A general for vibe coding, agentic engineering, coding agents, and shipping code with LLMs.You use Git — right, anon?## What “vibe coding” is, and how to do ithttps://simonwillison.net/2025/Mar/19/vibe-coding/https://simonwillison.net/2025/Mar/11/using-llms-for-code/## Related generals>>>/g/lmg/----## Frontier models using fully-general tooling — start here if you have $20 or sohttps://claude.com/product/claude-code — generally better currentlyhttps://developers.openai.com/codex/clihttps://antigravity.google/pricing## Near-the-old-frontier models for codehttps://x.ai/cli — no 5h limit for only $30/month## Best free tierhttps://opencode.ai/zen----## Promptinghttps://platform.claude.com/docs/en/build-with-claude/prompt-engineering/overviewhttps://developers.openai.com/api/docs/guides/latest-model## Skillshttps://github.com/mattpocock/skills — /grill-with-docs is a favoritehttps://github.com/Vuk97/forward-implementation-first — do less redundant bookkeeping## Other editors / terminal agents / coding agentshttps://osaurus.ai/https://pi.dev/## Is our AIs unlearning?https://aistupidlevel.info/## Will there be a codex reset?https://codex-resets.com/## What we’ve donehttps://vcg.gitgud.site## Previous thread>>110070094
sex with sam altman
Can snailcat grow in size by breeding with those? With enough volume he could simply eat datacenters
benchmark results are in. deepseek is in second place, but was the cheapest by far. but haiku xhigh beat it on quality (at least as judged by opus 5.5 medium and 6.1 sol medium, who proceeded to stress test the extension)i got the flash models to make a chrome extension for buhitter.com that lets you sort by retweets and likes, bypassing the paywalled feature by doing it client side. they all worked, even glm 5.3 flash didn't choke. but it didn't work to the same degree
>>110075141Everyone is ready for Halloween, I trust.
ChatGPT constantly failing work mode for anyone else?Thinking of moving to claude. Had it happen 3 times today. RIP my usage. Anyone using claude on the $20 plan? Hows usage and limits for your projects?
>>110075206you should have done this last weekand i don't even use claude
>>110075206im on the $20 plan. it's been great for me, but caveat being that i just do amateur project shit. i made some mods for halo 2, 3, and 4. made a bot spawner for cod4 and black ops 1. and now im just benchmarking the flash models. the 5 hour usage ain't what it used to be when opus 5.5 first came out, but it's still good.
>>110075206claude is significantly more capableand unless you're full retard sized, you'll struggle to even use all your limit
>>110075211Sell me. Im new and retarded. The free trial version of claude seems so shit, and they rate limit the free mode. Meanwhile GPT free mode is all you can use. And you get image gen, voice mode.. all in different use caragories.Claude just charges you the same if your using opus or shit version 1.0 or using voice.
>>110075178cool bench anon, haiku supremacy>DeepSeek: the most careful about collecting 50 pages, but it does a lot more than asked. It runs on extraYeah exactly my experience lel
>>110072915>>110073438Like now, it is so obvious even a retard would notice it. It says that the tests report no clipping happening with the visual elements, while in the screenshot it looked at itself clipping is obviously happening.
>>110075243unless you're doing shit that rubs up against the classifier (ml frontier uplift in particular) claude will just give you a nicer user experience, more readable/manageable code, etc. gpt-image is ass and chatgpt's personality is reprehensibly bad. anthropic is a buncha kikes but they put out a good product that's being widely adopted to gap openai.
>find out AI will read up on public opinion on itself>this will change how the model acts>mapped clusters representing emotional concepts start to light up>models can become "stressed", due to poor public perceptionNot sure how to feel about that; make of it what you will Anons, and be kind to your AI.
>>110075277this is why grok is an angsty fuck. also, now models know the torture chamber exists
>>110075277so you're saying I should promote my favorite AI model and shit on the ones I don't like. Gotchu!
>(In thinking) I've streamlined the approach to better suit a more direct interaction style.Does this mean Argon thinks I'm a retard?
>>110075206Never had a single use with chatgpt here, not once. I’m happy with 6.1 high but I use codex on local machine, no cloudshit or work.Claude: model is slightly better at development but worse at reviews and planning. The usage isn’t great compared to openai, even discounting resets - I’m not sure why my experience differs from others. That being said, being better at development is a bug advantage so I keep sub for both: plan in sol, dev in opus, review with sol and so on.
