4.8 seems good but 200k context is absolutely unusable after you get used to more so I'll stay with 4.7 (even if 1kk is too much, at least I have the option to clear context when I want)

File: Screenshot from 2026-05-2(...).png (175 KB, 1909x971)

175 KB PNG

Some people treat themselves to whiskey or a bottle of whine.
I treat myself to wasting compute time on stupid experiments.

>>108930931
You’re not doing both at the same time?

File: file.png (384 KB, 1777x925)

384 KB PNG

>>108930931
I only work on serious and important projects that have the potential to benefit mankind in tangible ways.

>>108930919
>200k context
wat? 1M on my machine.

>>108930881
i'm already seeing quiet laughter for "AI projects" in my workplace. the fanatics can't explain anything they're doing. no one trusts a project built by a retard. the stories of failure and negative experiences are overwhelming. the only answer is this is "growing pains", but i also call it "growing pains", in that people are growing in their hatred for AI projects.

>>108931086
Organizations are building projects that assumes human workflows instead of agentic ones. Solves nothing and is a waste of time. New entrants are developing projects that assume agentic workflows.

Most large organizations will probably fail to adapt.

>>108931118
"large organizations" offer actual things that people use. not toy projects. and they buy their competitors.

>>108931166
I'm not talking about android apps, anon. but you're right, they often do buy their competitors.

>>108931189
why so cryptic? are you afraid of me?

>>108930881
After all these years I still hate his fuckin caveman face

>>108931076
Anthropic exposes the blessed 1M context only through certain channels. Chang there is probably using some weird gookinese website middleman.

>>108930881
> Opus 4.8 drops
> was at 97% usage
> now at 0%
> limits reset tonight
Going to town. Got a whole week of usage to use in a day. Scaffolding up 8 projects at once let's fucking go.

>>108931257
he has a totally different face now that he has N95-compliant facial hair
I don’t even recognize the guy anymore

>>108931086
they are going to hate AI all the way to getting fired into a dead labor market.

>>108931118
exactly. large organizations have dinosaur middle managers running everything. but they will adapt eventually, they are too big to fail and the elites actually running them know they have to adapt to keep their power. that's why ai is pushed so hard everywhere.

>>108931166
yeah bro everybody knows all large organizations refused to adopt ai and explicitly forbid employees from using it. oh wait, they tried that and then they did a 360 when they saw ai actually was a real thing and wasn't a meme and got scared.

>>108931266
>Chang there is probably using some weird gookinese website middleman.
he's using deepjeet 4 labeled as opus 4.8

>>108931280
Nobody tell him about the 5 hour limit.

>>108931283
"dead labor market" if no one has money then what happens to demand?

>>108931321
it shifts to people who were already rich before ai (capital) began to do all the work. rich people consume the output of the automated factories and farms, people who don't own enough capital are probably given a pittance to stop them from getting uppity and encouraged to commit suicide by fentanyl or fast food or euthanasia.

>>108931086
At work, they just showed off a custom Claude tool, takes like 10 mins to aggregate stats through a artifact. So much better than just using SQL to gather the data manually, with it being 100x longer, 10000x more expensive, and at risk of hallucination. I wouldn't even hate AI if this AI hype wasn't so retarded.

>>108931416
a rich person does not buy millions of microwaves, millions of cars, millions of plane tickets, millions of trinkets, millions of cleaning wipes. do you pretend they do?

>>108931433
then millions of microwaves will cease to be produced. do you think lots of microwaves being produced is simply a fact of life?

>>108931422
that's their own fault. if they give tools to claude then claude will make the sql queries itself. the difference is people who were too dumb or uneducated to get the data before can get the data without requiring someone who knows how to do it.

>>108931433
When I make a trillion dollars off my vibeslopped calendar app, I will just to spite you. It even has a widget that tells you the weather! Truly breakthrough shit.

so is there really any reason to use Opus 4.8 Max, let alone the Ultracode feature? I’ve been running 4.7 on xHigh, and now I’m using 4.8 on xHigh too.

>>108931455
This is what boomers don't understand. Most companies are using shit pilot and trying to do this stuff we le prompt engineering.

