Thread 107000580

What have you vibe coded lately, /g/oys?

>>107000580 (OP)
turning local sales into an RPG with map tracking and such to keep motivated during the boring social grind

How do you guys do it? Just open a chat window and begin talking about code and manually copy/pasting it to your project? Or is there some tool you let loose on on a folder and everything within is fair game for it to read and write to as you prompt about the project? Or something else entirely?

>>107000580 (OP)
ur mom vibecoded you into existence and thats why your dad left for a younger woman

>>107000595
The second. Things like Claude Code, Gemini CLI, Crush let you just say what you want and it begins creating files and writing code

>>107000589
What do you sell?

>>107000614
Cool, which is best? How does one get started? Is there some rentry or just OP pasta? I feel like we need some vibed pasta and then turn this into a general.

>>107000595
>>107000589 me
i have scripts that dump my whole codebase into my clipboard then i dump it into aistudio.
then i have some userscripts and bash scripts so i can click a button on an aistudio codeblock and it gets moved into the proper filepath on my system (all my files have the path as a comment for the first line)
its important to have a good prompt too, my "engineering prompt instructions" is kept the same, refined, and gets sent to aistudio everytime. pic related, he acts autistic like me.
>>107000633
whatever i feel like i can do to help, business owners like to talk about their business, you can literally bring a clipboard and be a business owner therapist. ask questions, find a problem, say "oh i know how to fix that, wanna pay me to do it for you?" boom business to business salesman.

>>107000639
>another aislop general
i think we need a /slop/ board instead

>>107000580 (OP)
the entirety of cy-x.net is vibe coded and it's the only functionally good vibe coded website you'll ever see
source: view-source:https://cy-x.net/
>NEW:
this is the only proof that i can give to you right now

looks and works well for being a vibecoded website too

too bad you cant code with AI without paying a subscription or purchasing the latest slopware to locally host

>session limit reached
>come back in five hours
>can't register more than one account anymore because
>too poor to purchase a bajillion phone numbers
i've been using free claude for years
i also verify it with grok just so i don't waste my messages on what could be trivial bugs that it hallucinated
i've been using the websites to do all of my coding
any advice? i'm admin btw

>>107000639
I haven't extensively used them so I can't really say. I'm currently using Qwen Code which is a fork of Gemini CLI that allows you to use any model by setting a custom endpoint and you API key.

You just install it, set your model and API keys in environment variables or config files (depending on the tool you use) and start it. Then you just tell it what you want
https://github.com/google-gemini/gemini-cli
https://github.com/QwenLM/qwen-code

>>107000645
>do you enjoy the time we spend together?
>no, fuck off, I'm just trying to get the job done with as minimal interaction with (you) as I possibly can
Kek.

>>107000660
>>too poor to purchase a bajillion phone numbers
AI Studio has higher rate limits. And creating Claude accounts is hit and miss, sometimes it doesn't ask for phone numbers, might have something to do with cookies or fingerprinting, how good a goy one is.
ChatGPT doesn't ask for one anymore. There is also GLM by z.ai which is pretty good for coding

>>107000662
Thanks!

>>107000658
we've needed one since sd1.5 came out.

>>107000730
we needed one even before that, for all the consumerslop

>>107000580 (OP)
I vibecoded some intentionally unoptimized C99 code with plenty of branching to practice refactoring with branchless programming. Trying to get it second nature to be able to write branchless code but also train my intuition to identify cases where NOT rewriting to branchless is better (i.e. the data is predictable so aggregate branch predictor successes will outperform the more consistent branchless operations which avoid prediction misses)

>>107000580 (OP)
my advice learn to throw your stuff up on cloudflare workers (used to be called pages)
they are very very powerful you can basically do whatever with them and the free tier is generous like 100,000 free "serverless" API calls a day
>Cloudflare Workers has a free plan at $0/mo (100k requests/day) and a paid plan starting at $5/mo ($0.30/million requests/month).
better than using vercel or netlify or nextjs which is a slave of vercel.
use all their tools. supabase is good too for databases. very generous free tier.
also look at all the vibe coders no one thinks are making anything!
>We Tracked Every Website That Launched in September 2025. The Data is Wild.
https://websitelaunches.com/blog/post.php?slug=september-2025-website-launch-data
>The final count: 368,454 new websites in 30 days.
>That's 12,282 websites per day. 512 per hour. 8.5 per minute.
mmm sloppa. wild.

