A general for vibe coding, coding agents, AI IDEs, browser builders, MCP, and shipping prototypes with LLMs.►What is vibe coding?https://x.com/karpathy/status/1886192184808149383https://simonwillison.net/2025/Mar/19/vibe-coding/https://simonwillison.net/2025/Mar/11/using-llms-for-code/►Prompting / context / skillshttps://docs.cline.bot/customization/cline-ruleshttps://docs.replit.com/tutorials/agent-skillshttps://docs.github.com/en/copilot/tutorials/spark/prompt-tips►Editors / terminal agents / coding agentshttps://opencode.ai/https://cursor.com/docshttps://docs.windsurf.com/getstarted/overviewhttps://code.claude.com/docs/en/overviewhttps://aider.chat/docs/https://docs.cline.bot/homehttps://docs.roocode.com/https://geminicli.com/docs/https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent►Browser builders / hosted vibe toolshttps://bolt.new/https://support.bolt.new/https://docs.lovable.dev/introduction/welcomehttps://replit.com/https://firebase.google.com/docs/studiohttps://docs.github.com/en/copilot/tutorials/sparkhttps://v0.app/docs/faqs►Open / local / self-hostedhttps://github.com/OpenHands/OpenHandshttps://github.com/QwenLM/qwen-codehttps://github.com/QwenLM/Qwen3-Coderhttps://huggingface.co/bartowski/Qwen_Qwen3.6-35B-A3B-GGUF►MCP / infra / deploymenthttps://modelcontextprotocol.io/docs/getting-started/introhttps://modelcontextprotocol.io/exampleshttps://vercel.com/docshttps://mcp.desktopcommander.app/►Benchmarks / rankingshttps://aider.chat/docs/leaderboards/https://www.swebench.com/https://swe-bench-live.github.io/https://livecodebench.github.io/https://livecodebench.github.io/gso.htmlhttps://www.tbench.ai/leaderboard/terminal-bench/2.0►UI/FrontendFigma MakeLovableClaude designhttps://uiverse.io/https://ui-ux-pro-max-skill.nextlevelbuilder.io/https://stitch.withgoogle.com/https://gamma.app/https://github.com/nextlevelbuilder/ui-ux-pro-max-skill►Previous thread>>108691806
>>108706320The best corporate use of AI generated code is never really talked about. But it is savage.
so sonnet 4.6 is just as smarter as gpt-5.5?
>Now I have the whole picture>(doesn't understand at all)
https://www.technologyreview.com/2026/04/27/1136466/elon-musk-and-sam-altman-are-going-to-court-over-openais-future/This week Sam winsNext week? Sam also winsIPO? Sam wins. Buy the $100 plan NOW
chatgpt in xcode is unusable lmao>please fix this>got it, here's 4 solutions! (massive wall of text) (that will be 10,000 tokens)>would you like me to implement them?>sure>okay! >it doesn't work>hm... I understand. (reverts all the changes)
>>108706420Just use the Codex app, Apple's integration is useless right now
>>108706352No. GPT5.5 (or 5.4) and Opus 4.6 can solve things Sonnet 4.6 can't.
>>108706372kek
If you think you're a hotshot programmerand you think AI is shit and worthlesssee if you could code this if left to your own devicesI know for a fact that I couldn't
>>108706709sir you have red squigglies in your codeplease press F7 to run the spellchecker
>>108706716thats because I put it in as blank file
>>108706320which ai shits has autocomplete?id rather work on code and get autocomplete suggestions, rather than AI burning up all my tokens
>>108706771If you had asked that question a week ago I'd have said GitHub CoPilot but they've just introduced usage based billing, so fuck you.
Thread challenge: get an llm with a 1930s cutoff to do fizzbuzzhttps://talkie-lm.com/chat
>>108706709btw even a mathematics phd would struggle to write this codethey would use numpy to do use huge matrix multiplicationsbut this has bitwise manipulation and manual indexing which is much more programatically faster and memory efficient than the libraries out there. they would understand the theory of it, but struggle with the implementation in a programming language. Only a very competent engineer could write this code
>>108706771Cursor did that a long while ago (18 months?) but the quality of thinking that can be done in tab-complete time isn’t very good
Tibo just reset rate limits lmao
>>108706806yeah, what I really like about having robots write code for me is that they can apply very-rare autism to an entirely new domain and be _very_ good at it
>>108706815…ChatGPT just reset rate limits?
>>108706840Looks like my weekly limits got reset 12 hours early
>>108706840Yes. And he's managed to do it on the day they were gonna reset for most people anyway. If we're lucky he resets them again tomorrow like last time.
