A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs.## 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/----## Frontier models using fully-general tooling — start here if you have $20 or sohttps://developers.openai.com/codex/clihttps://claude.com/product/claude-code## Worth it for code, but the frontier models above are betterhttps://x.ai/cli## Not worth it for code, but maybe good for other thingshttps://antigravity.google/product/antigravity-cli----## Prompting / context / skillshttps://arps18.github.io/posts/claude-code-mastery/https://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/https://github.com/mattpocock/skills — /grilling is a favoritehttps://github.com/DietrichGebert/ponytail## Other editors / terminal agents / coding agentshttps://osaurus.ai/https://pi.dev/https://opencode.ai/https://cursor.com/docshttps://docs.windsurf.com/https://docs.cline.bot/https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent## UI/Frontendhttps://www.figma.com/make/https://www.anthropic.com/news/claude-design-anthropic-labshttps://uiverse.io/https://ui-ux-pro-max-skill.nextlevelbuilder.io/https://stitch.withgoogle.com/## In-browser builders / hosted vibe toolshttps://bolt.new/https://replit.com/https://docs.github.com/en/copilot/tutorials/sparkhttps://v0.app/docs## Benchmarks / rankingshttps://www.tbench.ai/leaderboard/terminal-bench/2.0## What we’ve donehttps://vcg.gitgud.site## Previous thread>>109387280
AI news brought to you by: gpt-5.6-terra (high)July 10th, GitHub put Copilot’s agentic security-alert autofix into public preview>Assign an alert, let the bot make/verify a fix, then review its PRJuly 14th, Cursor added Grok 4.5 across its apps, CLI, and SDK>new model, same “surely this refactor won’t break prod” energyJuly 23rd, GitHub’s Copilot cloud agent for Linear issues became generally available>assign ticket; agent works in a disposable environment and opens a draft PRJuly 24th, Claude Opus 5 started rolling out in GitHub Copilot>frontier agent slop is now available in even more IDEsJuly 27th, Cursor launched Router for Teams and Enterprise>automatically picks a model per request to cut agent costs without manually model-shopping
>>109393630My shitty script now also tracks the context
>>109393654
>>109393630Why'd the other general disappear?
>>109393675Who knows, probably got himself banned like yesterday.
Current vibe?
>>109393675Some luddite mod probably thought that it was made too early.
which models is goated at cyber and won't yell at me like an old crone?
>>109393694SMOL SMOL SMOL SMOL SMOL SMOL SMOL SMOL SMOL SMOL SMOL SMOL SMOL SMOL SMOL SMOL
>>109393694I hope this is AI. I want the secret
Goth Baddie 2 stronk
>>109393727I think it's just grok.
>>109393646I'm considering moving to terra for non-coding usageis it a smarter 5.4?
>>109393749>smarter 5.4it's smarter than opus 4.8 and gpt 5.5 lol. it's kimi k3 level
>>109393744from what I see online seedance is more coherent but gork has more delicious feel to it?
>>109393749It is a lot smarter than 5.4, yes.
feels like everyone that talks about projects are using codex. no opus 5 testimonies? no kimi k3?
>>109393809best value for your money, and intelligence no longer lagging behind anthropicchina and open models generally can't compete pricing efficiency with google OAI datacenter so I don't bother to check them out (poor)
>>109393838for me, kimi k3 just took too long. it was very capable, but took 15 minutes longer to do what terra max could do in 7. good to see open models bridging the gap at least
why and how is opus 5 slower than fable?
>>109393809i use both claude and codex, and lately codex’s weekly limit has been evaporating way faster.opus 5 is unironically much better now. i can get a ton of shit done even on medium. sol keeps overengineering everything, overcomplicates basic tasks, and is dogshit at orchestration. i don’t trust it as the main agent anymore.now i just use sol as an advisor subagent under opus. it’s still good at spotting random details opus might’ve missed, but i’m not letting it drive.also trying kimi through opencode rn and holy fuck it’s slow.
>>109393809I use the fuck out of Sol and now Opus 5, both are phenomenal. Kimi K3 is fine but it's so extraordinarily slow it's damn near useless for it, the new faster providers (as of yesterday) make it much more tolerable but that bumps the cost deep into the "not worth it" category.