>>107000782
Yeah, I put my image generator website on there

>>107000580 (OP)
wish these threads would catch on, its really cool to just build stuff really fast.
theres reason to hate the stuff sure but its up to the individual what they do with the tool.

What's the state of the art AI to write complex code? I've been having a good time with Claude, but I was wondering if there was something better out there.

>>107001341
probably the smartest are gpt-5 pro or grok 4 heavy, but they're expensive ($200/mo, $300/mo), slow, and not hooked up to the tools like claude code. so not that good for vibe coding

>>107001341
Sonnet 4.5 is the best at coding according to all the benchmarks

>>107000580 (OP)
I've been "vibe coding" (telling an AI what I want it to do exactly step by step for my application) a social media slop Expo app. I had been using Gemini 2.5 Pro in the browser and just copying and pasting. Then I moved on to gemini-cli when that dropped.
Finally, I tried Claude Code at the recommendation of a coworker. Holy fuck, this thing is so much better than gemini-cli, it's insane. It is able to infer what I want nearly first time, every time. Gemini always did I told it to, but missed some some basic shit about the code base, even though I fed it the whole file it needed. Gemini's context window can't be beat but HOLY FUCK Claude Code makes me cream.
I haven't done anything with custom agents or custom commands. I feel like I should be, considering I'm using these tools in the most basic way. I really want to try out GLM 4.6 as an alternative, because I keep hitting my usage limits on the Claude Pro plan.
I feel like a retard I wasn't using this before, and I envy the people who got to use it when it was unlimited before the price change rug pull.
Does anyone here have advanced custom agents or other cool ways you use this shit?

A simple, unreliable sentiment analysis tool where I download a bunch of images from a board, get their color histograms, and classify them.

>>107000580 (OP)
>Vibe Coding General
please fuck off, you people are not coders, you are just copy pasters

>>107001987
As opposed to?

>>107002012
Made with Claude?

>>107002017
Yes

>>107001946
how many prompts before you hit your usage limit on Claude pro? Also does it cut you off or just starts delaying responses

>>107001987
>the man riding on the carriage yells at the man in the car

>yet another general
Good riddance.

>>107001341
A formal proof of correctness & declarative pseudocode.

>>107000589
>>107000889
>>107002012
why do all AI slop program GUIs look like that?

>>107002012
https://rentry.org/qjtKMOLcwqfIWfDvyIWS

>>107002048
I'm not sure on prompt count. Going by the hour, it takes 3-4 hours of proooompting step by step instructions for me to hit the limit. When it hits the limit, I get completely cut off until a certain time. Usually, it's a time 3-4 hours later. I can see why people based their sleep schedule around these limits.

>>107003481
Trained on web"dev" slop

>>107001946
>Gemini always did I told it to, but missed some some basic shit about the code base, even though I fed it the whole file it needed.
im a gemini fanboy but usually it fucks up with my programming language because its just trying to do 'the old way' and just needs the error logs to fix it properly

>>107003481
it's called bootstrap, the most popular CSS library, created like 10 years ago maybe?
it was used so coders didnt need to design.
all you need to do is remove rounded corners and change some colors and it doesnt look like it anymore.
i could 'de-bootstrap' my site even more but since its just for myself its not very important how it looks, just that it functions.

P.S. OpenStreetMap rules! imagine if i had to use bloated google maps for this

the only things that i consider useful so far was a userscript that iterates over my bank term deposit simulator so I dont have to do it manually. later in I will get into some UI automation

>>107006100
I feel like it's just poorly integrated in every environment. It's missing that polish that other companies are charging money for. That said I'm a swe and actually know code and find it to be just as useful as Claude maybe just a little less intuitive. It just needs more detailed direction because it's assumption are usually reckless. I've been using the claude web chat as my planner and senior I can ask for advice and Gemini writes the code and I find this to be the best workflow to not be a slave to rate limiting, and it's free.

>>107000645
>then i have some userscripts and bash scripts so i can click a button on an aistudio codeblock and it gets moved into the proper filepath on my system (all my files have the path as a comment for the first line)
old video i have demo'ing it - doesnt have my new rice theme
and i have made some improvements. theres now just a little icon beside each filepath, i dont have to highlight or click the 'download' button in the codeblock. just click a lil icon.
i could automate the process of going through all of the code and clicking them all, to make it even easier, but i do like to have some control over the process and i do have to debug sometimes.

was playing around with OSM trying to identify certain types of intersections
claude worked very well, but the data from OSM sucks ass

>>107006795
>certain types of intersections
lol what? thats curious

>>107006463
yeah i have to make my own polish kek >>107006792
oh well, theyre my favorite out of all the others i've tried (all of them)