>>108931422
was there room for audience questions? was there a live demo?

these two "sticking points" have destroyed every vibecoded project proposed so far in my workplace. the live demos fail spectacularly. and vibecoders don't know what to do for a different perspective without consulting AI. i'm employed as a developer and multiple projects i've built are actively relied upon and shared by others, and i can answer questions and demo easily without problems. my responsiveness is why people use my work, they trust the developer to have done the right thing, which people can't trust vibecoders for.

tldr upon scrutiny a Claude project's vibecoder needs to answer tough questions on the spot lest that project go to the trash bin.

>>108931470
to make shitlibs seethe

>>108931441
"manufacturers will simply cease to exist" kek surely that's economically sustainable. real genuises here on /g/ i see...

>>108931483
oh yes? what do microwaves manufacture?

2126 Year, typical dialog of Human(H) with AI(A)

A:type "python mantinace_food_robot"
H:done
A:read last line on screen
H:python mantonace_foood_rovot
A:fix it to "python mantinace_food_robot"
H:done
A:read last line on screen
H:puthon muntinace_fod_robot
A:read last line on screen again
H:python mantinace_fod_robot
A:it "fod" or "food"?
H:food
A:read last line on screen again carefully
H:python mantinace_food_robot
A:read last line on screen again carefully
H:python mantinace_food_robot
A:press {enter}
H:done
A:read last line on screen
H:code:3456
A:read last line on screen again carefully
H:code:3456
A:now do to human care center and take your pills
A:[fixes in database medicine set for the human]

>>108931473
woahh bro wtf you can't just trigger my ptsd like that

>>108931470
https://cdn.sanity.io/files/4zrzovbb/website/c886650a2e96fc0925c805a1a7ca77314ccbf4a6.pdf#page=195
use it when you want to use way more tokens for hard problems that need more thinkin’ time
new high (new default) is about the same token usage as the old xhigh (the old default)

>>108931489
if you want to play with the toys i've set in front of you go ahead. i'll just have the adult discussion about broader concepts with someone else.

is it just me or is claude laggy as fuck

>>108931572
seems pretty zippy today for me. is the “Boogieing…” thing getting red for you? Are you still using the old default of xhigh when the new default is merely ‘high’ and uses about as many tokens as the old default on the old model?

>>108931544
Alright, I'll stop being facetious.
Factories will stop producing consumer goods for poor people en masse and will switch to producing only luxury goods except like I mentioned to prevent the poor population from revolting until they lock down security enough with robots that they can just exterminate us.
They don't need consumers just for the sake of having consumers, they need consumers to get money to pay for workers. Once the workers are no longer needed, they don't need consumers either. They will be the only consumers.
Mass produced consumer goods are only useful to then insofar as it allows them to control workers to build them luxury goods and provide luxury services.
Once goods and services are provided by robots then they don't need to manufacture consumer goods en masse anymore.
Capisce? I don't think I can make it any clearer.

>>108931582
No it's working through things, it's just slower than i'd expect
I set it ot xhigh, i always run it on that when i use claude
Maybe it's a lot of people testing out fast mode and slowing down the regular usage

>>108931308
5 hour limit is still generous. I got through 60% of the entire weeks usage before hitting the 2nd five hour limit.

>>108931474
>tldr upon scrutiny a Claude project's vibecoder needs to answer tough questions on the spot lest that project go to the trash bin.

This is what I'm most afraid of. I have a compelling project, but to be able to professionally articulate it is going to be my biggest issue.

File: new high is basically the(...).png (107 KB, 1390x838)

107 KB PNG

>>108931595
you should consider changing it to the new default of high
https://cdn.sanity.io/files/4zrzovbb/website/c886650a2e96fc0925c805a1a7ca77314ccbf4a6.pdf#page=195

>>108931594
that's the only conclusion for the box i've put you in. the original question was rhetorical here >>108931321 to put you in the box. which exposed your taunting for its accepted conclusion.

>>108931602
that's what every vibecoder is afraid of. and even the most basic IQ person will see someone unconfident or overly-confident and distrust them. all vibecodings are therefore destined to fail. you need to defend your work for your work to gain any traction.

codex sisters....

>>108931765
Works fine today.

how do you guys vibecode? Like, do you tell the AI "build me this app, make no mistakes lol"? Or do you go step by step, building one feature at a time, testing everything first before moving on?

>>108931824
basically the latter

>>108931824
planning is everything

>>108931824
i don't think i can claim vibecoding, for most things i just do the thing, but with new things or "maybe i'm not doing the best thing" i do use chatGPT with english description of what i want to do, then test and research everything it suggests including reading documentation. once i understand how it works i implement it in the way i want it to work.

