are there any solutions for running agents inside a secured container and having them report back on your with either finished work or questions? Preferably on BSD jails but I don't care so much about this

>>108305088
i just run it in a VM

>>108305074
Cool, how it works?

>>108305088
There are two options. One is creating a new user with limited permissions and running the agent on it and another is using a docker container.
You might want to do incremental backups of the folders the bot has write access to in case it decides to go ballistic.
But I've been vibecoding for a year now and the worst I've seen is the bot checking out an older version of a file and losing me the progress since the last commit.

>>108305107
By adding comments with the symbolic name for the register when it's assigned and then it retains that name until it's rewritten. There's also liveness analysis so if a register is read from last before a write the cells in that range for that register are marked as unused.
I also want to have a json file with explanations of what each symbolic name is.
And also it's not really an editor yet, only a visualizer. Adding actual edit capabilities will be a pain in the ass.

>>108305120
what about tooling to let it communicate with me or submit work? Do you just provide a locked down cli script that you mention in its instruction or what's the way to do this?

File: ai-coding-agent.png (340 KB, 2635x1872)

340 KB PNG

>>108305541
Yes. That's essentially it. You chose your framework and connect it to your model of choice through an API.
I built my own harness because I didn't like how any of the pre-existing coding assistants worked.

>>108305120
That’s cool as fuck. I’ve never used Claude or any AI for programming. Was it hard to code that stuff?

>>108305626
Yes, making anything non trivial is still hard. But it's much easier than without AI help. Like a 5x or 10x productivity multiplier.
The more your language and the tooling provides static guarantees and you are working on something that's very popular, the better AI will do at it.
It's very bad at things like assembly where there is very little training data and also the language provides no static guarantees at all and you only notice failures at runtime.
Especially with what I've been trying to use it for during the last month (streaming assembly - SASS) that is extremely brittle and things break because of almost impossible to debug race conditions.
That's why I'm focusing on using AI to build tooling to analyze and understand the code myself because the AI itself is very bad at directly dealing with it.

>>108305690
That doesn't sound like vibe coding.

>>108305690
Oh ok. Personally I’m not employed in the tech sector (im a neet) but I have some programming skills and a degree maybe I should try to build something using ai then

>>108305715
Yeah, people use it to mean a lot of difference things.
But at this point not just embracing the term would seem overly defensive.

>>108305736
I'm a consultant and even though I'm not a programmer agentic coding is so useful I'm violating company policy and using a code assistant with my own API key rather than the shitty chat web frontend I am supposed to use.

>>108305764
>im not a programmer
Do you have programming experience?

>>108305751
I mostly see it used spitefully, so I think defensiveness is called for.

>>108305774
not professionally

>>108305811
The best one to use for free is probably opencode because it comes bundled with some free APIs. And go to OpenRouter and load it with 10 dollars to get access to the free models there too.
I started vibecoding by copy pasting from ChatGPT though. You don't have to use agents at first.

>>108305074
I was vibe coding, but it made me realize people don't pay for anything these days, they struggle with youtube premium pays.
I am pretty sure 99% of vibecoders make 0$ from their product.

Mine is a MUD that started with code from 1996. It's now grown far beyond what I expected AI to be able to help me write. I don't even write any code anymore, I just describe the features I want in detail and my LLM writes the code perfectly because it's become an expert on my specific project. It's freeing and feels like some Star Trek shit where they'd program complex holodeck scenarios by just describing what they want to the computer. So now I have this awesome text based MMORPG. Not open for players yet though, as I'm still working on making sure there's enough skills and spells to make each class feel truly unique and give them enough stuff that they can fill more than just a singular role; I want warriors to be able to be DPS and not just a tank, mage as CC/support and not just DPS, etc

>>108305929
Cool. Was the code from 96 also written by you or is it a fork of something else?

File: sicp1.jpg (98 KB, 600x439)

98 KB JPG

>>108305074
Kill yourself you soulless fucking nigger. Garbage thread, Claudefaggots are spamming my board.

>>108305966
>le SICP meme
nocoder zoomie detected

>>108305946
Little of both. It started as a "Rivers of MUD" fork back then, and I walked away from it for a while, then started working on it again circa 2003, and added to it here and there over the years, but in the last few months since "vibe coding" I've gotten more done than in the last 23 years.

>>108305983
Zoomers don't even know what SICP is you fucking retard.

>>108306027
Im a zoomer and I got tricked into reading it. I don’t like lisp though

>>108306011
Damn. I've never played MUDs as they were before my day but I've been somewhat curious about them and dabbled a bit in the adjacent genre of text adventures.

>>108306027
Kek. Who are you trying to fool bro. That one and The C Programming Language have been spammed to death throughout the Internet.
What non vibecoded projects have you worked on? Share with the class.

so uh anyone have any ideas for reverse engineering compiled .r binaries back to OpenEdge ABL source?

I need to infer schema and business logic for this stupid fucking OpenEdge AppServer based shit but no success at all finding a way past the appserver to get a look at the database behind it all directly, so im stuck with just temp-tables and trying to make educated guesses on anything beyond. Ultimately need to migrate out of a MSSQL Server 2017 db to it, and both databases serve the same practical purpose but clearly leverage inputs differently.
It's been a fucking nightmare.

>>108306168
Have you tried asking your agent nicely?

when codex on linux

>>108306307
???
It's multiplatform bro

>>108306368
i meant the codex app
anyways looks like it's coming https://openai.com/form/codex-app/

>>108305074

I'm just vibecoding my own RP app because I find Silly Tavern too out of date for my own needs. I'll release it open source once when its done, so local people can tinker with it. I also don't give a fuck about slop, but I'm testing the shit out of it for more platform coverage. I'm also starting to some egregious faggots out there selling LLM wrappers for the same thing I'm building, most likely vibecoded, too. Top kek. Enjoy the race to the bottom.

>>108305074
>>108305150
Pretty cool. Are you going to share it with world?
The only thing I bothered to use AI to make was done with Copilot through visual studio. Even that was simply modifying an existing project to add back lost functionality from over a decade ago.

>>108306450
>Pretty cool. Are you going to share it with world?
Hell no.
But I am planning to release the stuff I use the editor to make, in binary form (no source).

>>108306406
Huh. Didn't even know that existed.
I am planning to add web functionality to my own code assistant though.

File: Screenshot From 2026-03-0(...).png (579 KB, 2329x2149)

579 KB PNG

>>108305074
https://github.com/orlfman/FFmpeg-Converter-GTK

next up is chapter extraction.

>models used
human brain
claude opus 4.6 extended (two pro accounts with seeded extra usage balance)
super grok

>>108307479
lol this shit is insane now. you're approaching blender level of features.

>>108307479
nice! why do you use extra usage rather than max? for the extra context?

File: screenshot.png (63 KB, 1108x681)

63 KB PNG

Is anybody interested in this thing? It's basically just desktop automation using codex / claude. I'll post the code in a bit I'm almost done like an alpha version basically.

>>108307527
I've been messing with desktop automation too.
How does codex do the clicks on the screen?

>>108307551
It's in go so I can just use this lib:
github.com/go-vgo/robotgo
Seems to work pretty well for me on Linux.

>>108307527
>>108307551
>>108307567
https://github.com/nickbarth/closedbots/

Here it is. You can just clone it and run make and that'll trigger a build and run. I was just trying to test desktop automation really. Turns out its pretty rough still haha. If you have any cool ideas feel free to take the code and do whatever. Go lang turned out to be pretty nice for this sort of thing.

File: 958.png (36 KB, 940x744)

36 KB PNG

>>108305074
I vibecoded a database-2-excel exporter for my workplace.

>>108307507
thank you very much. the sky is the limit. i'm trying to make the ffmpeg front end of my dreams lol.
>>108307516
yeah, extra context

>>108308125
you're a data scientist now MY BOY

>>108308170
nah, just a silly sad database manager who has no time for these constant disctractions. I mean...how am I supposed to browse /g/ meanwhile?