>>108706849damn. if i knew they were about to reset I would have bumrushed the 30% I had left
>>108706855same here, I was saving my last third or so because my limit was gonna reset in a couple daysnow it’s 100% and I’ve got a week to use it in
>>108706855>>108706859And that's why they're never gonna announce it in advance
>>108706859>>108706863NOW I FEEL JEWED CHEATIED OUT OF MY LAST 30%
>>108706865I kneel
https://chilitown.orgI vibe coded this bros, any feedback is greatly appreciated.Will fix up the dark mode stuff when I get home from work.
>>108706902>accountsmegacringe bro
>>108706816What's so great about that codeto write it if left to your own devicesyou'd need to be an expert in 3 different academic domainsor at least 2, and have a basic understanding of a 3rd enough to do research in itwhen people say AI is shit, and I write code like this with AI people really have no idea what's coming
>>108706865Claud Code bros.... our response???
>>108706909You can vote and see memes without an account.The plan was originally to make it pure anonymous but plans shifted.
>>108706902I tend to think of “spicy” as more or less “annoying to the regime” and “not spicy” as “having no political content whatsoever”commie propaganda is neither>>108706942I have zero reason to believe that a vibe-coded app securely stores login credentials and IP addresses so they cannot be recovered by hackerssee also: the app where women say bad things about the men they’ve fucked, and how it got hacked
>>108706955The most spicy content you can make is commie content about murdering billionaires.Being a uppity natcuck who spends all day secretly jerking it to blacked porn is neither here nor there.
>>108706977>billionaireshelicopter cargo only complains about “billionaires” because otherwise millionaires like Hasan Piker would find themselves having to fend off accusations that they’re part of the class that deserves the rope
Is Codex broken? Mine can't compress context anymore.it's fucking over bros
>>108707018beats me, I ate up my 5h limit in one go
>>108706806>>108706816You two are so retarded it's painful. Linear algebra is by far the most optimized set of software liberties in existence and basically all modern compute devices are designed with it in mind. Which is particularly important because keeping your CPU fed is difficult due to memory speeds. Not to mention your "clever" AI code is using fucking Python lists instead of C++/FORTRAN numpy arrays if you did a linear math approach. I like AI, but God damn, to no-coders and morons AI is some sort of magic God.
>>108707053hah, I vibe-coded a thing and got a program that used numpyI had to throw out the Python part and switch to a different language because numpy couldn’t do streaming data and I didn’t want to have to use 110% of my RAM to process thingsbut numpy is good software
>>108707053You are a moron. Even basic variations like integer quantization with scales requires you to write a kernel from scratch and cuBLAS or whatever boomer software you have in mind will be completely useless there.
Did my codex limit just get reset early or something? It was meant to reset on May 1, now it's back up to 100% and resets in a week.
>>108706994Pinochet was controlled by the CIA which is an institution dedicated to taking people's freedom away you idiot.The people who supported throwing people from helicopters won, and you see how fucked up everything is.You can become a millionaire through being exceedingly brilliant and in the right place at the right time, though it's rare. But becoming a billionaire absolutely requires being a scumbag.
>>108707125yup, that’s what we were talking about upthread
>>108707141
>>108706955Paid a good amount of attention to security. Argon2id hashing etc etc.Worth another look to avoid an AshleyMadison situation.
What did AI mean by this?
>>108707183“I don’t know how to use triple-backtick code fences”
>>108707179that’s good but there’s a lot of “oh shit, I wouldn’t have thought about that” in security
>>108707176She probably wouldn't let me touch her tits anyway so idgaf
>>108707183Heh. Reminds me of the ramblings Opus 3.5 would come up with if you let it work for too long unattended
>>108707176>>108707245>implying those are the only things that are big
Lmfao I'm testing a bad word filter for the first time and Codex 5.4 just KWAB'd meTotally unprompted or without me ever mentioning those words in any conversation or instruction ever
is it feasible to use a Claude code or a codex subscription without using their harness?
>>108707403Sure
>>108707446Will it result in me getting banned
>>108707459Nah.Anthropic doesn't like it but hasn't actually banned people for it that I know of, OpenAI explicitly allows it.Don't enable extra usage though, they are trying to charge you API prices if they detect you're doing it. The idiots were trying to detect it by searching for a HERMES.md in the project folder and they were billing people API prices for using Claude Code just because that file was present lmao.
How do big harnesses solve the problem of AI stuck in a loop trying to fix it's hallucinations?Writing my own local coding loop from scratch, picrel: the AI over and over again tries to fix wrong parameters for a function call, while it was wrong library to call. Will comparing the last 4 errors and saying "bro, it's the same error, you're must be doing something completely wrong" help?
>>108706320I wish Grox is actually good, I wanna sexting and vibe code at the same time :(
>>108707576You shouldn't mix your context like that, it lowers coding performance.
>>108707565I believe Claude Code has some basic thinking loop detection, Codex doesn't fall into loops to begin with.But when using big models it should be very rare, like once every 10 hours.It sounds like you have some major bug or using a model that's too small or quantized.