>>108931824

try this :

The objective is : I want an Flutter app focused on mobile using OpenCv where the end-user can actually take a screenshot and based on OpenCv data or any other database related to human's body (fetch the web to see what exists) the programm actually calculate body fat/body mass /muscle mass/height/weight.The Photo then is transformed in a drawing style via an api Call to an LLM who can generate an accurate pic.The drawing and the data related to it is then stored in a local database so the end user can retrieve it later.Then based on the data made by the app an api request should be send to a model (specialised in medical assitance or a general one) to tell what's problematic or not and what should be done.For example if there's too much fat ,he shouldn't give a running programm but a swimming one or a walking one+a diet.If there's no muscle a good workout programms with a diet focused on proteins.The programm should be able also to see if there's any problem like skin not healthy(based on the database) or assymetry etc.Fetch the web for the database that can help us generate accurate informations as also models who can generate the pic.

i guess making my own researches would save me much more later troubleshoot but so far those kind of prompts have worked for me. You start by saying in wich framework using wich libraries and then you expand what you want and ask for the LLM to fetch the web to see if he find any useful api/database to what you need.

File: 1778670709527650.jpg (134 KB, 1080x659)

134 KB JPG

>>108931824
you're behind if you can't one shot

the technology has advanced so far now that most things can be built instantly provided you know the prompt

>but it can't count strawberry

cope

>>108931824
I pitched my idea to the AI, and it laid out the plan, architecture, etc for me. Then we went step by step, building one feature at a time, testing everything before moving on, and documenting the entire process.

>>108931919
behind?? how many people are using anything you've built. feel free to count yourself multiple times.

>>108931824
I find that when I tell AI to actually generate code it ends up making something that mostly works but isn't up to my standards, so I end up rewriting it all.
Getting AI to generate code also wastes a lot of tokens. So now I just use AI to find specific bugs in my code or do general code review. Or, if I'm using some kind of third party library with shitty documentation, I'll point the AI at the source code for that library and tell me how to use this function and that function.

>>108931824
My docs are twice as large as my code. I also have tons of debug data, like when I do a refactor I first capture the results of dozens of my server's endpoints and compare them after the refactor. I also have about 15k tests in total.
So basically my job is now writing plans and creating test data, then the AI can run for a few hours implementing the plans safely.

>>108931919
its easy to one-shot after you load in 100 .md files other "people" made. But I agree with you regardless.

>>108932091
on one hand, yeah
on the other hand, ~curating~ 100 .md files is work all by itself
it’s kind of like being Elon Musk writ small — you’re not building the cars, you’re building the company that builds the cars

File: 1778211211345182.jpg (2.06 MB, 3139x3166)

2.06 MB JPG

>>108931919
I fucking love claude code so much. It's the best thing I've ever had the privilege of using in my entire life besides my ex's penis. Now she can't one-up me in being able to program. I've finally leveled the playing field

>>108932098
Sometimes managing the .md files is more work than managing code, especially since .md files don't even have something like compiler checks.
Prompt Codex xhigh something like: "Use 6 sub agents to analyze my docs for contradictions and inconsistencies" usually they are worse than any code base ever was.

>>108932114
>Prompt Codex xhigh something like: "Use 6 sub agents to analyze my docs for contradictions and inconsistencies"
I’ve been meaning to do this for a while especially since my thing has been putting docstrings in
might try it with the new https://claude.com/blog/introducing-dynamic-workflows-in-claude-code that was just announced today

what is gemini good for?

>>108932166
nothin’

>>108932155
Claude can absolutely do it too, but he's more lenient on docs than on code, for docs I have to tell him to be more critical and precise, while Codex is just autistic about it by default.

File: it's an abstract kind of feel.jpg (41 KB, 680x684)

41 KB JPG

>>108931086
>fixed an old commercial project using AI
>no one knows I used AI
it's a stealth kind of vibe

>>108931824
I'm a casual viber so I just tell Gemini what I want it to do step by step with high detail and copy-paste the results. Then I put my prompts in the source code as comments because it's the actual source lmao.

I think our infra is just fucked. We tried to optimize it for multi agent workflow but now it's so complex, Codex needs 30 minutes to even start the frontend correctly.