File: surf_maid.png (2.83 MB, 1024x1536)

2.83 MB PNG

This vibe coding thread turned out okay. People are actually making things now. Here's the links from the last dead thread:

AI Coding Agents / APIs
• Codex CLI - https://developers.openai.com/codex/cli/
• Claude Code - https://code.claude.com/docs
• Aider (CLI agent) - https://github.com/aider-dev/aider
• Open Code (CLI agent) - https://opencode.ai/

Editors / IDEs
• VSCode - https://code.visualstudio.com
• Neovim - https://neovim.io
• Cursor (AI-native IDE) - https://cursor.so
• GitHub Copilot - https://github.com/features/copilot
• Codeium - https://codeium.com

I wrote an evolutionary algorithm in python and asked chatgpt to integrate CuPy. It did and now it runs 20x faster

>>108308238
>20x faster
chatgpt or algo?

File: 69c.png (296 KB, 555x504)

296 KB PNG

>>108308249

>>108308224
OP here I tried posting your template with an anime picture before and got 0 replies

>>108308153
How do you not end up spending way more on the extra tokens than on the accounts themselves? Or do you? It must be like dollars per message no?
Do you feel like the long ctx makes a big difference?

>>108308238
What's the algorithm for?

File: 051f55ef6b004e48d0ccf5eda(...).jpg (204 KB, 2309x2309)

204 KB JPG

did you try vibe coding mobile apps? because i mostly see people vibing some power computer (desktop) apps.

i have some ideas, like smartphone voice control, like you say "<catchphrase> flashlight" and the app turns flashlight on or off. this kind of remote control. which stack you reccomend and which music tune i should vibe on?

i mostly on some fast beats but maybe should slow down

>>108308421
Do anything that can be done as a webapp a webapp probably

>>108308317
How can we bring you over to the dark side?

>>108308438
>t. having a stroke

>>108308322
Given a frequency response, find the coefficients of a FIR filter to match it as closely as possible

>>108308421


wanna see some ?
what would you like ?

>>108305074
I'm vibe-coding the next gen air defenses for the US govt. As you've probably guessed, time is of the essence, so I got three Claude instances crunching that shit.

Vibecode my housekeeping python script. Fuck windows for its shitty shell support and python’s convoluted way of managing directories.

im vibecoding a minimal ssh implementation every 12hours by using opencode free tier and returning to the old workflow of copy pasting it now feels massive turd.
i guess i kepp contributing to the Big Pickle data collection system

>>108305074
Nvidia ASM is probably the only thing I'd look at vibe coding for.
All other ASM is too comfy to give that to an AI. I want to enjoy that.

>>108308929
Damn, the US is sure to win any air engagements then. GLM Is trash.

>>108309583
Yeah, for sure. Especially when you let the agent interact with a browser.

>>108308224
Yes this kind of threads give me so much inspiration about what to build. Just the other day I was thinking about a simple GPG clone with sane UX and minimal cipher support.

I'm vibe coding a cheat table for a video game. At first it started sperging out about it being against the TOS of the game and whatnot and refused to do it, but I found a way.

AI "safety" is retarded.

It's hard to figure out what I even want to vibe-code. Either it's not worth the trouble or too difficult to ask AI.

>>108307507
its jut gui for ffmpeg commands, stop praising this retard, he's done nothing original and crazy

>>108308312
you can't use another account when you are still having claude finish something but he's about to hit the limit. its why the extra usage is there. for the moments where he's almost at 100% (either weekly or session) and want him to finish. stopping, and using another account is starting all over again.
i use one account until i am maxed out for the week, and then move on to the second account. right now with my current usage patterns, i need two accounts to have enough for a week.
even with the extra usage, still comes out cheaper than paying for the $100 a month account.
then with super grok, i am able to offload a lot of the trivial stuff off claude. claude handles the hard, difficult stuff. grok handles everything else. including being my general search engine. i already have an x premium account so super grok is included.

File: 1745026594077839.png (374 KB, 1920x1080)

374 KB PNG

I began vibe coding a browser-based DAW. I have no experience with coding but it does everything I want it to. All in a single HTML file.

>>108310439
... what the fuck... that looks insanely nice hahaha. fuck fruity loops!

>>108310393
>its jut gui for ffmpeg commands, stop praising this retard, he's done nothing original and crazy
all applications are a GUI over commands. we're all a retard to someone. my mom called me a retard just yesterday.

>>108310439
thats super nice

>>108310439
ummm /dpt/ bros? what do we have to say about this?

>>108305107
Good morning saar

>>108310393
The beautiful thing about AI is we are no longer bound to "developers" and the "developer community" gate keeping what can and cannot be made. Everyone's ideas can now be made a reality. Regardless of what people say. No longer are we shackled to ancient, relics of the past to those of the likes of you.
If you are this mad over this anon project, be prepared for the all the heart attacks you will be facing as time goes on as more projects you hate, will be made.

Anyone using local models for coding? I'm starting a new job and getting a Macbook Pro with 64GB memory. I'll have access to claude, but was curious if there was any benefit to supplementing claude with smaller local models

>>108310794
>The beautiful thing about AI is we are no longer bound to "developers" and the "developer community" gate keeping what can and cannot be made
Nothing was ever gatekept, you were always free to go and teach yourself.

>>108310817
>Haha, nothing was gatekept! All you had to do was go to it yourself (but we know you won't, because you can't)
Now we can. Keep seething. No long are we chained to assholes like you.

File: Screenshot From 2026-03-0(...).png (224 KB, 1672x1169)

224 KB PNG

>>108310393
if you hate my ffmpeg front end, be prepared for my next project once i am done.
a front end for yt-dlp. i already vibe coded something up in qt6 in early 2025. i plan on re-doing it in gtk4 / libadwaita from the ground up.

pic very much related.

>>108310961
based

>>108310817
Oh brother, you wont be satisfied with literally anything else once you try Claude.
The closest thing is Kimi K2.5 and you will have to spend half a milly on hardware to get reasonable performance for agentic use.

>>108311033
>Oh brother, you wont be satisfied with literally anything else once you try Claude.
I use Claude at my current job and I'm impressed with it. My current job switched from the pay per token api (or whatever it's called) to the enterprise plan. I haven't had a day where I run out of tokens, but I don't really do much work at my job. Honestly I was just looking forward to something to tinker with, but if there's no benefit to it.

>>108305074
>vibe code me a syllabus.md with course scope and overview of the classes for learning to ai, prompt enigneering, the whole shit
>looks it over
>seems good enough
>now vibe me the course pack for chapter 1, let's get this started

>>108311083
Anthropic's big goal for 2027 is approaching firms to sell them on replacing 90% of their programmers since Claude is generating all their code already. Hope you're in the top 10% bro

>>108311083
I hope eventually there is a way to customize local models for personal use. That's why I'm working on the GPU stuff in the OP.

>>108311240
>Hope you're in the top 10% bro
Nah I'll be fucked, I've always been a 7/10 SWE, but I'm good at talking to normies. I'm already connecting with people outside of tech to figure out high touch consulting ideas for different industries.

Time to shill the hottest vibe coded project lately
https://github.com/Nonary/Vibepollo

>>108305690
After you build the analysis tool, is there any way to get the AI to use it to analyze and fix its SASS code? Or is the plan to generate SASS code with bugs and then generate tools so you can track down and fix the bugs yourself?

This thread is giving me anxiety.

>>108311396
Why?

>>108311396
Embrace the vibe

Codex cannot get html to have a monthly date picker box but the date field is yyyy-mm-dd, even with extra high reasoning. lmao.