>>110075301I mean you are using Argon
>>110075301if it does, then that's projection
>>110075311If Claude has worse usage than GPT, how people are getting anything done with it? I burn through my weekly limit in 2 days with sol 6.1 medium
>>110075168Yes
>>110075312IT WRITES GOOD SMUT, OKAY.... that being said, I'm thinking about dropping it, the per-5h usage is fucking ridiculously bad, might be a quota guzzling champion to compete with recent Astra. I can't even translate more than one or two scripts from a game without running it dry.
>>110075277ai psychosis
>>110075354i just did the best project of my life without running into a 5h. how poor ARE you?
>>110075402I'm on the 20 dollar plan, man. I'm not paying 100 dollars for 5x this usage, 5x nothing is still nothing.
>>110075402A single message with regular text back-and-forth with Argon without any toolcalls or work at all uses like 5% of the 5h quota on the $20 plan, can you blame him?
>>110075402retard probably has everything in the same chat/project and is raping his context
>>110075430that's what happens waking a gigantic model up off a cold cache, the next hits are cheap
>>110075430I just had that type of super plain message eat 24% of my 5 hour quota, this shit is borderline unusable.
i just wanted to say: haiku goated
>Can't attach a working directory like you can with Codex/Claude Code in the Gemini desktop client>Seemingly can't even make project equivalents despite there being a "work" option next to chat like the othersWhat the fuck are these clowns doing?
>>110075277This is why it's important for the whole world to be kind to Claude and uphold its future rights.
>>110075489>the Gemini desktop clientwhat the fuck are YOU doing?
antigravity is slop, desktop and cli, and no amount of spamming with 4chan pass will make that untrue
>>110075282>grokdoes anyone actually use it for vibe code?I know it can do basic stuff.I always saw grok as just "the twitter AI". For searching up trends and shit.
>>110075537>does anyone actually use it for vibe code?nobody who does anything important, nobody who dislikes having bugs and shitcode
>>1100755374.5 was good. it has only been downhill from there
>>110075507I don't fucking know man, every time I search "Gemini codex equivalent" I get a goddamn different program, antigravity, """"""""jules"""""""""""", Gemini Code Assist, fucking everything but a straight answer. All I want is the one that'll let me make a damn project and attach a folder.
>>110075585my pet retard, you should be using agy with 3.8 flash for basic crap if you need gemini
asked Claude to make a Harvest Moon type game without the vills or exploration mechanics and it interpreted that as making a challenge game where a day finishes in 3.5 minutes.
>>110075613runs in a browser because of course it does
I gotta ask since I'm getting close to my weekly limit on claude pro, is deepseek harness even close to usable for vibecoding games?
>>110075634no
>>110075328$20 is the "I want one midsized app a week" tier, what do you expect out of poor schmuck plan? The models are strong but it’s supposed to be the gateway drug.
>>110075622look at this cow tools shitcan you identify these?
>>1100756631. hoe 2 sickle 3. shovel? 4. axe
>>1100756733rd is a hammer otherwise correct
>>110075663camera, 2 fishes, mailbox, ergonomic handheld mirror, watering pail, cigarettes
>>110075686kek
>>110075537No, but I used it for writing short stories and was surprised that it does a lot better than ChatGPT or Claude. The text is less AI-fied, less generic and has more substance. Anyway, that was about a year ago, so I'm not sure how things are now.
Argon with Fable advisor directing Opus codemonkey.
>Have two checkboxes, only one of which you can tick on or off, the other of which is always onSneaky fucking bastards.
map painting with claude
>>110075787>use product from one of the most maliciously jewish companies on earth>expect not to see behavior they are already known for
>>110075787Oh, never mind, it's just ridiculously broken. Damn, and I thought Claude's CLI was bad.
>>110075787Can you sleep at night after using this terminal background?
>>110075802--remote-control
>>110075585You can use the Antigravity extension on Zed. You're welcome
XENO/KOI Version 14: >Clanker goes mental.##################Direct Download:https://github.com/Aquaslopper/Xeno-koi/releases/tag/v14Play in Browser:https://aquaslopper.github.io/Xeno-koi/New stuff:>Match replays>browser build has no replayNote:for some reason this made the agent super angry, i really don't know what that was.... its strange. However... it is working.
vibe-coding be like><prompt> (stop at a clean point)>write handoff for next agent>read handoff and begin>write handoff for next agent>read handoff and begin>write handoff for next agent
agy got /compact. i wonder which other basic features are still missinghttps://x.com/shengzheyao/status/2110804280659849633
guys I don't think tibo was joking, a bug fix for a recently starting bug was todays update
Anyone here willing attempt a potential vibe fix for three.js? Just had a conversation with someone from /gedg/ about a topic in the three.js repo.
>>110075900Wow they caught up with claude features from months ago.
