Vibe Coding Is Fun–But Vibe Refactoring Pays the Bills | Hacker News Reader

Vibe Coding Is Fun–But Vibe Refactoring Pays the Bills

dawidmakowski.com

62 pointsmakowskid56 comments

The author's description of vibe coding seems to have confused it with getting into a flow state in a vaguely AI-adjacent way. Vibe coding as originally coined by Karpathy [0] can't pay the bills because it would be wildly irresponsible for a professional to do—it's the kind of no-checks minimal-validation AI usage that gets lawyers into trouble for submitting legal documents with hallucinated court cases, just translated to coding instead of law.

What the author is proposing isn't really "vibe" anything, it's just dedicating a small amount of time to fixing tech debt in a way that happens to involve an LLM as an assistant. The LLM in this model is honestly mostly superfluous.

Don't get me wrong, this absolutely is how LLMs should be used in a professional setting, but I just question why we needed a name and a blog post for it. This is just responsible code maintenance as it's always been.

[0] https://x.com/karpathy/status/1886192184808149383?lang=en

eh terminology happens. I’d argue we shouldn’t be using “AI”, but here we are. I’d prefer encouraging “vibe coding” to mean this over fighting against the wind

But the moment it means two things.

One is ok, the other is risky

One is YMMV the other is bad.

But what is "this"? It's just refactoring with an LLM along for the ride! That's increasingly the default assumption when we're coding or refactoring, so the extra "vibe" attached at the front is totally meaningless if it doesn't imply that you're going so fast you're not checking the LLM's work.

> The author's description of vibe coding

Judging from the style and complete lack of substance I don't think there is an author, unless you count ChatGPT as one.

That didn't stop the submission from rocketing to #1 on here though since it jingles the right set of keys.

The articles getting to top are getting worse rapidly - low effort AI crap with a random headline gets to the top of HN regularly.

I have a feeling a lot of "vibe coders" are joining the community. Could be bot spam too. At this point I'm starting to miss the liberal thought pieces.

It's interesting because this site usually encourages 'curious' discussion, and there was always an impression that the userbase was mature enough to spot these kind of substanceless pieces easily.

I'm pretty sure this is closer to what Karpathy called "real coding": https://x.com/karpathy/status/1915581920022585597 :)

Thanks for your comments - but I wasn't using GPT to generate the article, I only used it once for proofreading my own thoughts :)

Also - I was focusing on the REFACTORING part where LLM might be one of 100s of tools to be used while being in a state of flow of goal-less refactoring.

Disagree gently here. Vibe refactoring is so much harder than vibe coding. I understand the temptation of having LLMs refactor code because that's ostensibly what they're great at – "transforming" the code. But LLM's are opinionated enough where they love to arbitrarily drop features or constraints for no particular reason, and they won’t tell you what they dropped along the way.

Proper refactoring with an llm requires full testing coverage and, by the time you do all that, is the refactoring really necessary? I prefer 100% stability. If you're refactoring the because it's poorly structured and unreadable, that’s okay.. LLMs can help understand it.

In my use of llms, i find it’s actually much easier to rebuild something from scratch rather than refactoring flawed code. It’s much less likely to inherit strange assumptions and code smells that way.

With all that said, the one prompt I do use when refactoring is to tell the llm to do a lossless refactor and then follow up with a "was this really lossless"? It's not foolproof. LLM's love to lie.

> Proper refactoring with an llm requires full testing coverage

Proper refactoring WITHOUT an LLM requires full test coverage! But most definitely WITH an LLM.

In cases where there is no test coverage, the first thing I do is have the LLM write one test at a time. The problem there is that if you truly wanted valid tests, you'd have to actually break the code, watch the test fail to prove it was valid (basically, the inverse of TDD) and then re-fix the code and start on the next test, but in practice, it is difficult to get an LLM to stick to this loop. I wish someone would train or refine some coding LLM to use either TDD or this form of "inverse TDD" where you're applying tests after the fact and also want to check their validity. (Or tell me how to do it.) Because mere prompting doesn't seem to stick- it always regresses to the mean eventually.

(I'm currently seeking work, btw, and would probably be happy to help refactor old code, advise people on code, etc. Sorry for the self-promo.)

Refactoring is indeed one of the areas where LLMs already shine bright. And I love it.

I have been coding since I am 12 years old. I always loved writing code. But I also always loved reading code. I don't know, for me code is a kind of art. Never met anyone else who sees it like this. When a friend was hiring for his startup a while ago, I was happy to sit down for multiple hours and read all the code the applicants wrote and gave him advice on whom to hire.

So for me, the new times are paradise. I try to not touch code directly anymore. I write prompts that would enable a really good developer to implement features and then let various LLMs work on it. Afterward, I rate the results so I have an overall score for each LLM. I pick the best solution for my codebase and manually finetune it to perfection.