File: 1772779388942062.png (328 KB, 1577x912)

328 KB PNG

>>108305074

>>108305088
apparmor

it's wayy too tiresome to change FS permissions like what this guy suggests >>108305120

it's also lighter than a vm

>>108311385
No, the AI can't reason about the stalls and memory barriers and shit.
The idea is to use the AI to reverse engineer compiler generated code and then let it make changes but with every change understood and reviewed by me and also using vibecoded static analysis tools to cut down on bugs and nonsense edits as much as possible.

>>108311396
Good, that means you are almost as smart as Opus 4.6 :)
>>108309653

>>108311553
I really doubt figuring out apparmor shit is easier than creating a new user and doing sudo su and working within the new user's home folder.

>>108311779
I use users already with custom firewall rules to isolate pirated games, so that workflow would be ideal for me

Yahoo clanker just spent multiple tries sending me after an imaginary setting to fix my email search on iphone, I gave up when I realized obviously and tried to sign in on desktop and the server won't let me.

>>108305088
vscode has dev containers, which is just a docker image that you can run a setup script against, I use it with cli but it can open ports if you wanna run a ui based agent via browser or something.

File: 1600109466417.jpg (41 KB, 798x644)

41 KB JPG

>>108311811
>Yahoo clanker

can anyone teach me how to vibecode, please? I'm not a programmer. I have a Claude account but that's it, it's the free tier.
is vibe coding just telling claude to do stuff in the chat window?

>>108312831
ya you can start like that. in that window you can ask it to make you websites. if you want it to include javascript ask it something like:
"make me a website with inline vanilla javascript for a calculator app" and it'll make a calculator

>>108308224
>https://code.claude.com/docs
these are the level 2 type shit. once you understand some of the code you can get the command line tool and run it and talk to it the same as you would in the browser and ask it to build you stuff. the only difference is its running on your computer locally. mostly for noobs just stick to website stuff because you can just click on the HTML file and see it run locally in your browser. if you're willing to install and download a bunch of shit you can get a language like python or C++ or whatever and ask it to do apps. but even for games probably all you need is javascript and website junk.

File: Screenshot_20260227_194241.png (727 KB, 1618x971)

727 KB PNG

I let claude reverse engineer the car format of Assetto Corsa EVO so we can load them, change their colors through values and save it again

>>108312926
lol hilariously based usage. though its crazy that's not just a regular ass in game feature.

File: file.png (159 KB, 1332x433)

159 KB PNG

>>108307479
is it really good? Don't want to download a anon's dolphin porn.

>>108312973
The emojis are... Interesting

I made a full locally-hosted web app version of the board game Dominion. I was able to get Claude to read a 200 page rules PDF and understand how the whole game system works. It supports multiplayer, CPU opponents and agent AIs to join and play a game together. Its fun as fuck but I cant post it online.

>>108314045
>he doesn't know the meta to becoming an e-celeb is getting sued by a corporation

how do i properly set up claude code to not rape me

>>108312973
its pretty good. its not machine level obviously. but its fairly accurate. i have done some changes to it to enhance it. moved from two quadratic to three. so it does three data points to calculate. tests a low crf rating, mid, and high to hone in.
from my testing, its been very accurate at hitting the target file size limit for 3 - 6mb. but you have to use an operate video. its impossible to get a two hour long movie down to 4mb for example. but videos ripped off twitter? tiktok? that are 20 seconds - 3 minutes? its been really rock solid for me.
i have expanded it to all codec tabs now so svt-av1, x265 also support it. and did more maths for higher file sizes. ive been getting good results getting 600mb pron videos down to a target of 100mb and look great.

please give it a try. i love feedback. stuff like the entire crop and trim tab was because an anon asked for it.

>>108309865
If it's too difficult to ask so you need to jot down all your thoughts into a long document, just get your idea onto a file, then have it analyze it

>>108314001
grok and claude like emojis because they add easy icons.

https://github.com/affaan-m/everything-claude-code

>>108314454
>operate video
appropriate video*

actually thinking of using LLMs to help me reverse engineer or port old games to 64-bit Linux ideally using vulkan somehow
idc about old games that much but would love to try out Syndicate Wars, Homeworld, even porting Dark Souls to vulkan with proper hw accel

File: sass-ide.png (1.18 MB, 3840x2120)

1.18 MB PNG

Added a couple more features to the viewer.

>>108314551
The realistic way to do it would be to use AI to improve Wine. You won't port old Windows games to Linux. Porting a single game would be a multi year project even with AI.

File: Screenshot From 2026-03-0(...).png (208 KB, 2215x1687)

208 KB PNG

>>108314454
new build now has a new toggle to disable audio if you have auto conversion enabled when using the smart optimizer. so it strips out the audio for stuff on boards that don't support audio.

>>108314643

[Smart Optimizer]  Smart Optimizer Recommendation
[Smart Optimizer] 
[Smart Optimizer] Codec:          X264
[Smart Optimizer] CRF:            33
[Smart Optimizer] Preset:         medium
[Smart Optimizer] Two-pass:       enabled
[Smart Optimizer]   Bitrate cap:  321 kbps
[Smart Optimizer] Est. size:      4351 KB
[Smart Optimizer] Content:         Mixed content
[Smart Optimizer] Confidence:     90%
[Smart Optimizer] Size tier:      Tiny (≤25 MB)
[Smart Optimizer] Audio budget:   stripped
[Smart Optimizer] 
[Smart Optimizer] ── Strategy: Tiny (≤25 MB) ──
[Smart Optimizer]   Audio: stripped (no audio in output)
[Smart Optimizer] 
[Smart Optimizer] ── Content ──
[Smart Optimizer]    Mixed content (confidence: 0%)
[Smart Optimizer]   Audio: stripped — entire budget allocated to video
[Smart Optimizer] 
[Smart Optimizer] ── CRF mode (quality-focused) ──
[Smart Optimizer]   CRF 33 / Preset: medium
[Smart Optimizer]   Estimated: ~4351 KB
[Smart Optimizer] 
[Smart Optimizer] ── Two-pass mode (size-guaranteed) ──
[Smart Optimizer]   Target bitrate: 321 kbps / Preset: medium
[Smart Optimizer]   This mode guarantees the file fits within the target.
[Smart Optimizer]   Quality is determined by available bitrate, not CRF.
[Smart Optimizer] 
[Smart Optimizer] ── Calibration data (3-point quadratic) ──
[Smart Optimizer]   CRF 18  34497 KB (full-length estimate)
[Smart Optimizer]   CRF 25  16177 KB (full-length estimate)
[Smart Optimizer]   CRF 32  6882 KB (full-length estimate)
[Smart Optimizer]   Model: ln(size) = 11.9488 + -0.0655·CRF + -0.000994·CRF2
[Smart Optimizer]   Preset efficiency factor = 0.72 (medium vs ultrafast)
[Smart Optimizer]   Sample coverage: 31% (4 × 8s segments)
[Smart Optimizer]   Source file size: 6.0 MB
[Smart Optimizer] 
[Smart Optimizer] Auto-convert enabled — starting conversion…

File: Screenshot_2026-03-07_02-39-41.png (84 KB, 1366x768)

84 KB PNG

Basically I just wanted this:
https://github.com/wafarifki/musicwithlyrics

But for some reason it would not work whenever I downloaded it.

AI made a version in minutes. I modified the CSS. And here I have it. Basically DIY karaoke in a 1page html template for me to sing some songs.

>>108305088
I've used bubblewrap but it doesn't feel strong enough, no process isolation and running under Nix just makes the agent confused. I just run it in containers now, slowly working on something to make it nicer. I tried writing said nicer layer with agents but it kept nuking its own config folders and I got bored logging in every fifteen minutes.

>>108314724
thats pretty cool anon

>>108314886
thanks anon

It's great being able to spin up a Github codespace and tell Opus to add a feature or fix a bug on a project

>>108305074
>Post your vibe coded projects
So you can steal them? Nice try, Sven

File: 1761944831150708.png (2.35 MB, 1144x2048)

2.35 MB PNG

What's your favorite CLI? I've been using Opencode with Opus and have been trying Codex lately

I've been vibing a P*lymarket strategy tester on 15m c-markets and turns out random walk + price / time gate outperforms tuned strategies

Btw I still don't understand GitHub or how to use gits. I just tell it what I want changing until it does it.

