I don’t really care about vibe coders but as a dev with just under 2 decades in the field:
- Your vibe coding shit will not go to prod until humans fully review it
- You better review it yourself first before offloading that massive mental drain to someone else (which means you still need to have some semblance of programming skills). Don’t open a PR with 250 files in it and then tell someone else to validate it.
- Use more context. Don’t give it vague ass prompts.
- Don’t use auto-accept. That’s just lazy asshole shit.
I can’t stress this enough: if you give me a PR with tons of new files and expect me to review it when you didn’t even review it yourself, I will 100% reject it and make you do it. If it’s all dumped into a single commit, I will whip your computer into the nearest body of water and tell you to go fish it out.
I don’t care what AI tool wrote your code. You’re still responsible for it and I will blame you.
When I see a sloppy PR I remind people “AI didn’t write that. You wrote it. Your name is on the git blame.”
I like this mentality. I might start telling people the same thing
Vibe coding is useful for super basic bash scripting and that’s about it. Even that it will mess up but usually in a suler easily fixed way
I don’t think it has much to do with how “complex or not” it is, but rather how common it is.
It can completely fail on very simple things that are just a bit obscure, so it has too little training data.
And it can do very complex things if there’s enough training data on those things.
As a software developer, I’ve found some free LLMs to provide productivity boosts. It is a fairly hairpulling experience to not try too hard to get a bad LLM to correct itself, and learning to switch quickly from bad LLMs is a key skill in using them. A good model is still one that you can fix their broken code, and ask them to understand why what you provided them fixes it. They need a long context window to not repeat their mistakes. Qwen 3 is very good at this. Open source also means a future of customizing to domain, ie. language specific, optimizations, and privacy trust/unlimited use with enough local RAM, with some confidence that AI is working for you rather than data collecting for others. Claude Sonnet 4 is stronger, but limited free access.
The permanent side of high market cap US AI industry is that it will always be a vector for NSA/fascism empire supremacy, and Skynet goal, in addition to potentially stealing your input/output streams. The future for users who need to opt out of these threats, is local inference, and open source that can be customized to domains important to users/organizations. Open models are already at close parity, IMO from my investigations, and, relatively low hanging fruit, customization a certain path to exceeding parity for most applications.
No LLM can be trusted to allow you do to something you have no expertise in. This state will remain an optimistic future for longer than you hope.
I think the key to good LLM usage is a light touch. Let the LLM know what you want, maybe refine it if you see where the result went wrong. But if you find yourself deep in conversation trying to explain to the LLM why it’s not getting your idea, you’re going to wind up with a bad product. Just abandon it and try to do the thing yourself or get someone who knows what you want.
They get confused easily, and despite what is being pitched, they don’t really learn very well. So if they get something wrong the first time they aren’t going to figure it out after another hour or two.
But if you find yourself deep in conversation trying to explain to the LLM why it’s not getting your idea, you’re going to wind up with a bad product.
Yes. Kind of. It takes ( a couple of days) experience with LLMs to know that failing to understand your corrections means immediate delete and try another LLM. The only OpenAI llm I tried was their 120g open source release. It insisted that it was correct in its stupidity. That’s worse than LLMs that forget the corrections from 3 prompts ago, though I also learned that is also grounds for delete over any hope for their usefulness.
So there are multiple people in this thread who state their job is to unfuck what the LLMs are doing. I have a family member who graduated in CS a year ago and is having a hell of a time finding work, how would he go about getting one of these “clean up after the model” jobs?
It makes me so mad that there are CS grads who can’t find work at the same time as companies are exploiting the H1B process saying “there aren’t enough applicants”. When are these companies going to be held accountable?
Never, they donate to get the politicians reelected.
The difficult part is going to be that new engineers are not generally who people think about to unfuck code. Even before the LLMs junior engineers are generally the people that fuck things up.
It’s through fucking lots of stuff up and unfucking that stuff up and learning how not to fuck things up in the first place that you go from being a junior engineer to a more senior engineer. Until you land in a lofty position like staff engineer and your job is mostly to listen to how people want to fuck everything up and go “maybe let’s try this other way that won’t fuck everything up instead”
Tell your family member to network, that’s the best way to get a job. There are discord servers for every programming language and most projects. Contribute to open source projects and get to know the people.
Build things, write code, open source it on GitHub.
Drill on leet code questions, they aren’t super useful, but in any interview at least part of the assessment is going to be how well they can do on those.
There are still plenty of places hiring. AI has just made it so that most senior engineers have access to a junior engineer level programmer that they can give tasks to at all time, the AI. So anything you can do to stand out is an advantage.
Answer is probably the same as before AI: build a portfolio on GitHub. These days maybe try to find repos that have vibe code in them and make commits that fix the AI garbage.
Answer is probably the same as before AI: build a portfolio on GitHub
You really think that using GitHub falls in the usual vibecoding toolbox? As in: would they even know where/how to look?
You think vibe coders don’t love the smell of their own shit enough to show it to the world?
Has he tried being a senior developer? He should really try being a senior developer.
He needs at least a decade of industry experience. That helps me find jobs.
No idea, but I am not sure your family member is qualified. I would estimate that a coding LLM can code as well as a fresh CS grad. The big advantage that fresh grads have is that after you give them a piece of advice once or twice, they stop making that same mistake.
a coding LLM can code as well as a fresh CS grad.
For a couple of hundred lines of code, they might even be above average. When you split that into a couple of files or start branching out, they usually start to struggle.
after you give them a piece of advice once or twice, they stop making that same mistake.
That’s a damn good observation. Learning only happens with re-training and that’s wayyy cheaper when done in meat.
Link?
It is not useless. You should absolutely continue to vibes code. Don’t let a professional get involved at the ground floor. Don’t inhouse a professional staff.
Please continue paying me $200/hr for months on end debugging your Baby’s First Web App tier coding project long after anyone else can salvage it.
And don’t forget to tell your investors how smart you are by Vibes Coding! That’s the most important part. Secure! That! Series! B! Go public! Get yourself a billion dollar valuation on these projects!
Keep me in the good wine and the nice car! I love vibes coding.
Not me, I’d rather work on a clean code base without any slop, even if it pays a little less. QoL > TC
I’m not above slinging a little spaghetti if it pays the bills.
AI used extremely sparingly is sometimes helpful to an experienced coder. “Multivac, generate a set of unit tests for this function.” Okay, some of these are dumb, but it’s easier getting started on this mess than just looking at a blank buffer. Helps get the juices flowing a bit. But man, you try to actually do anything with it, and suddenly you’re lost chasing a will-o’-wisp.
Oh man, I love ChatGPT for one thing in particular: “Hey chatbot, is there some library or standard library function for that very specific, yet still kinda generic thing I’m trying to do, so that I don’t have to write it myself?”
It does frequently give a helpful answer. That is, it doesn’t give me working code, but a helpful pointer to some manual where I can find good instructions for how to use the thing to solve my problem.
I will usually google that kind of thing first (to save the rainforests)… Often I can find something that way, otherwise I might try an LLM
Google has become so shit though