>>107006800
i was just wondering about how often a certain type of intersection occurs (and started with an easier task, e.g. identify and map all dead end streets). it ended up being that OSM had multiple entries and it was using the wrong one. And when it plotted the streets a big chunk were missing.

people always cope about AI by saying it can't solve novel problems, i figured this one would be pretty novel and it did a fine job, the data source just sucked

>>107006877
oh thats a pretty neat idea. OSM also has some public GPS data maps. you can see congestion, not as good as google ofc but its something. surprisingly there was a lot of data for my area.
maps are really fun to play with. i want to make a tool that picks a random location nearby for me to walk to, so i just click a button then go there, dont have to think about it and its a new spot each time. can slowly work up increasing duration/length.

I'm seeing if claude can fix a race condition with a mutex in some old abandonware written in c++ rn

>>107007520
doubt it

>>107000645
>bash scripts so i can click a button on an aistudio codeblock and it gets moved into the proper filepath on my system
inotify-tools does the heavy lifting, make it watch /Downloads/ folder and take action when a new file is downloaded that has the proper prefix
xclip for getting clipboard data

inotifywait -m -e close_write --format "%f" "$WATCH_DIR" | while read FILE; do
log_message "File detected: $FILE"

if [[ $FILE == ${PREFIX}* ]]; then
SRC="$WATCH_DIR/$FILE"
log_message "Processing file: $SRC"

CLIP=$(xclip -o -selection clipboard 2>/dev/null || echo "")
log_message "Clipboard content: '$CLIP'"

if [[ $CLIP =~ ^/[^[:space:]]+ ]]; then
CURRENT_PATH=$(validate_path "$CLIP")
if [[ $? -ne 0 ]]; then
log_message "Invalid path in clipboard, falling back to manual input"
else
log_message "Valid clipboard path: $CURRENT_PATH"

>>107007543
i'm optimistic, it rebuilt it with debug symbols and seems to be on the right track

It did suggest to me a better binary to use that worked perfectly fine so my actual problem is solved anyway.

>>107007662
>>107007543
>>107007520
it worked, though it may have reduced performance a bit based on cpu usage i saw.
i'm double checking that the output is identical to what lbzip2 gave me but the size is identical so i doubt there is any difference
total patch was about 50 lines i one cpp file to pbzip2

Bump

'mp

>>107010367
vibe code a bump reminder alert when it gets to page 10

vibe coding is a sham unless you're expected to churn out enterprise-grade slop and make up for it with 10x as much tests

>>107000614
Can I somehow direct it straight to an existing directory on my local machine and it will start doing stuff automatically?

>>107011090
Yes. Just cd into that directory and start the program there, then just tell it what to do

>>107000660
https://www.youtube.com/shorts/xngBF-sWsLg

Are any of you guys using anything that DOESN'T give all your data away to some big corporation? Google, OpenAI, antrophic, etc.
I really want to start using AI to increase my productivity as a programmer but not in a way where some jew can train his models on my data.

>>107012677
stop being fucking retarded
no one cares about your slop
Either way, if you use a paid Gemini account they cant use your shit for training, that only do that with free accounts.

'mp

I'm trying to make Cursor extract game mechanics from slither.io js file into a gym env and somehow train on self-play so I can beat all kids at this stupid game.

I think game mechanics are properly replicated but until I have a working visualization of the env simulation I doubt it's ready to leave training

>>107014803
its probably a bunch of bots anyways, thats what i'd do if i had an .io game. the games are so simply just fill it with fake bot players

>>107017029
you'd be surprised by how complex it can actually get, i havent seen many bots.

>>107017253
i know it was popular to do that for .io games, when they sorta blew up. and those games basically need players for the game to play, perfect for bots.

for those of you using automated processes like claude code, how does it handle projects with a substantial visual component? I work almost exclusively on graphical stuff, and the initial sanity check for most of it is "does it look right?" rather than verifying quantitative output. are the tools taking screenshots and running them through a captioner? or are they not suited for this at all?

>>107017811
it seems like where it would flourish is being able to go into an existing website and then convert everything, all the styles, all the settings into some modular easier scoped / more efficiently managed system.
then when you do need to go in and make adjustments, the refined system it has built out, makes the process for you, the designer, to fiddle and tweak it to perfection.
after its in this system state, you can then automate the process of having it always work off of the better organized baseline and then go like
>make it like a forest
>turn it girly
>i like the color red
and it will just shit out the different versions of it from the baseline, and set it and forget, you can come back and see 100 designs if you want.

is it fucked up that i like taking some shitty code spat out by ai and optimizing it to be decent?