>>108707634> using a model that's too small or quantizedYeah, seems like. Restarted the loop and it solved the task in 4 iterations, never encountering that error.
>vibecoding in anything but rustShiggy diggy
>>108706709I wouldn't but that's because I've always been a mathlet
>>108707497Ok thanks for the heads-up
GPT 5.4 High vs 5.5 Medium, has anyone been comparing them?GPT 5.5 High is probably too much of a usage-guzzler to use even with my two accounts.
>>108706320Anyone else hitting usage limits rather quickly on Codex 5.5 medium? Im just using their Codex IDE and debugging a web app.
>>108708077Yeah it's possible that 5.5 just isn't even worse using even in comparison to Extra High on 5.4
>>108708100worth* using
>>108708100Yeah... Im gonna see what using 5.4 on high is like and try it that way, because I don't feel like waiting 5 hrs until usage resets.
>>108708130I've heard that some people even use 5.3 just for the way better usage. But I really wish we had a simple chart for intelligence and usage ratios.The charts I've seen list token usages and costs as separate things and I just have no idea what I'm looking at. I don't even know if API costs/usages line up with subscription costs/usages. I know you get better value with subscriptions but that's all I know.
>>108708208I would use 5.3 on Cursor for free through pro but it would get stuck on debugging things. 5.5 so far has mostly been able to fix deep bugs in one or two tries on high mode, but it takes like 20 mins sometimes
>>108708208A gpt plus subscription probably pays for itself within the first one or two 5 hour blocks if you max them out. It's insane how much better value it is.
>>108708255>pays for itself
>>108708260You know what I mean. You'd spend $20 on API usage doing what you get done in 2 blocks with chatgpt plus. Faggot.
Using web ChatGPT to form coding plans specifically set up for use with codex definitely has reduced my usage and codex spends less time trying to reason (if it even tries). I could probably set up desktop commander so I don't have to keep telling it to pull my git repo and make it faster
>>108706709This isn't programming as much as it is domain modeling. I couldn’t write this code but if the AI gave it to me I couldn’t prove its correct either so what’s the point? Most programmers wouldnt be able to write that code and those who would be able to wouldnt be able to write other specialized programs from different domains How do you even know the code works and does what it says?
So it seems github copilot will be useless starting june, you get credits based on your subscription and get charged API prices when using them, I see no reason to subscribe then.
google's so fucking infuriating.nanobanana limits in flow are very generousyou can generate images in antigravity with your subyou can only use api keys in the with the gemini-cli extension ???there's no official mcp or cli that lets you use your sub quota for image gen ???>>108708321i'm tempted to do what you're doing and just build a cli that's just wrapped around an authenticated headless chrome that's running flowthis is bullshit
>>108708472Just the free tier stay? That's the only thing that matters.
>>108706902Fixed up light mode default (also respects user's default theme)Fixed OG images and titlesAudited account creds securityFew icon shifts and other minor changesAny other feedback appreciated anons.Upload a meme or 2 if you feel like it.
Some of the things AI does are beyond human comprehension.
>>108706919a $400 tier plan
>>108706709AI is good at writing individual pure functions. It's also generally good at unblocking you. You might not understand 1 math equation, or need to reverse engineer 1 function, that's all good.AI still writes bloated code and bad architecture. Even though I use it, there are many things I could do better. For highest quality, a mix of AI and human code is still the best, but often I get lazy.
Some anon had an MCP server with the chatgpt web UI to have what codex does cheaper, how do you set that up?
It's tough going back to something like gpt 5.3 after using 5.5 for a while. I'm worried it won't do things correctly or completely break things that 5.5 setup.
Which of Cursor llms have best price/performance? Composer 2 is pretty good, liked it performance so far, but is it the best?
>>108706865I MISSED OUT ON THAT, I DIDN'T KNOW
>"I want to do X">"I don't think doing X is a good idea, let me waste tokens by telling you why...">edit prompt, add "Don't add commentary on whether this is a good idea or not.">"Here's a step-by-step plan for how you can do X by yourself...">edit prompt, add "Do it for me.">"Before I do this for you, let me waste tokens by explaining to you what else you could do other than X...">"Do X anyways.">"Here are several ways to do Y, which can do something similar to X..."
https://xcancel.com/llmdevguy/status/2048736363600613776
>>108708993I have never experienced this. Perhaps the issue is your skill.
>>108708993Why not reply instead of rerunning the initial prompt, I don't understand.
>>108708993Stackoverflow simulator
>>108706352absolutely not
>>108709001retard
Going to sleep, if anyone has any other thoughts please share:https://chilitown.org
10 hours passed in the blink of an eye.
>>108709021Replying leaves useless information in context.
>>108709186It is not useless in that context, you are telling the model you are aware of X and Y concerns and that they don't matter, go ahead anyway.