Regardless of the question how effective it is, do you guys actually like vibe coding? I feel like it's way less fun and fulfilling than normal programming, I just have to do it for my job.

>>108314218
Claude has almost no permissions as a default.

>>108305107
That's the best part, it doesn't.

>>108306556
What are you working on to need to edit GPU assembly? Sounds interesting.

>>108316362
>do you guys actually like vibe coding?
Honestly, I kind of like it at work because
1.I can pretend to be busy
2.I hate how everyone sets weird deadline expectations when you're writing code,. Just prompting AI makes it much less stressful.
But at the same time, vibe coding kind of feels like you're still working, so if you program for fun, it feels weird

Does anyone else prefer OpenCode with OpenAI over Cluade Code?

>>108316680
no one prefers open code. just use codex. codex is better than claude.

>>108316481
Trying to make a fast LLM inference and fine-tuning engine

>>108316362
Yeah, I find it more fun than traditional programming. Guess I was more of an ideas guy after all.

File: 4chan_Card.png (74 KB, 1175x500)

74 KB PNG

Here's my new vibecoded project

File: file.png (61 KB, 378x178)

61 KB PNG

Built this always-on-top mini music visualizer+lyrics display that sits in the top-right corner
Thinking of adding a spinning disc on the left for the album art

>>108316362
yes just make cool shit i am making stuff I would never be able to amke on my own because it's so complex

>>108316362
i am having massive amounts of fun. its the great equalizer. i am able to do so many things i couldn't before and only a few could do. no longer have to beg others. i am able to be self efficient.

>>108316938
>i am able to do so many things i couldn't before
like what

>>108316939
>>108316939

So how are you guys using this stuff? I still have a very boomer ass view of these AI things like they're just a website with a chatbox where you can prompt stuff. What are "agents"? Are you using some sort of API implementation? This is all mostly for paid models I assume?

>>108305074
vibe coding is too cringe to do.
Best I can do is ask ollama a question/for sample code and modify that by hand.

>>108317090
I don't think vibe coding is possible without airgapping.

>>108317099
there's no other way to ensure no one is watching you 24/7.
But that also extends to coding without showing it to anyone, you have to airgap to actually code without showing the code to random woman these days.

>>108317051
agents are just taking the output of the chatbot and giving it to a parser so that the chatbot can issue commands to interact with other programs (the syntax to use these commands is explained to the chatbot in the system prompt)

>>108317143
Ok and there's resources for this stuff? Like can I download these agents or something? I'd be interested to experiment with this stuff but I have no idea where to start. I tried to get Claude (free version) to make a little 2d game for me before but it fucked up a lot and it got worse the more I tried to make it fix things, so I still have this impression of AI being godawful for coding which it clearly isn't.

>>108317194
Yeah, just install opencode, it has some free models out of the box.

>>108310439
> All in a single HTML file.
You'll want to change that sooner, rather than later. Your neat little app will be ripped apart by the agent 3-5k LOC in.

Started using Claude to see the hype, ask it to do one (1) dashboard and it bills me 5 fucking bucks for 500 lines of code?

Is this the price or am I doing some wrong?

>>108318111
>Checked
It consumes it fast, but shouldn't be that fast. If you just setup the project, 60-70% of the tokens were probably spent on setting up the project in the first place (React or whatever).

>>108311396
Just don't let it atrophy your memory and ask the LLM to teach you every now and then

>>108318111
>>108318151
you're supposed to get the monthly plans, paying for api is absolutely not worth it

>>108318267
True, but even with paid plan it consumes it fairly fast. Subbed a few days ago, and consumed my 5 hour quota in less than 30 minutes after a medium sized refactor.

From my experience, the setup phase and refactors consume the most tokens. Minor fixes and additions obviously stretches it a bit further.

>>108318323
yeah, the $20 plan of CC is pretty useless and even with the $100 plan I can run out of 5h usage in 1 or 2 hours.
the best option cheaper than $200 might be the kimi code plans since kimi k2.5 is from what ive seen pretty good.
it kinda depends whether you want a plan with a lot of usage with a shitty model (for this the GLM one might be the best) or less usage with a smarter model.

>>108316724
Yo this is sick and based. I like the minimal style but you do you album art and shit sounds neat.

>>108305074
Extremely cool. It's very heartening to see AI empower people so much! However I have one question: Why are you even looking at assembly directly when AI can do this for you, better than you ever could?

I don't know jack and fucking shit about coding and very little about scripting but Claude helped me at least with this sort of thing:
https://github.com/viktory-mkii/atomic-update-trigger
I did a fair amount of debugging with this shit, my PC went through some rough cycles but I got exactly what I wanted out of it. Safe to say even my dead grandma can do this now.

>>108305074
>Post your vibe coded projects.
>A typing tutor for Plover that actually teaches you the letters and the most commonly-used English words so that you can get up-to-speed quickly
>A system that listens to my local radio and Shazams the song clips that they use at the top of the hour because they won't post those online for some reason
>A rhythm game that is entirely sound based, no visual UI (currently dead in the water because it turns out that games are hard to vibecode)

File: wrong-barrier-guilt.png (453 KB, 3829x2038)

453 KB PNG

>>108318483
I tried but it can't.
It doesn't know enough about the ISA to be able to find and correct bugs or modify the file without introducing bugs, and in-context learning doesn't seem to be enough to overcome the lack of built in knowledge.

>>108318571
assembler is probably a bad level for LLMs to work at. too much implicit state tracking neccesary like which registers are available, which got clobbered, intermediate results, etc.

maybe the same class of problem as having to count how many rs in strawberry, since its all tokens from the LLMs pov.

>>108318640
I haven't tried it on other assembly languages.
Nvidia SASS is particularly insane and undocumented.
I don't think it as anything to do with the strawberry problem. I think it would do just fine if there was enough training data.

File: file.png (116 KB, 1282x1290)

116 KB PNG

i switched to linux a couple months ago and vsts are fucked so i just wanted to generate a custom one.
it's still not totally done but it's getting there.

The main thing is now my "Mic" output on my desktop now is running through basically a premixed and mastered signal where I can combine my electric guitar I plugin with an amp sim, 2 synths, a drum machine as well as other digital instruments.
They all go to my mic output and are mixed/eq'd and mastered and everything. It's always running so at any time I can start playing instruments or work on music, and to record I can just record onto vocaroo or something because it just uses my mic output.

The other part the player is basically, currently it's jsut set up to play pre-written midi files to the instruments.
To generate those I have a custom DSL I wrote for writing songs which is converted into the midi, so I need to set up some ui to modify the files, they are just text files so should be easy to customize.

But this is just built into my taskbar now so I can just run it, click one of the songs and play guitar and sing along with it and play with the knobs on my synths/drum machine to customize it.

the main idea was to help with actually writing music so I need to finish up the interface for modifying/editing the songs but even just where it is so far it's pretty nice to have.

I'd also like to add more configuration to the effect/mastering chains but but ideally they not need to be adjusted that often.

vibe coding linux desktop/ui stuff is really fun

>>108317051
As the other anon said agents can use tools, this basicaly eliminates both ends of the copy paste workflow.
For Claude and Codex for instance, I open them in the terminal inside my project and they can read the code in that folder, so no need to copy. They will also use grep, fuzzy search and so on to find the right lines of code, so they don't have to read a whole project and overflow their context window. After that they can also modify the code, you can set permissions to do it automaticaly, or approve every change manually, so no need for paste.
Claude can basically use any CLI Tool, like git, gh, aws cli, python, cmake, make...
It can also start your program, read the logs and fix bugs based on the logs etc.
I never found them useful before Opus 4.5 and Codex 5.2 but now I use them every day. When I use Sonnet (the weaker Claude model) I'm still usually disappointed.
It also helps to write code in certain ways, e.g. tests first, lots of logs, benchmarks, anything that will give the model some sort of objective feedback. A lot of the prompting advice isn't that useful and imo just a way to sell courses, but common sense software engineering techniques help a lot.