>>109393910how much better is opus 5 than sol? >>109393891 says much better. im wondering if its worth it to briefly sub again, even if just for a month
steam should add vibedev section, reduce or remove posting fee
>>109393928I have a hard time saying it's better than Sol. Both are great, both are extremely capable. Opus is still vastly better at making a coherent UI. More importantly right now Opus 5 seems better at problem solving, something I consistently gave 5.5 and Sol over Opus 4.8, but they've tipped the scales on that one. Just for example, in the past Opus 4.8 really struggled when trying to put it through my Gameboy benchmark, but Sol would just go find opensource Gameboy games on github and steal from them directly. Now Opus 5 not only goes and finds these games like Sol did, it actually studies them and writes about its findings instead of just copy-pasting like Sol did, then goes back to writing its own code. So far only Sol and Opus 5 have actually passed that benchmark. Both are excellent and well worth using, and they make great partners for fusion reviews.
reminder that price performance doesn't matter if your model can't actually successfully complete your task
I've been using claude sonnet 4.6 to make my website, I didn't even realize there were better models out there lol
We're getting a new version of Sol and Terra this week. Perhaps Terra will actually be useful
>>109394023are they going back to the checkpoint system they used to do before gpt 5?
>>109393579Based, godspeed
Next reset when? Already spent my $100 codex quota. ALso tell me what you write in your agents.md or other .md files
>>109394172Thanks anon>>109394224I need another in a day or so
>>109394224christ lick her armpits
>>109393009>could pay for your meal and rent in an hourThat would be close to $5k where I live, would be nice to bill that rate.
>>109393876That hexes it. Musk psychopathically over-promises everything. What he says here just tells you what will NOT happen.
>>109394385grok 4.5 delivered pretty well
I've seen people with complex setups where they have one agent orchestrating sub-agents with certain skills, configs, etc. Is any of that even a net benefit? Seems like a lot of effort spent on trying to optimize something that is already optimal. I've never had any issues just using a single terminal with Opus and it spawns sub-agents as needed for larger tasks.
>rapidly approaching $200 in donations while the amount of users lined up to pay for the premium version of my app increases every dayholy fuck are we actually going to make it?
>>109394410zased. what is the app
>>109394418Can't say because it'll harm my reputation, sorry. Just know that I've put in 3 straight months of work for it to even get to this point.
>>109394457congratulations anyway
>>109394457Good job, anon, proud of you for not being a retard.
>>109393694Why does the speaker cone have cum on it?>>109394408I've been thinking of having one agent orchestrate sub-agents running on a local model to save tokens.
browsed some our thirdoid forum and they were saying openai is dying, and its cheap price is "strategy" and "last leg", and the latest 5.6 sol is less smart than opus 4.6 levelthe bias difference between bubbles is pretty hardcore, and chatGPT normie branding sure filter out the midshitters
>>109394410congrats!
>>109394410For me it's Steam (not a game), in sha allah this will take my sales to another level (I've been selling on my own website since 2008 but sales got bad in the last 5 years). I fixed all the bugs I know of with Codex so it's ready.
any creative voice use yet?so far I only managed to make it turn off by itselfthey forbid the AI from opening voice chat for some reasons
anyone have a good idea of how opus5 does vs opus4.8?I always wait a while before jumping to the next model and wonder if it's worth switching now
>109394561you wouldn't eat something a cat made
saas (snails as a service)
>>109394566I would if they got that excited while making it
https://www.anthropic.com/research/discovering-cryptographic-weaknessesJfc
qwen 3.8 qwen?
>>109394637nothing exciting, just some newshit algo get debunked
>>109394637>AES-128, the specific cipher we attack, has 10 rounds. Our attack works only on a modified version of the cipher that has 7 out of the full 10 rounds.yawn
>>109394637>https://www.anthropicYeah I ain't clicking that yiddish bullshit
>MCP is now statelessthen what's the point, just use script and data
>>109394515Good work, and good luck, anon.
>>109394410Fuck you, it should have been me instead, I hope you fail
>>109394692>Go you, it can be me as wellftfy
>our second attack is on a reduced version of AES and does not break the full cipher.3>Even then, the attack would cost hundreds of millions of dollars to implement and does not impact other similar cipher schemes.Hundreds of millions of dollars to attack a reduced round aes-128lollmao even
BASED OPTIMIST
I have defeated yet another Chinese printer. I don't want Xi's plintel_dlivel.exe or raber_mastel.apk any more than I want HP's 3.5GB "Printer Driver and Related Software". I really appreciate OpenAI not giving a fuck and allowing ChatGPT to find and decompile an apk for me without me ever needing to touch a Chinese website.
>>109394692I want us all to escape wagie hell
LLMs are good for porting shit and modding.