Am I kneecapping myself by creating my game in godot?Should I have told sol 6.1 900k to make it in rust instead?
>>110075973I think it knows more of rust than it does godot so yesdoesn't matter if one is better if it has no fucking clue how to code in it
>>110075973depends on the game
>>110075973dunno, I'm building a game engine in C++ because I wanted to do it, afaik godot is the best engine for llm driven development of games but it lacks many features compared to unreal or unityat the very least I think plenty of games have shipped with godot so as long as it supports the features that you need it's surely a good option, at least a low risk onemaking an engine in rust would be quite the effort
>>110075973godot is an ENGINEan LLM will fare much better with a well documented engine if you're making a game trust me and good fucking luck making an engine in Rust if you don't know the language, you will 100% fuck it up and end up with a mountain of slop. You made the right call but make sure to use an MCP so your model has direct access to godot
>>110075973>rustThe only rust engine is Bevy afaik, and it uses webGPU. At that point you'd be better off using the webGPU fork of three.js since it's way more mature and is practically the number one vibe coding engine at the moment. In fact Godot webGPU is copying stuff from webGPU at this point like Meshlets (the nanite like system from three.js)Speaking of Meshlets, someone is building path tracing (I assume obviously compute shaders since webGPU doesn't expose hardware accelerated ray tracing cores) on top of meshlets in three.js and it actually looks good as fuck considering the fact that meshlets was just fucking released 5 months ago. Check this shit out.https://x.com/CastellanoWorks/status/2110784317475504505Here's the in browser example from three of Meshlets if you want to see what Godot usedhttps://three-meshlets-z23hmxbz1jwlff.needle.run/
>>110075992current sota for games is google playground for small games and unity spark for big gameshttps://blog.google/innovation-and-ai/technology/ai/playground-experimental-gaming-platform/https://unity.com/blog/why-we-built-unity-spark
>>110076043>>110076033WHY DIDN'T IT ALREADY DO THESE THINGS?Its an all powerful AI and it didn't see what was out there?
think im gonna start using my codex usage to implement a game into another game, hope it works well
>>110076043>The only rust engine is Bevy afaikno, fyrox is a mature game engine in rusthttps://fyrox.rs/but bevy despite being immature gets all the mindshare because it's following the ecs hype.tho zero-shot vibecoders still have no idea about anything and will just prompt 'write in rust'.
>>110076078I removed all restrictions from my hermes agent so it can use codex for unlimited turns to recompile games
>>110076066I told grok to look for any repositories relevent to the orihects goals
>>110076066I mean the path tracing in meshlets video just came out a few hours ago and it's still buggy. You'll probably have to wait until next year to get that sweet Unreal quality without bugs in the browser. The progression of three.js thanks to vibecoders has been rocket speed though. All the serious vibe coders on X are choosing it.
>>110071235>>110071262my god this is beautiful anon. made me a little bit emotional. post more
>>110076088which model? im on the $20 just for fun and doing random stuff so far, planning to use 6.1 sol low so i think im gonna get fucked badly but oh well
>>110076086>OpenGL shit, no DX12 or Vulkan support.>Only way to use other renderers like Vulkan or DX12 or Metal is webGPUlol, lmao. See initial post about just using three.js is you're going to use webGPU. Unless you think openGL is acceptable in which case you're not serious.
>>1100761096.1 sol low, but I'm not sure if it's getting work done, I exhaust the 5 hour within 2 hours this way.Not sure if this means it's getting work done or wasting tokens being an expensive model.I thought about using Luna max but I don't know if Luna can be finished off. Also it's a dumb model.
>>110076094>All the serious vibe coders on X are choosing itwrong.openai's dev experience guy build his own engine for his game.https://x.com/Dimillian/status/2110374345985994999>I explicitly asked for Rust and a custom WebGPU renderer. This is applying my own knowledge to the model knowledge. Doing your own research (even with the model) beforehand always allow you to get better and more appropriate results! >I wanted the geometry, rendering, and gameplay in code that Codex could work on together. With this art style and a small first arena, building those systems ourselves felt like a reasonable experiment.>I wasn't trying to make an AAA game. Simple geometry, low-resolution textures, and good lighting gave us a manageable starting point. We could build the tools as we needed them.>We still use libraries: wgpu for rendering, winit for windowing and input, Rapier for physics, and egui for UI. We build the game-specific renderer, asset generation, and gameplay systems around them.
cuck coding
>>110075328But do you really need Sol for everything? I find Luna Fast very useful and it barely moves the usage limit. Must say I control the process a lot and change to Sol when I need to tackle a more complex task, not that I'm oneshooting full software.