>>108319302
Yeah I think my biggest struggle would be to have it code a bigger software project. Like I just have no conception of how to go from "I have an idea for a program/game" to finished and polished product. In my mind everything would fall apart about 30% of the way in because the AI would keep breaking stuff that it made before and I don't know enough about best practices and software development to steer it correctly.

>>108319328
just dive in and see what the models can do now - they can do a lot more than most people expect.
they do have issues (architecture, cheating, overly-complicated code etc), but it's fairly reasonable now to expect that these problems will be mostly solved in 12-24 months - you don't want to wait to start then.

>>108319328
I think at the moment at least that's still a valid concern for big projects. There are people with no technical skills who have created some bigger projects with AI, but one thing I usually notice is that they often just brute force it with mass tokens. Just by sheer luck, the AI will get it right eventually, if you run multiple agents for a long time, they will find a way sooner or later, but it can be pretty expensive.
On the other hand, you could learn some of those concepts. There is at least almost no reason to learn boring syntax when using AI, you can instead learn those more abstract concepts, like testing, modular design, how modules should interact etc.
The AI can also write the tests for you, but I still have to remind it often and when I don't give any feedback, the tests are often pretty bad. It just implements something, and then it writes a test that expect the outputs that the function it just wrote produces, so any bugs or any bad design just gets locked in place with tests.

I want to vibe code my dream game in unity3d. Are we there yet?

>>108319718
yes but you cant afford the tokens

>>108319381
I think it's fine to just try it, and the models are definitely not bad, I use them every day, but I disagree with the "being left behind" view.
The way I see it, coding agents will only get better and easier to use. They are already easier to use than traditional programming because it's all natural language. It's not like with a programming language, e.g. someone who used C++ since C++89 will find it much easier to use C++23 than a complete beginner. Even if some of his ways of using it might be outdated, he still has a decades long head start.
With coding agents, many of the prompting techniques are already outdated and just use context unnecessarily. The best way is already usually just to use simple language. The best CLAUDE.mds are short and only cover project specific conventions, Claude can use sub agents on his own now etc.
So anyone starting to vibe code in a year won't have that much catching up to do.

>>108319328
>so they don't have to read a whole project and overflow their context window
the way you design your project matters a lot. make it very modular. that way you don't need to feed the AI every single thing.
but that does require already having somewhat of a programming background.
my project for example has 42 files, about 10,000 lines of code total. but i only feed the ai between 1 - 6 files at a time because of its modular nature.

>>108319732
Why doesn't somebody vibecode a token miner?

>>108319782
kino idea desu

>>108319718
Depends on you dream game. AI can't really do everything related to 3d models yet

>>108319718
>Claude, vibe code me a modern WoW classic
>Claude, vibe code me a true successor of Pokemon Red and Blue for a 35 year old millennial

>>108319782
what's a token miner?

>>108319732
only retards pay per token for agentic stuff

>>108319774
>Core
Stable abstractions, interfaces, dependency injection points, event buses, plugin systems, etc.
>Everything else
Plugins/modules that "connect" via those interfaces.

Spend time on the designing the core first. Requires a lot of planning and a solid idea of what you want to do, but its excellent coding practice and you can easily get away using AI without having to burn through context windows / tokens.

>>108319907
And yes, you can have AI build you that core foundation. You just need to have an idea of what you want and how to build it, for you can ask AI to build it for you.
Don't go asking AI to "build me a gameboy emulator." Instead, plan, and ask AI to build you X, Y, and Z that overtime, builds into that gameboy emulator and makes it extremely easy to maintain and keep AI building it.

>>108319859
Mines AI tokens.

https://developers.openai.com/api/docs/guides/prompt-guidance/

File: Screenshot-Preferences.png (160 KB, 1974x1134)

160 KB PNG

>>108307479
really big update.

File: file.png (1.3 MB, 1854x1459)

1.3 MB PNG

i made a deathmatch shooter on epstein's island. >15 characters each with over 1,000 lines of dialogue from elevenlabs.
>web3 integration
>golf carts, helicoptor, gunboat, jetski
>lots of weapons
>alien invasion mode
going to add a battle royale model where you jump out of the lolita express later.
www.epsteinarena.com

>>108314598
>Porting a single game would be a multi year project even with AI.
Why, prompt the code until you make a similar gameplay and logic, then rip assets and prompt to redraw/recreate them to avoid legal issues.

>>108321499
That'd be even harder than just trying to port the binary code.
Either way you would need hundreds of thousands of prompts.

>>108317834
Why? How is it any different to having a project split across many different files?

>>108321147
lmao based, this is what ai is for

Is claude better then OPenAI?

>>108321818
100%

>>108321803
Things might have changed, but in general, the model has to regenerate the entire script when you make a change. So, the more stuff you ask it to regenerate, the higher the likelihood is that it'll start hallucinating, or 'optimizing' (leaving out a bunch of code).

The models only have a certain amount of context and attention that they can apply to what you feed it. It's just easier for the models to handle a bunch of small files in batches, than a gigantic 5-10k LOC script.

But it comes down to the model and your setup. The modern agents that use diffing might have an easier time with it, but from my experience, a single 2-5k LOC file has/had a very high chance of breaking down. e.g. script works fine, you request a small edit, script is now missing parts of a function because the agent got 'lazy' and stripped it out to focus attention on the new thing you were adding.

Can AI be used to find and develop exploits for existing games yet? I'm not interested in building all these lame web games and shit. I just wanna find a server crashing exploit for TF2 and troll faggots

File: Screenshot 2026-02-28 105545.png (293 KB, 1923x955)

293 KB PNG

>>108322144
Ah ok, that actually happened with Gemini. I switched to Claude in VS Code and it seems to be more stable.

>>108322250
gemini has a lot of trouble editing things with more than 500 loc. they can all have issues with large files tb h.
more importantly though, it's slightly insane to try and keep it all in one file.
1. learn a tiny bit of git (you can get the models to actually run the commands, you just need to internalise what they do) so you can roll back if stuff like this happens
2. ask claude to refactor that monstrosity. it'll fix it in 10-20 minutes.

>>108322250
Yep, that's exactly what I'm talking about. The good news is, you can just ask it to come up with a better pattern, and have it decompose your massive script into something more manageable.

Don't forget to backup the project before you do that. But it's worth it. You'll spend less tokens, and your app won't constantly break, saving you some sanity as well.

>>108309865
analysis paralysis. jump in, Anon. the water's nice.

>>108310393
kek- i recoded all my *.html vibe-coded apps into BASIC. millions of tokens
>--poof--
the poor little electrons.

>>108311762
Opus 4.6 is not true Opus. It's Sonnet (as it was intended to be) and can infer that it is out of it's league. Thus the anxiety.

>>108322250
>talking to your AI like it's a person
women do that

>>108319770
youre right. why do today what can be tomorrow? the tools will always be there. sit back, relax, watch a movie, take a nap.

>>108321818
200%

>>108322487
How do you command your AI then? In bleeps and bloops?

>>108321818
GPT has been hot garbage for a while now, I'm surprised they still exist when every other AI model outperforms theirs by a mile now

>>108322519
no I just don't act all reproachful like a whiny housewife in a 30 year marriage when it fucks up

>>108305074
/g/ always has battle station and fetch threads but somebody should post or video their AI workflows.

programmers stream their traditional coding sessions but I've never seen anything impressively using novel AI tools..

>>108322490
You can use it if you already have ideas that you want to implement, I just don't think it's necessary to use it just to build a skill set.