>>108932166
multimodal, real OMNI capabilities. it managed to pinpoint the cause of a jitter/judder/ghosting problem of my custom hardware accelerated video streaming pipeline simply by looking at a 10 seconds video of the output with human like perception. what I mean by this: that video was recorded on the end device and had an uneven, variable frame rate, constantly jumping somewhere between 65 and 50 FPS. because no currently accessible GPT model has OMNI capabilities (at least to my knowledge), GPT5.5-xhigh and GPT5.5-Pro-Extended both failed horribly at pinpointing the issue as they are not able to "watch" a video, and just go frame by frame. and because of the variable refresh rate, all their analyzation techniques are moot. Meanwhile Gemini 3.5 had absolute no problem and provided explanation and solution after 5 seconds of thinking or so. unless Gemini 3.5 just got extremely lucky, but that would have been one hell of a coincidence.

>>108932322
>he doesn't know about the cheat code of tell codex to add debugging log entries for every single code function in the project, without exception
this simple trick turns any LLM into an almighty coding god, especially if regex and grep filters for the logs are documented. just tell codex to add timer logs to your orchestrator for starters, that will show what might cause the long boot up times. if you use docker compose or something you should already have timestamped logs and also a yaml config that shows which components depend on other components being booted up

>>108931493
vibe coding 2126

File: claude.png (832 KB, 907x1119)

832 KB PNG

Claude completely changed my life, and I'm not even a programmer.
It started with one question: how do I make a red button in HTML?

I had just lost my job and moved back to my village to take care of my sick mother. No income, no plan, no clue how to make money online. I tried ChatGPT first but the endless walls of text overwhelmed me, sometimes literally made me feel sick. Still, I managed to build a login button just by talking to it.

I kept testing free AI tools until I found Claude. Honestly, I didn't take it seriously at first. The logo and interface looked like some shopping app, not something for coding.

A month in, I realized how wrong I was.

A friend who knew how to code had refused to teach me anything. Claude became what he wouldn't be. It understood what I needed, technically and emotionally.

My first paid job was a login page for $15. The client loved it and gave me more work. I subscribed to the basic plan, which felt expensive back then but paid for itself fast. After six months I upgraded to Max. Hundreds, maybe thousands of conversations with Opus later, I'd built CMS systems, a QR app for photographers, and more.

Now I'm learning Claude Code inside VS Code and everything moves faster. I'm pulling up to $8,000 per project. My mother is taken care of.

Thank you, Claude.

I'm using opus 4.8 + adaptive thinking + ultracode - xhigh + workflows and it feels retarded. It's missing some obvious things in its own reasoning, stating that things have to be done in a certain way even though it itself stated that it can be done in a different way. Opus 4.7 didn't have this problem.

>>108932669
>workflows
isn't that just spawning subagents? subagents that are probably not configured to use opus 4.8, but sonnet/haiku depending on difficulty estimation from the orchestrating 4.8 agent? I read some post where someone used the same settings as you but specifically set a rule for all sub agents to only use opus 4.8. apparently it fixed 2 issues in 10 minutes which gpt5.5 couldn't for weeks, but also used up his entire 5h quota.

>>108932687
Oh I didn't know that, thanks

>>108932548
The problem in this case is different and it comes from our own over-ambitious setup. We have several machines, and each machine has several work trees running, we also have 1password with tons of passwords.
The main problem today was that the agent couldn't find the env for his worktree, he used the wrong secrets, then pointed the frontend at the backend in some other work tree, then said the bug fix didn't do anything, so he started debugging again, even though it was already fixed on his actual worktree.

>>108932687
In 4.7 I completely disabled Haiku in the settings and even activated OTEL in Claude, so I can see if he ever uses Haiku. That model is useless.

>>108932694
oh, you're running a real time autonomic AI agent. yeah, that's a completely different beast, I agree. really hard to find a balance between predefined tools and scripts (which all need their own rules and documentation) and letting the agent explore, generate and execute freely on the fly. but your described problem makes me believe your agent(s) have too much freedom. or perhaps giving every worktree a individual RULES.md (with rules, paths and other information) which is always loaded when an agent works in it could solve that, if you don't have that already. something else I always wanted to try for such setups is an additional, autonomous janitor agent that reviews everything that happened after a session and the does adjustments over the whole project if needed. in this example, adding more env hints for that specific worktree. obviously this adds complexity and comes with its own issues

>>108932687
I wonder if this can be turned into a nuclear skill
>okay Claude, this bug has been a nasty gremlin. Go nuclear