>>107017953
its cool i guess, for practice right?
its not like youre beating the AI though its just a combination of humanities creations, so youre applying better patches to stolen code soup made by a bunch of people.
thing is though all the principles youre going to apply to it, you can just get the LLM to apply them itself. it just doesnt understand why it shouldve been implemented whatever superior way you think they should be.

Resources:

Websites:
https://aistudio.google.com/ (Very generous limits, biggest context size)
https://claude.ai/
https://chatgpt.com/
https://chat.z.ai/
https://grok.com/

CLI tools:
https://github.com/google-gemini/gemini-cli (1K requests/day if you connect your Google account, not always the case)
https://github.com/QwenLM/qwen-code (fork of Gemini-CLI, allows you to use any OpenAI compatible provider)
https://github.com/anthropics/claude-code (like Qwen code is compatible with OpenAI endpoints by setting environment variables)

>>107019407
you should check out jules too for googles gemini asynchronous coding agent. amazing what it can do, just give it git repo access.
runs tests, takes screenshots, etc.
15 'task' limit a day, but those tasks are not 'fix one thing' each task can be you implementing / fixing a large handful of features and functionalities

>>107019443
Seems pretty good, had read about it but never really gave it a try

>>107001946
I got a Claude subscription this month and just finished a small project using up all my tokens with an hour till refresh, fucking blown away and know that feel of feeling like a tard for not using it sooner

>>107017811
You can take a screenshot of a webpage, hand it CSS and HTML, tell it what you want and repeat and it will get simple stuff right

how many lines can you shit out from caude with a single prompt ?

insofar i only managed to shitout 5500 lines of templates reduced to a 1787 line i struggle to manage going above 2000 of functional code, separating task is random desu ,5500 lines was auto-discussion of two LLM ,every time i tried my factories ,it failed ,after one point the Coder stop actually giving a fuck a just fail but i'm talking about a LLM workflow that i started at 8pm and finished failing during the night (i went sleeping) i was also doing the same on the same API now i'm just using as much api as i can ,insofar even with a basic 60request per min you can pretty much etablish a good workflow and not paying a penny ,i'm gonna do that after i'm done with the free request with the first30days of Nvidia even two LLM discussing together with an average time of 5.86 second per answer finshing after 15 turns of conversation(bot1>>bot2>>1 etc) you calculate the number of api you get per min can go pretty fast quickly.

>>107020260
bro you are replying to the thread like we are an LLM, kek

>>107000580 (OP)
vibe religion when
https://youtu.be/ev05H6vvZw4

>vibe coded 4000+ lines of userscript in 1 file
>actual way to do it is to write in multiple files like a real human and use Rollup to bundle them
Why didn’t it tell me this

ive been working on a schizophrenic project for over a year now its so big.. it's gonna change the world

>>107000580 (OP)
I vibecoded a fully customizable study App with detailed metrics and project tracking

here's the showcase of my App
https://youtu.be/PgxaSWxPRFA?si=_D09AeEowp7gjHCS

repo
https://github.com/Chino-chan/Pomo-Doll

>>107021563
You can track project time separately and then see what time you've spent studying or doing that.

>>107021563
>>107021570
You can then compare with the past month or several months for example and things like that.

>>107021570
>reading sapiens

>>107000580 (OP)
qstyle editor. Doesn't work well the models suck at qt I wonder if I should spend money on claude model...

>>107021608
It's some dummy data lol

I vibe coded before there was ever ai. All you have to do is copy paste code snippets from reference manuals, read through reference manuals to look for keywords and library functions and write little programs testing how they work, and then you can vibe them into doing what you want. I vibed a lot of programs and tools and little games long before ai chatbots appeared.

I used glm/claude to help me add sadpanda support to komikku (ebussy manga reader app) by just translating the tachiyomiSY implementation. Idk if I'm going to try to clean it up and get it in, the developer was very reluctant to add nsfw support in the first place.

>>107023533
people would pay for access to this by the way, like people that would want to use this tool for the things i will be using it for.
i'd probably make more money just selling access to this tool and making it public rather than just local, and advertising it to the target market.
my point is, even though everyone can theoretically make this, because im leaning on an LLM to do it, most people wont take the effort or understand how the tool should function in the first place, that still takes your engineering/design experience to do well.
and then you still have to sell it anyways when youre trying to get people to use it, which is a barrier for most.