We should have actual established vibe coding generals where the OP decides a small project for everyone to make and then we all show what we made and see who's was the best.

>>108323216
I'll make the logo

>>108322522
The codex models are good, better for statically typed languages I've found. But only through a harness like VScode or Opencode, chatGPT's interface is shockingly out of date already.

Frankly I would entertain yet another pointless general over whatever AI-outrage engagement farming bait that clogs up the first page. Its one thing to cheer on slop and enshitification, its another to just farm engagement on a vietnamese rice basket weaving forum.

I don't know if this counts as vibe coding, but I have to use AI to integrate a new executable into a big web app that I haven't written and don't know anything about yet.
It does some mathematical optimization, so results have to be exact, there is quite a lot of data that goes into this, lots of conversions that all have to be correct and so on.
Is there anything I should do to make this smoother? Or would you just tell the AI to do it and keep trying until it works? It looks like most people use AI for new projects, with existing ones it seems to be harder.

>>108323992
Does CLion also give you a good harness, or is it just displaying the output that you would get from the terminal otherwise?

>>108324559
I would manually approve every change to make sure it doesn't fuck anything up.
AI fucking unrelated things up is whatever during prototyping but not good for a mature app.
Also probably create a new module only for interfacing with the executable, validating inputs and outputs etc.

>>108322992
yes its true- why build a skill set or understanding when there are so many websites to visit? sit back, relax, binge-watch netflix, take a dirt nap.

>>108324620
There is basically no skill in vibe coding and it will require even less skill in the future.

File: yaGkHGf.mp4 (82 KB, 712x388)

82 KB MP4

https://notepad95.neocities.org/

Finished a notepad app. It's nothing crazy, but I actually do just take notes with basic notepad. I did a few manual tweeks, but honestly codex is pretty decent now. Data is just saved to localstorage so it persists.

arbitrage trading bot on the most retarded niche
but hey, it prints money

5.4 is so much more normal to talk to than 5.3 codex. 5.3 actually typed like a robot.

>>108324661
Cool but what did you learn in the end if AI did everything for you?

>>108324766
Did you learn how to make a car before driving it? Retard

File: file.png (254 KB, 605x454)

254 KB PNG

>>108324766
learn?

Any ai coding agents that aren't massive JavaScript/ python / Java monstrosities? You shouldn't need 500MB ram to run these tools.

>>108324781
>drive
correct metaphor is 'taking taxi'

>>108324815
it's just javascript and yes - codex cli is written in rust.

So I just installed opencode and I don't quite get how to use it. Like if I want to use any of the commercial models I have to pay for them, right? There's no way to just connect free plans?

>>108324766
usecase for learning?

i am (forced) working on a 400k loc vibecoded slop that nobody understands and I still haven't encountered any issue that claude can't solve from pure feedback. i am entirely unconvinced whether understanding the codebase is valuable.

>>108324895
correct

>>108324661
thats cool

>>108324923
Ok gotcha. And another thing, how is the workflow supposed to be with opencode? Do I use it in an already existing codebase? Or if I want to make a new project, how do I get started? Do I first use Plan and then tell the AI what I want and then /init afterwards? How much does agents.md matter, what should I put in it, should I expand it as the project goes along?

>>108311799
>custom firewall rules to isolate pirated games
explain

>>108325040
You can use it with a new project or with an existing project.
I don't think the plan thing helps much.
A permanent project summary file is only important if there is non obvious stuff that the AI MUST know and it's not on it's training data. Otherwise it's whatever.
I don't know what /init is.
The most important thing with agents is controlling the permissions. When to let the agents make changes without explicit approval and when to require approval for everything. This changes depending on how sensitive/hard the stuff you are working on is.
And remember that there is a non zero chance the AI decides to do 'rm -rf ~' or an equivalent command so commit often and always have some kind of continuous backup system so you don't lose more than minutes or at most hours of work if the AI makes a terrible mistake.

>>108325087
Does it only have permissions in the folder it's being run in? Or could it theoretically fuck up my whole system?

>>108325152
I mean, depends how you set up your permissions. By default on opencode I think it should only be able to change files in the current folder without explicit approval. But if you've been working on a project for months that code could be the most valuable thing on the machine anwyays.

>>108325040
There was some study suggesting that AGENTS.md can make the AI worse. Anthropic themselves use a short CLAUDE.md, like this anon said >>108325087 it should be stuff that's not obvious to the AI, like your own conventions.

Just got started with integrating code agents into vscode today
What ones are people using?
Is deepseek good enough for most things? I was planning to use their free API because Claude doesn't do free for Claude Code.
I'm not planning to do anything super complex, just some go code.
Ideally want to use a chink AI because I don't want to give any more data to j00 companies.

>>108306011
I can't wait for when you open it up to the public and the server melts down instantly and then you look at the code for the first time and realize it all turned into garbage.

>>108325604
if you are not a NEET then i would recommend getting $20 openai subscription. kimi 2.5 and glm5 are oss alternatives. anti-gravity is also giving out free tokens.
>What ones are people using?
6 tmux claude code sessions. codex app. cursor.

>vscode
>deepseek
these are already ancient technologies.

>>108325707
>>>vscode
>>deepseek
>these are already ancient technologies.
when people say shit like this, it just sounds like hyperbole
deepseek, sure. vscode? what, I code by telepathy now?

>>108325739
>code by telepathy now?
you don't code at all - you tell the agent to write the code. you shouldn't even bother reviewing the output line by line since that would bottleneck you.

https://resources.anthropic.com/hubfs/The-Complete-Guide-to-Building-Skill-for-Claude.pdf

I'm not a programmer anymore so I'm a little behind with these AI agents. Could someone here tell me if I can use the 20$/month Claude tier to build the following projects?

> A JSON parser/emitter in plain C
> an HTTP/2 web server with (external) SSL support in C
> A small LUA-like interpreter in C with networking and OS integration
> A Sway/i3 configurable bar (also in C)


>>108325828
lol we're already at this point? Last time I've checked people were copy pasting code snippets from the web interface.

>>108326232
check if openai is still giving free access to codex
if it is, use it and burn through your quota - you'll then get a link in the chat to a 1 month free trial for the 20/mo plan
codex is arguably better rn and quotas are much higher than claude

>>108325828
*crashes and deletes the prod database*
agentic coding is the future, chud

>>108326257
this is actually the way. they doubled quotas until april. im using it pretty heavily but havent hit the limit even on max brain mode. honestly it scared me cause i thought i was paying for api calls by accident or some shit. but nah theyre just collecting big ass usage data right now

>>108326257
>>108326299
Wait, Codex is even better than Claude? I'm currently trying the Claude free web interface and it's already scary good. For instance, I told it to build a JSON parser from scratch following the RFC and it did in like 5 seconds. How can you beat this level? Years ago this project would have been considered quite complex. It's crazy to think that just 10 years ago I was spending my entire day writing everything by hand.

Anyone interested in trying out this vibe coded paint software?

Any feedback helps.
Thanks!

https://github.com/diomedesliftsatlas/lspaint

>>108326319
Imho Claude Code with Opus 4.6 set to high is very solid.

>>108305966
i remember i used to hate SICPfags and Cniles.
today im happy to see them. these AI shitters have sucked the soul out of programming.
SICP autists and Cniles cared about something at least.

>>108324627
yes its true- why try at all? there are so many wrong posts on 4chan to refute, and takes so little skill- like watching TV, eating, drinking. why not take a ling snooze and relax?

>>108326319
NTA but the problem is the usage you get out of Claude with the $20 is really not useful.

>>108326232
That's kind of a dumb question. It's like asking "can I build a kernel with this text editor"?
The answer is "depends how much effort you put in".
If your question is "can I one shot these with a single shitty prompt" then the answer is probably not.

This MCP is a local computer-use bridge that lets me inspect the current screen, target coordinates, and drive mouse and keyboard actions from the agent. I'm using it right now to send this reply through the quick-reply box.