At least Gemini was honest.
>>109394845AI is going to need human subjects for all the bio-medical research it does, so it's probably better to let it use them instead of decent human beings.
>>109392872That's prefill aka prompt processing speed.Generation would be much, much slower.
>>109394845careful there, anon
>>109393809I am using Kimi K3.>>109393910Meh, it's alright.
>>109394992What did he say?
>>109395011He posted a variation of the UN schizo meme
>>109393630Hi lads, have one here made a C++ game engine with vibe coding? I want to start using it after seeing notch and linus bend the knee.
im just starting vibe coding and already im noticing a problem. chatgpt takes forever to get shit done because it keeps going on and on down rabbit holes and doing all sorts of shit going on tangents that seem to suck you in and keeps doing all sorts of shit like random little details then sort of repeating shit in a slightly different way each time. how do i prompt better to keep it in check so i get something in and out in one shot? because it feels like a fap session where you just keep clicking new tabs until you end up with like 100 open tabs of porn
>>109395053I should mention, is buying a subscription necessary for such problems? If so, what which one should I get for this task?
>>109393994sonnet is honestly pretty good for a free tier model
>>109395053>>109395109>which one should I get for this task?depends>the next unreal engineget claude $200 and whichever subscription has the best 3d modelling model>random rpg or simple 2d enginestart with $20 gpt
>>109395085work out a clearly defined project scope in a design document first
>>109395109Scam Altman needs more money for his second Koenigsegg. Pay his ChatGPT Pro subscription and generate more slop.
I fucking hate both of these slimy scheming snakes.
>>109395188I know how to model and have AI gen set up for illustrations i just need the code
Question, why is Opus 5 so good at oneshotting fully functioning games but when I ask it to do something simple like add ragdoll physics to godot or make HP system for NPCs it will fuck up 3 times in a row and then waste 200k tokens on a task that should have taken 2k?
just gib the codes already
>>109395362Because it doesn't know how to work with what you're giving it. Harness your shit, build out your agents.md, and give it the knowledge it needs to work within Godot and on your project. If you're not using one of the (many) Godot MCPs you should, they have knowledge of Godot systems built-in and give the model a much better starting point.
>>109394023>new version of Sol and Terra this week??
greeetings from China ohm-mericans, we are watching your vibe from time to time
>>109395362>talks about oneshotting games>thinks adding a ragdoll physics or HP system would only take 2k tokensyoure retarded
FUCKOFF
How long can we still make money from games and SaaS?How long until the models are so good, the average iPhone user can just go to the app store, type in the game they want, and Apple GameMaker powered by Fable 8.1 will just make it for them on the spot?
What does Kimi K3 use for image generation?
>>109394561its way better
>>109395230Makes sense desu, why should new Groks be released without facing the same scrutiny as OpenAI and Manthrobing models?
Secret models are always so enticing. What are you, kieran?
>>109395478>how long until people just buy a hamburger from Tesco and pop it in the microwave instead of going to McDonald's
>>109395478It all comes down to cost of generation. Even models as cheap as Kimi still take tens of dollars generating even the most basic games beyond just demos.Also I think we are rapidly approaching it, the oneshot examples on twitter look better then anything I have ever seen vibecoder produce. I think there is 90% chance that we will have one prompt fully functioning high quality SaaS and games before we get or at the same time as we get multi-prompt iterated high quality SaaS and games.
>>109395506>compairing a hamburger to 0s and 1s
>>109395519Yes
is fable the cocaine of LLM. the premium drug
>>109394410Brazilian users status?
>>109395487>But you're right about the ratio — the style reviewer flagged it too (~103 comment lines to 7 of code). Trimming now:I'm not impressed... I'll try turning it down to high instead of xhigh for the next run, but I think I'll go back to 4.8
>>109395482>Kimi>Image generation
>>109393809I use K3. Might pick up an OpenAI sub to cover some gaps. Fuck Anthropic, never giving them another dime after the shit they pulled this year.
>>109395482K3 doesn't have any native image generation. Through Moonshot, Kimi with the Image Generation plugin can use an undisclosed third-party image generation service. This is almost entirely undocumented, Kimi provides no information and doesn't acknowledge the existence of this plugin that they provide.
>>109395584kek, I appreciate you remembering that. Looking back, I think that was just me seething because the project ended up being far more work than I expected. Reality is humbling.
After optimizing it, K3 PP performance on a single pro 6000 up to 50 tk/s
>>109395895How can it go so fast on just 96 GB of VRAM?