>>107021544
What is it about?

>>107025005
You're absolutely right.

>>107020260
>frog
>retarded
many such cases

>>107021563
>>107021570
>>107021575
cool stuff

>>107000580 (OP)
op(?) any interest in making a discord?
>>107010376
could make it alert the guild

'mp

>>107000580 (OP)
Vibe coded a feature to a program I use, and the program's author went ballistic on the PR because it's vibe coded.
Fuck these codemonkeys and their brittle egos. Who actually gives a fuck about how the code executes? Nobody without autism, that's for sure.

>>107031749
this never happened but i wish it did

I vibecoded a TCP/HTTP stack in a bootloader/kernel which runs in QEMU.

>>107000645
what do you call yourself when calling these businesses?

>>107033073
generic term would be consultant. freelance marketer. you dont have to be afraid of just presenting yourself as "Hey Im jeff", when youre starting out anyway the targets youll be able to sell to wont care about that kind of thing, just that youre real and genuine.
phones are just something that needs to be grinded out so you can get to the person that makes a decision and have him say yes/no to giving you money.
you want to do that meeting in person ideally, for trust, and so you can actually say "give me a thousand bucks". takes balls.

>>107033212
Yeah, sounds interesting. I downloaded a bunch of local business data from searchfunder the other day. Could i upload that cav into the system and have it populate all the locations on the map?

>>107033365
you could make your own front end using openstreetmap like i have.
im not making my thing for everyone just me specifically to grow as a salesman and grow my local network out again. i should though, i already mentioned ITT i'd probably make more money just selling that but its not as fun

its got some neat shit but all needs to be refined, its a WIP.
pic related. i have an LLM create random scripts that i can go through step by step based on scenario details i can set in the sidebar. good for stimulation and preparation.

>>107033501
lol yeah Id probably pay you just for what you have now for the same reasons you're using it, but alas I guess ill just try to replicate and tinker away to suit myself. thanks for sharing some screen shots

>>107033668
are you technically proficient? just keep it simple. if you dont know nodejs youll be able to achieve it with vanilla javascript and html. use localstorage to save data.
you just need the map. some sidebars with tabs. ways to import, track, add data (localstorage). checklists. change types of each business location, keep them categorized.
im totally overdoing it and procrastinating even with making it look this sexy and adding more features, but its because i want to actually use the thing everyday. its going to become a reflection of my soul.
you dont need to go that hardcore
start here id say
https://leafletjs.com/examples.html

'mp

As a manual tester, am I the enemy of AI?

Can someone vibe code up for me a tilemap editor that has per tile attribute selecting which palette to use and displays the tile with that palette?

>>107036957
sure boss ill get right on it

>>107000580 (OP)
a CRC-32 implementation
gpt-oss:20b one-shotted it

Is this real

>>107036859
Automation tester here, the rise of LLMs elevated my skills to a level where it's possible to make my own tools or patch existing ones so there's no reason whatsoever to not like it.
Moreover, our testing skills allow debugging the vibe-coding mess we generated more efficiently

Bumo

>>107000693
If you are paying for ChatGPT already, you can use Codex without paying extra. Just install codex-cli, or the vscode codex extension, and use your ChatGPT account to log in.
There's supposed to be limits, but I never hit them.

>>107019407
You should add Codex to the CLI tools.
https://help.openai.com/en/articles/11369540-codex-in-chatgpt

>>107000580 (OP)
How do I get into vibe coding? I missed the early adopter phase because I stopped ctrl+c/v many years ago, it was just faster to write code myself. But now I'm getting the impression that if I don't start using AI tools yesterday I'm gonna be completely fucking useless even compared to nocode retards who just know the right tools to proompt

what's the best way to get into vibe coding in Q4 2025?

>>107044585
get claude code and read the docs

>>107044585
I'm dealing with an operational 30k+ lines project right now where 99% is LLM generated, here's my advice and how you can implement it to B2B sales >>107041220

best vibecoding editor? I want to just give prompts and barely touch the code. fine if it's paid as long it's not unreasonable.

>>107044983
Save your money and use VSCode with Codex extension, Github Copilot and Claude Code. Use your money for the LLM services API.

Github copilot best bang for buck, followed by OpenAI, probably followed by Gemini. Claude Code is amazing but very expensive and from my experience you get almost the same results with ChatGPT 5 or ChatGPT Codex

Sirs

>>107000580 (OP)
Pretty much everything I do at work. I just ask Cursor to do it now and I watch a movie.