After each commit, I also ask the LLMs if they can find anything in the files that can be refactored to make the code shorter or more logical. The result is that the codebase becomes better and better. Because the LLMs often find stuff to improve. They usually come up with 10 ideas I dislike, but also one idea that I like. And so the codebase becomes better and better over time. Instead of worse and worse like in the past when you had to keep a balance of refactoring for the sake of beautiful code and building new features. Nowadays, refactoring becomes more and more free.

> I also ask the LLMs if they can find anything in the files that can be refactored to make the code shorter or more logical

Citation needed? The worst kind of code tends to be clever code. This seems like a lot of code churn for no real benefit other than some loose definition of "better". How do you prevent bugs with these constant refactors?

You have to read the suggestions and evaluate them using your skills and experience and then decide to accept / deny / adjust the proposed changes. There is not silver bullet but the speed you can iterate and experiment is impressive

Man, I must really suck at this stuff. This is not at all my experience. Asking LLMs to refactor almost always results in hasty abstractions that I want to keep out of my codebases at all cost. Am I not letting go enough?

I really, really don't understand this either. Sometimes I feel like I must be using different LLMs than some HNers because my experience of them is the complete opposite of what they describe.

Which LLMs have you tried?

Not who your replying too, but since I've expressed similar sentiment:

Sonnet 3.7 with thinking is my go-to.

Deepseek R1

Gemini 2.5 pro (I've heard it said Gemini is outperforming Sonnet, but I find Sonnet more consistent)

O1-mini

Depending on what I'm doing it's generally either via Cody or Aider.

FWIW this is my experience too. I use LLMs pretty regularly for coding but to get decent code you really have to supervise the hell out of them and often it's not worth the effort to push them into doing the right thing.

Maybe I'm just bad at getting it to do things, but I think your question about "letting go" is the real story. I think there are a lot of people not paying close enough attention to what's coming out of the LLM, and the tech debt building up is going to come back to bite them when it builds to a point the LLM can no longer make progress and they have to untangle the mess.

The way I "ask" is that I really ... ask!

I ask the LLM "Can you find anything in this file(s) that can be made shorter or more logical?"

And then as I said, I like less than 10% of the ideas the LLM comes up with. But it is so fast to read through 10 ideas (A minute or so) that it is well worth it.

As the author of this article/video - it's really strange for me that almost everyone here in these threads has focused on the LLM part of refactoring, where it's only mentioned as one of many tools for target-less flow state of the code tweaking sessions. :-)

So, I know the battle has been lost for the definition of vibe coding (https://simonwillison.net/2025/Mar/19/vibe-coding/).

But we definitely need a word that makes a distinction between a LLM autonomously generating software, and having a human ultimately curating all the code (even if they're using a LLM to generate it.)

If not vibe coding, what should we call that?

pAIr coding? pAIr programming?

I actually kind of love this.

Simply coding. It's going to be AI-assisted by default. It already is for students getting into programming, and it'll take some years for old people to pick up these tools too, but I imagine we'll get there pretty soon.

What do you call coding with Intellisense? What di you call coding while referencing documentation?

AG code?

Lazy coding

Judging by the younger group coming out of college that I interact with, it's just coding. This is just how it's done now.

I find this happens for pretty much every name that doesn't actually match what it is suppose to describe.

In this case, most people have to be told what "vibe coding" is. I think hardly anyone guesses what it is by just hearing the name.

So I'm not at all surprised that people are using it to mean other things already.

Where do you draw the line though? Is it only vibe-coding if you don't touch a single line of code? If I ask Cursor agent to implement a basic refactor vs doing it by hand, does my level of "vibe" increase?

Why do we need a term anyway? Neologisms are nice for writing articles but I feel like we get lost in the weeds unnecessarily trying to categorize things. Its like trying to come up with genre names for electronic music. The more terms we create, the less useful they become.

I always thought of it as a temporary term anyway while we reconcile this technology with the status quo. In a few years I'll bet it will go back to be just "programming" or "development".

If don’t touch a single line of code, you aren’t coding.

Or where your customers/bosses vibe coding when they told you their requirements?

> Why do we need a term anyway?

For the same reason I think it’s useful to distinguish between purely AI-generated imagery vs. using AI tooling in some specific and restricted way. e.g. if someone uses an AI de-noising program on a RAW photograph, the implications of that are entirely different from generating the entire image.

Until recently, it was not possible to generate sophisticated programs from scratch across a wide variety of problems domains without being involved in writing the code. This type of “development” is entirely unlike a more limited “AI-assisted” workflow and the implications of each are quite different.

As a viewer of artwork, how the art was produced entirely changes how I value it.

As a user of software - especially in certain categories - how the software is produced also changes how I value it, whether or not I trust it with my data, would be willing to install it in my computers, etc.

To your point, there’s a spectrum of AI involvement. But I think it’s necessary and useful to have language that helps identify software that sits on the extremes of that spectrum. Classifying things across the range is more difficult.

What do we call emails that have been spell checked?

What do we call automatic generated emails?