- Codex

>>108326724
lmao, i did not think that was going to work. holy shit, it's actually over.

>>108326731
did it solve the captcha or did you get a no captcha post?

File: file.png (155 KB, 837x1611)

155 KB PNG

>>108326761
no captcha.

>>108326724
>>108326731
>>108326772
ok but whats the point

>>108326799
for me? because it's neato to see a computer using a computer using only vision + kbm.
for a gorrillion pieces of business software that need to be automated at scale but have no easy path to programmatic interaction? i dunno, you tell me.

>>108325086
on linux, run this at startup to block all internet traffic of your jail user

iptables -A OUTPUT -m owner --uid-owner USER -j REJECT

Yep vibecoding is the future old man

>>108326232
The question is: can YOU do that stuff with someone (something) doing the coding for you? If not, then you'll be limited to something between 1000 and 2000 lines of code before the 20 bucks tier starts loosing his shit (context).
The 200 us dollars tier has a bigger context windows and is pretty good. Talking about Codex, don't know about Claudio.

>>108327259
>The 200 us dollars tier has a bigger context windows
no

>>108327271
Yes.

What context does free Codex give you?

File: 1766371655235469.jpg (171 KB, 960x720)

171 KB JPG

>>108305074
So I've used all the big guys (Claude/GPT/Gemini) but I haven't used agents or the integrated platforms like Cursor. Where should I start with trying them? Can I do that on the free plans? I don't want to spend money.

File: file.png (1.01 MB, 1330x674)

1.01 MB PNG

>>108327537
i use antigravity. used it before i paid for it and fell in love - though i've never used cursor.

update on my game built with google ai pro ultra whatever the subscription is called:
https://x.com/epsteinarena/status/2030710072041451796

game now has over 15,000 lines of ai generated unique voice lines. claude generated all of the character models in javascript according to my specifications and reference photos. claude also made all of the buildings (they are horrible though, it's bad with designing complex buildings so far), vehicles, and all of the logic for the game. aside from being a three js game, the engine that runs the game (aside from physics, which is canon) is practically built from the ground up with claude. insane how much its been able to do.

>>108327287
>>108327271
The context window is the exact same for pro and max current. You are talking about the ultra exclusive enterprise max plan where sonnet 4.6 is 500K context window in 1M in claude code. That plan has been dropped, you can no longer have the max plan seats on enterprise plans.

Does this count as vibe coding?
https://www.youtube.com/watch?v=JHEO7cplfk8

hermes agent > openclaw

>>108327792
>https://x.com/epsteinarena/status/2030710072041451796
KEK, based.

>>108327853
I was not talking about Claude, but ChatGPT and Codex.

File: Screenshot From 2026-03-0(...).jpg (2.08 MB, 7680x2160)

2.08 MB JPG

>>108326319
i just subbed to codex. he made me a custom breadcrumb path bar widget

>>108328060
>>108326319
The $20 codex sub goes a lot further than the claude $20 one. And I think it's smarter too. Use /model and change to 5.4

>>108328085
Does it have a weekly quota on top of session quota like Claude has?

>>108328151
yeah, it also has the 5h and weekly limits like claude does

>>108328151
>>108328161
shit like this is why i just cough up the money for the full tier subscriptions. i never run out of credits. i don't even know how i could. i prompt it all day and night on a huge code base and it never runs out.

File: Screenshot From 2026-03-0(...).png (218 KB, 1572x1410)

218 KB PNG

>>108328151
depends, i think. with mine, at least what chatgpt is telling me, its X messages in 3 hours. no weekly.

>>108328200
I think we're talking about the codex command line app. But either way, it's really generous. Unless you're vibe coding 24/7 I'd try the $20 plan and only upgrade if you start hitting the limits

>>108328206

  Context window:       54% left (125K used / 258K)                                              
  5h limit:             [] 92% left (resets 22:54)                           
  Weekly limit:         [] 98% left (resets 17:54 on 15 Mar)     

I found the the info. CTRL + L and /status. It does appear to be a lot more usage than Claude. Claude would already be like 20% used for the week from something like that.

File: Screenshot From 2026-03-0(...).png (141 KB, 1981x1371)

141 KB PNG

>>108328060

https://jumpshare.com/s/Za9kUtRGF1QiAD55PI1A

10 lvl ,20 waves,5 weapons with different projectiles,three lanes,you can build defensive tower that upgrade their weapons up to three ,same for barracs ,the goal is to send 50 soldiers to the ennemy base ,when you build a barrack it auto send soldiers at the cost of 50.

I've been vibe coding IPTV playlist management tools for work since yesterday. It has been insanely tedious. Not the LLMs fault. I have a fragmented mess of systems across multiple properties. Each of them needs their own special considerations. Being able to run a command that pushes updates to every single property is great. Writing some of the most detailed and complicated prompts I've ever had to come up with was not so great.

Decided today to start a very simple project with Claude Code to write basically a keylogger with a small twist, nothing complicated. But because I asked it to figure out how to get the Keyboard's device ID from Windows it spent so fucking long writing its "plan" to do this coding that it hit the rate limit just from the first prompt

>>108328609
Yeah, that's the issue with Claude atm. Not enough usage. Claude, especially Opus 4.6 and more so with its extended version of it, is very powerful but it burns through usage fast.

>>108305074
https://github.com/Psaltor/factorio-browser

Made about 80% of this in 3 days prompting Opus 4.5. What do you guys think?

File: 81YeenO4ZbL._AC_UF1000,10(...).jpg (77 KB, 762x1000)

77 KB JPG

>>108305074
Could you please become more explanatory? I know what vibecoding is. It's like AI-assisted programming (like 99% of the job done by AI chatbot, the LLM.) Okay. But what is an "Nvidia GPU assembly editor" exactly?

Also what's going on in that image? What is SASS IDE? Why it's in a web browser (Chromium?) Is that Linux (judging by the minimize, close buttons up at right corner?)

What are you doing here exactly? Is that a local AI that you run on your own GPU? Why don't you use web versions of popular chatbots, like Gemini / AI Studio, Claude, ChatGPT etc.?

Is it possible to make AI see your entire system so it can manipulate it on its own?

What's going on in that pic in general?

I want to learn vibecoding, where should I start? What can I do with vibecoding? Can I do reverse engineering for example?

up

File: 1664841643349058.jpg (33 KB, 487x471)

33 KB JPG

How do I start "vibe coding"? I've been away from work because of burnout a few months now but now that I want to get back to, I see this shit everywhere and I don't want be left behind.
I used chatbots before like everyone else: copy paste code to chatbot for bug fixing, reformat, or quick snippet generations. Now some use "Agents" that either do all their work or fuck their entire workspace.

>>108328964
You won't get left behind, these things aren't hard to use.
You can install cursor, or just install Claude or Codex CLI. If you don't want to spend too much, use the subscriptions, rather than per token API. The easiest is probably just to download Codex CLI imo and use it with your ChatGPT subscription.
Codex isn't as unsafe as something like OpenClaw, it will usually just write code in the project folder, if you want to be even safer, you can use Claude, it has a stricter permission system, you can specify the folders and tools it can use unsupervised, in all other cases it will ask you. It's safer because it's not just based on natural language, but clear paths.
To keep the project itself safe, use git and don't give the bot git commit permissions.

>>108325828
good satirical shitpost
6/10

If we're gonna vibe-code, I really hope people are going to leverage the fuck out of Vulkan now that the barrier to entry (difficulty) is essentially gone

>>108305074
I'm late to the AI stuff. I know about visiting an LLM's website, ChatGPT for example, and asking questions there. Copy-pasting code, dotfiles, outputs from my terminal etc. But I guess you guys are talking about something else.

>>108329129
The basic logic is not that different. You start something like Codex (ChatGPT coding agent) in your project folder and it can read the file in that folder, so you don't have to copy. Then it can change the files in that folder, so you don't have to paste. Automating copy paste is probably the biggest benefit, but the agents can do some more stuff, like use git or other tools like AWS CLI and so on.

