for you, and that's the best case scenario. It's just a really good google search that gives you the results you're looking for instead of jumping through hoops to get something of lesser quality. Many times I've mistrusted the AI and did things the old way, and its solution contained more nuance, details and considerations I failed to glean with my initial attempts.

Before AI, the story of people copy pasting things from stack overflow, including bugs, and including good code that gets buggy in certain situations was a major trend. AI does that, except better both in terms of result quality, and in helping you avoid pitfalls. But imho, you still shouldn't trust it, you still have to vet everything it does and stop using it if that effort takes more time and it would have doing things manually.

I really appreciate all of his message -- responsibility and actual engineering are critical and can't be (deceptively) lost even though Pull Request and CI/CD workflows exist. I hate the term vibe-coding because it seems flippant, and I've leaned into LLM-assistance to frame it better.

I consider vibe coding and LLM-assistance to be distinctively separate things.

I am vibe coding, if I needed x, I lay that out task with any degree of specificity, and ask for the whole result. Maybe it’s good, I gave the LLM a lot of rope to hang me.

I am using an LLM for assistance if I need something like this file renamed, and all its functions renamed to match, and all the project meta to change, and every comment that mentions the old name to be updated. There is an objectively correct result.

It’s a matter of scope.

All of these “go back to handcoding” posts seem to be done by very experienced coders. The fact that a less than mediocre coder like me can have sql statements or Python ETL written and tested for me in seconds rather than hours is all I need to see.

How do you know the output is correct?

1) Testing for myself 2) Vibe code testing 3) Team members do work from scratch and try to break 4) can read and reason it out even if I come out with it myself.

Looks like an airplane, fails under real inspection. Same smell in diffs: bugfix turns into refactor. Review eats the "velocity." Revert rate keeps you honest.

I'm feeling AI is already falling apart dramatically in the first month of 2026, with so many negative news.

No.

i am new to llm assisted coding, but i dont like the way it try to add more stuff when fixing thing, instead of going simplest less code possible, it recreate bunch of already coded logic, it also mostly try to code workaround and accumulate spagetti-esque codes. any advise? appreciate it

Tell it to stop?

By this I mean give it precise instructions and ensure the instructions are always included in the context window. You’re responsible for managing the agents effective working memory, if that analogy makes sense. If you don’t tell it the way you want it to act, it’s effectively a blank slate.

I am trying to log in to my pet insurance's website. The log in used to work just fine about 3 months ago. For MFA, I have been waiting for 10 minutes and have not received the code in my email yet. I have reached the maximum attempts(3) at resending the code. The code is supposed to be expired in 15 minutes. I am pretty sure somewhere some dev using AI has f'ed up and I hate it. Code generation is not hard anymore, keeping a software working through the changing world is hard. I will update here once I have been able to log in.

I received multiple codes. Entered the latest one and a pop up message says, "Our dog’s gone off-leash… We’ve run into a problem and are working to fix it as fast as we can. Please try again soon."

I am not liking this era of infinite abundance.

I was able to get in after another attempt but I keep receiving the mfa codes after log in. Lol.

> "in isolation this code makes a lot of sense"..."what is this junk"

I mean I am left with two thoughts:

1. programming language skill issue. Some languages are simply much better at composition than others. I find that yes, this happens, but actually on the order of a day, and once the code is "good", it doesn't really change that much in the grand scheme of things?

2. Even for languages where composition is better, this is exactly what happens with human development too?

2. tends to happen less with more experienced engineers

I think the mistake is assuming you get high quality work with an agent regardless of the driver’s skill level. That doesn’t appear to be the case, but it’s the value prop that is being sold.

2 is true even for highly experienced engineers over the long horizon, you keep building and then you realize that there's a better pattern that you should refactor a bunch of stuff for, but it gets tabled because feature XYZ is more important, and then feature YZW, etc. and then one day you're faced with looking down the abyss of committing to a refactor while some JE wants to rewrite everything in rust.

And then you realize that most stuff is not worth refactoring if you can’t finish it in one go… it’s like a dunning Krueger roller coaster.

Contrary to what most devs belief: Most code is not shipped to hundreds of millions of users and passing the test and actually implementing a feature is worth more that drowning in backlog.

The video is spot on for codebases of products that are critical systems: payment, erp etc -> single source of truth.

Simple Crud apps/ Frontends for ecom that abstracted away the critical functionality to backend APIs (ERP, shop system, payment etc) benefit from vibe slop vs no shipping cadence

You're producing technical debt. At some point you will invest more time fixing the vibe slop than it would have cost you to do the work yourself in the first place. A lot of vibe-coding just feels like shifting responsibility and resources from "development" to "incident response".

The term "vibe coding" is less than a year old!

The weekly thread on AI coding.

I've been trying Google antigravity. Despite VS code being total dogshit compared to Intellij I'm pretty sure I won't be renewing any more (after 15 years, sorry jetbrains).

More like X-times daily. Where X is a function of how much money Company Y need to get from investors this week to continue training and running LLM models.

It’s now ship shit but it’s the right way to do. We need to figure out how to make it ship high quality code as possible as we can. Not just give it up.

Try implementing something that is too hard for you. Usually that'll involve implementing math in a high performance language or with parallelization. Then try going back to "writing by hand".

Nothing that is possible is "too hard" if you're willing to put in some effort. The only question is whether you will learn to do it, or press a button, hope the LLM did it well, and let it forever remain "too hard."

Honestly, without judgment, I think this is just a fundamental difference in how people approach their craft. You either want to be capable yourself or you just want the results.

> Try implementing something that is too hard for you.

This is almost the only thing I'm against when it comes to LLMs. You have no ability to figure out if it is right, and you will be overly impressed by garbage because you aren't qualified to judge. Has anybody come up with a pithy way to describe Dunning-Kruger for evaluating the output of LLMs, or are people too busy still denying that Dunning and Kruger noticed anything?

When it comes to implementing math, the main problem is that the tiniest difference can make the entire thing wrong, often to the degree of inverting it. I wouldn't be in any way comfortable in shipping something I didn't even understand. The LLM certainly didn't understand it; somebody should.

if it’s too hard for you to write, it’s too hard for you to understand and comprehend. how are you going to take responsibility for that code and maintain it if needed?

From my attempts, still, it takes more time to fix the code than it would if writing it all by hand. BUT! It's getting better and I look at it as just the next abstraction layer we will get used to. Think of it. Back in University I had to write code on PAPER. To think of memory allocation manually! Then came out managed code.. Then huge SDKs.. Then smart IDEs with intellij/intellisense..

And some us even remember riding the CPU with mov rcx, 5.

We were just shifting our focus byte by byte from the nits and bolts towards the actual problem we are solving. In other words going less "hard"-ware, more and more "soft"-ware.

AI is just continuing this evolution, adding another abstraction layer in soft dev process.

Reply on news.ycombinator.com