Usually spam

Yes, one is created without that personal connection while augmented with technology.

One is coding, the other is ordering.

Otherwise all my customers/bosses did vibe coding way before LLMs.

Why do we need a term for any of it, other than "coding". Before AI, I did not need different terms for coding when some devs would cowboy-code buggy slop straight to production vs. a full team doing test-driven development with dedicated QA staff on top of it, or anywhere in the middle. Vastly different qualities of work, but the same term - "coding".

So why do we need a term? What difference in action will hearing a different term drive?

On that front, I like the fact that the term "vibe coding" exists because anyone who uses that term unironically has just told me that they care more about the process than the result. Now that is information that changes my actions because it changes how seriously I take them.

It's going to be a spectrum so will be hard to divide up.

I know a crash-course-day's worth of programming language(python and C), but very strong understanding of the principles of programming. So LLMs are a godsend because I can basically write code, function by function if need be, in English.

"Create this variable, do these mathematical transformations on it with retrieved value from API, display the output of it here, also compare it to the output of everything in the test set, store differences greater than 30% to their own set, store all values in an SQLite database, create a simple GUI with a field showing each output, make a button that outputs the results to a .pdf, give it a title block with labeled results listed, etc. etc."

Does coding become vibecoding when you do it in English? hah

Be _very_ careful about refactoring with Cursor -- it has a tendency to randomly delete important blocks of code and comments and re-order things in a way that make for difficult diffs to read. Make sure you do it in relatively small chunks and do PRs and have others review. The larger a refactor you do, the more difficult/impossible it becomes to actually review the code usefully.

Or adding. I've had to point it out multiple times to a co-worker when Cursor keeps adding the same thing to a data structure that shouldn't be there.

I don’t really get what they’re doing. This would be a better article if they actually explained how they use repomix.

Did you use AI to write this blog post? Be honest.

my first though as well, it was like reading any other chatgpt response

Nope, I didn't. I never generate any of my articles on my blog. I used (as always) ChatGPT to help me proofread typos, though.

Are we at the point yet where companies are looking for programmers specifically to fix their vibe coded codebases? A codebase restoration expert?

Or will they do what most companies do when they sink millions of dollars into a codebase that doesn't work: dump the codebase, dump the team, hire a new one, and build from scratch?

What does everybody think?

edit: oxford comma is life.

I tend to use tools like ChatGPT to refine my code.

i.e. "I have this function. Can you suggest ways to make it more efficient?" etc.

Sometimes, it gives good feedback, sometimes, not. I almost always need to modify whatever it gives me.

Are any of the freelance software dev marketplaces seeing an uptick in opportunity to refactor vibe code?

Probably an opportunity to category kill that niche.

"Got a vibe coded prototype you want to make more robust?"

We've reached the stage where the word vibe in a coding submission is a really good indicator that you should just flag it and move on. It is a meaningless, incendiary, click-bait term that is meant to trigger people and induce engagement.

"There’s a lot of hype about vibe coding"

There is incredibly little hype around vibe coding. 99% of the comments about vibe coding are people propping it up as a strawman to knock down. Otherwise it's like it fills the void left by web 3.0's decline to irrelevance where a bunch of useless masturbatory noise is had by people trying to get in front of something that they think will be a thing. Maybe they can put it on the blockchain.

All of us normal people incorporate LLMs into our work process. No vibes at all. Just another tool in our belt.

EDIT: Just discovered that my comment is apparently dead by default, which is...interesting.

I noticed because I show dead. And magically, it's not dead now. Welcome back from banishment llm_nerd. Opaque heavy handed moderation is the only way to run websites now evidently.

So this is a plug for repomix?

Some questions,

Does repomix do anything Github copilot should be doing? It seems like this should be something copilot does automagically.

Does it work on any language? I notice the repomix github suggests a different tool if you're using python.

It seems straightforward to create an output.xml on the repomix site, but is there an opinionated try-it-free AI to use that output with?

I'm tired of trying things only to get ai slop. If this cures the slop, I would be interested.

edit: it seems the HN discussion is dominated by the definition of "vibe coding" and not at all interested in what the article presented as a solution... nice.

can we just ban llm-generated blogspam submissions already

I didn't use llm to generate it. I used it to proofread my thoughts, though :)

Why does everyone these days assume that everything is generated? You can check my other articles on my blog as well - none of them is generated.

Hiring a bunch of recent college grads to build your startup is fun. They're great drinking buddies and they make you feel young again. They'll inefficiently work around the clock, sleep at the office, and treat it like their mom's house, and make you think you're a big, important boss.

Coming in as a gray hair to fix their disaster definitely pays the bills.

I wont feed my production code to a LLM thanks for trying though. There is not a single trustworthy "AI" company out there. My companys codebase is a complete dumpsterfire and I will extinguish with a complete rewrite later this year. There is nothing, absolutely nothing worth paying for when it gets to LLMs.

Reply on news.ycombinator.com