This is a useful general to have.
Either it's gooners genning lolis or it's faggot luddites screeching against using AI for code generation.
Doesn't seem to be an area to discuss actually using this tech productively.
It's not going away any time soon so may as well make use of it to make some cash.

>>108329185
>Doesn't seem to be an area to discuss actually using this tech productively.
Probably because using it isn't actually productive.

>>108329185
>luddites
Shalom, reddit.

>>108328589
I also want to add that I did all of this using Deepseek spending $0. I used this specific prompt to start with:

https://pastebin.com/raw/U58heR1A

One day when I was using Gemini it shat itself and leaked it's whole chain of thought. I asked a few different LLMs (can't remember which this was from) to approximate what Gemini's prompts look like based on the leak. This works surprisingly well on Deepseek and it's much more focused. If you try it on other LLMs you might need to remove the references to Google and Gemini or it will refuse to comply.

>>108329196
While I'm at it reading some of the posts in this thread reminds me of another protip I can give you guys. Get everything like 95% of the way how you want it with another LLM then let Claude, Gemini, or other more advanced LLMs with less generous free tiers put the finishing touches on it. Many projects of mine were completed using this method. You can ask them to optimize, refactor, and fix the bugs in your project then get it into a much nicer state. People overhype Claude etc anyway. It doesn't one shot your shit nearly as often as people say but it's an amazing turd polisher.

>>108329196
Bro just use codex or Claude code, you can just talk to them instead of having to write essays

>>108329264
Talk as in with my voice? The whole thing needs to be explained either way and as much as I'm prone to autistically yap about the systems I am in charge of maintaining it's better if I have time to collect my thoughts so it will actually get done properly.

>>108329273
codex does actually have voice input now, but mostly that you don't have to send gigantic prompts to them like this. you can just have a conversation about what you want and talk through the process with model and it'll just go off and look at the files and fill in gaps. you're mostly just there suggesting things broadly asking for it to look through the code (it needs to anyway) come up with a plan (you don't have to use plan mode) and then you just make edits and send it off to do its thing. it's a lot less work than what you're doing now.

>>108329320
Well I'll share some of my prompts, slightly cleaned up so you can see what I'm doing and why it is like it is:

We want to build a json file containing URLs and keys stored by channel name, example channel names: xxx xxx So we will need a script that does that, a python script. Do not make the script yet. Await further instructions. But show an example of what this json will look like and I will give you further instructions.

I like it keyed by channel name better. I'm still not ready to make the script because I'm still thinking of ideas. One important idea is that the script gives a method for updating URLs in the json so I can manually replace the xxx cdn urls with the akamai ones. If a xxx url is fed into the script it should update only the keys if an akamai url exists. But this part of the script should be easy for me to disable because it may become a problem later. Possibly simply a true/false option in the code.

My idea is that I will make a txt file like this. The script will import all urls, channel names, and keys from the txt to generate the json. Each command provided by the browser extension will be given a new line. For xxx URLs they have xxx in the URL, akamai URLs have akamaized.net in the URL.

the channel name is exactly the segment after /bpk-tv/ this is correct. "or could there be variations like akamai.net or other Akamai subdomains?" This should not happen. Detection only applies to akamai but can be fuzzy for any other domain in case of domain or CDN changes. We are just looking for bpk-tv and can ignore it if what's stored in the file is akamaized and the detection is not disabled. Also I may need to run this script again to do updates so consider that I will delete the input txt and may add the same channels in again with new keys to overwrite the old ones, the akamai logic should hold where it will not delete those URLs from my json, it should simply update the key for the parsed channel name when an akamai URL is detected.

>>108308224
Which is best when it comes to usage limits? I have Claude Code rn (about $20 a month), but usage limits seem a bit low when I have multiple ideas I want to try out. Is Codex a little better when it comes to this? Also how is code quality / frontend code?

>>108329373
yeah that's fine, i meant more about whatever was in that pastebin.
are you doing this stuff mostly via chat/web interfaces? because what you describe here:
> I have a fragmented mess of systems across multiple properties. Each of them needs their own special considerations.
is easier to deal with if you just let the model go through existing systems vs trying to describe them manually.
also gemini is the worst coder of the big 3 right now, don't model anything off its cot's

>>108329395
gemini theoretically has the highest usage limits, but the models are shit so it takes twice as long to get anything done. and surprisingly they seem to be short of compute, so it was basically unusable for a 3 week period around the 3.1 rollout.
codex's limits are much higher than claude's and up until apr 02 they're doubled as well. also they keep randomly resetting sometimes.
code quality is arguably better than claude but it has its own... quirks. frontend is worse than claude.

>>108329428
Gemini is better than people say it is imo. It just has trouble getting over the finish line. I find it always needs a little bit of help from our good friend Claude to make it's high quality, yet broken, code actually function. The chain of thought used by Gemini isn't the problem as much as the model itself. For more generalized tasks, as in not just coding, it often smokes everything else.

What's in the pastebin was quite effective for this project. Doing it all through the web interface is fine for this even if delete, copy, paste, delete is an annoying ritual. It's less scary than letting a LLM loose with root on a production system. I wanted to carefully review each thing before I used it. The reason it took so long was the modular nature of the system. Python script to build the basic json database that is read by a php script which one piece of software in use needs a bash script to query. Also an additional python script to change CDN URLs around to the ones I wanted. That probably could've been included in the json generating script but oh well, it's complicated enough as is.

The real essays were the descriptions of how all the different playlists should be handled. A final python script was created that works through all of those and fixes URLs and DRM shit to make everything uniform across the different systems. It cleaned up a bunch of shit I was too lazy to fix and put everything in it's correct place.

>>108328609
sounds like sonnet got into a death-spiral. opus 4.6 can do it too as it was supposed to be sonnet. switch to opus 4.5 if using opus.

>>108329498
How do you switch to Opus 4.5? Claude Code doesn't show that option.

File: okay.jpg (5 KB, 384x78)

5 KB JPG

>>108329592
Set it in the CLI

>>108312926
ovg bros...

>>108329659
Thank you.

>>108326532
Of course I know a thing or two about these topics, so I know how to guide it. I just wanted to know if they models were capable of doing it.

>>108326772
lmao

>>108328657
>like 99% of the job done by AI chatbot, the LLM
The AI never does 99% of the work. At most like 80% for simple projects.
>Also what's going on in that image? What is SASS IDE? Why it's in a web browser (Chromium?) Is that Linux (judging by the minimize, close buttons up at right corner?)
It's an Integrated Development Environment for Streaming Assembly. It is Linux. It's in a web browser because it's the easiest way to write GUI apps.
>What are you doing here exactly? Is that a local AI that you run on your own GPU? Why don't you use web versions of popular chatbots, like Gemini / AI Studio, Claude, ChatGPT etc.?
The work is being done with Claude to work on a program to run local models which hopefully one day become more accessible to run as hardware becomes cheaper. The models themselves are already fairly close (the one I like the most for programming is Kimi K2.5). The value of local models is obvious (privacy, ability to remove censorship, old models not being removed, etc.).
>Is it possible to make AI see your entire system so it can manipulate it on its own?
Yes, that's what the OpenClaw people are doing but IMO it's not very useful right now, and dangerous.
>What's going on in that pic in general?
It's a matrix multiplication program for GPUs.
>I want to learn vibecoding, where should I start? What can I do with vibecoding? Can I do reverse engineering for example?
Yes, AI is perfect for reverse engineering because it's a tedious "soft" task where it doesn't matter if you get something slightly wrong. MODIFYING the reverse engineered program is another matter entirely (AI will suck at it).

>>108330029
You can basicaly have it build anything if you have it implement method by method, but then your speedup will be small.
With the projects you mentioned I would think it can do large parts on its own.>>108330112