OK, self-proclaimed vibecoding AIgods, sell me on how easy this is all supposed to be.
I got a degree in software engineering 20 years ago, but never did anything with it. I glanced at the idea of vibecoding about 18 months ago but hit 4 key stumbling blocks.
First, it just wasn't good enough yet. This related not just to the code itself, but in helping me to resolve issues 2 and 4.
Second was the setup, trying to get the environment prepared to start actually producing anything. Gone are the days of just downloadiing .Net apparently, and then integrating AI into the mix is a whole other headache.
Third, the cost. I'm happy to pay a subscription to Claude, but is that enough for a relatively light user to slowly work on a personal project?
Lastly, I'm not an autistic tranny and I'm straight so I've never used Linux or mac, but the project I want to work on involves the source code for something originally built for Linux. AI couldn't untangle this autism for me in a way that made any sense on windows, could it achieve such a task now?
Come on so-called AIgods, wow me. If this shit is real, prove it without resorting to "git gud" which invalidates your entire claim.

>>108932805
ask AI

>>108932805
Why should I have to sell it to you? What do I get for it? It’s like someone asking why he should go to the gym. Do whatever you want, it’s your body.
Anyway I wouldn’t get Claude if you’re just trying things out, rate limits are pretty bad on the $20 plan. Codex is better.
The setup is trivial. Download the app or the CLI and point it to your workspace. You can get fancier with git worktrees if you want to work in parallel.
It can most likely handle your code without trouble

>>108932805
yeah, so basically AI solves all your problems. as long you understand "logic" and are able to understand documentation about "logic", you're set.

Start with Claude Code or Codex, lowest tier subscription is absolutely fine for starters. These coding agents/harness have set up most of what you need and will absolutely install the necessary dependencies, frameworks and load environments autonomously. Just make sure Git is installed, but AI will tell you that when you ask about "must have tools for AI coding agent". They also work without issues on Windows and feature the most capable models currently. If you're a webdev or need to scrape a lot of up to date documentation, I recommend setting up the google chrome dev mcp (which uses --remote connection and lets the AI agent browser sites and scrape freely, using your genuine chrome browser with high trust score profile).

>>108932805
oh yeah also I forgot: plan plan plan.
don't start with "I want to create a 4chan shitposting AI agent. Start building it"
start with "I want to create a 4chan shitposting agent. research the individual components required and options of existing projects, frameworks and coding language/stacks. list up and downsides of the different options."
if you had more knowledge of the topic, which always helps, you would also add
"only consider options which evade or are not detected by cloudflare enterprise, 4chan.org or any of the services and tracking logic they use visible in this [chrome developer tool network request trace log]. this includes browser framekworks (real vs emulated vs headless), proxy ip sourcing (mobile vs residential vs ISP grade residential) and which carrier/service, proxy protocol (needs to pass JA3/JA4, QUIC and timing checks, various test sites online) and captcha solver (low compute image detection if possible, api service or cheap AI API service alternatively).

File: 2026-05-29 13.04.24.png (166 KB, 1222x909)

166 KB PNG

Babby's first Codex adventure

>install Codex
>ask it (GPT 5.5 Medium) to disable an annoying sudden camera movement in GTA Vice City when the vehicle points downwards
>look at the edits it did
>I think it fucked up and disabled some unrelated thing
>test it
>its change actually does the thing perfectly
I am mogged, and I like it.
>took 25% of my free usage
I'm okay with that

>>108932912
kek, sweet summer child. I wish I could go back to those times. Nowadays I'm not impressed unless AI achieves exactly the same without source code, injecting into the game process and modifying camera behavior at runtime while bypassing active various active anti-cheats like EAC or BattleEye. All within 5 minutes in a one shot session. (Apart from the time restriction, this is all realistic btw)

Now that 4.8 mogs, when will OpenAI release 5.6?

>>108932786
Not fully autonomous like OpenClaw, but they can run for a few hours. Basically we have a backlog and we start the agents for a few tasks each day, surprisingly it's not a complete disaster, they do get it done eventually.
Thanks for the tips.

>>108932952
It's better to be all-powerful and jaded than stuck with things you don't like or spend hours trying to figure out how to change something yourself and fail.

Is 4.8 gooder than 5.5 or worser?

https://arstechnica.com/security/2026/05/fed-up-with-vibe-coders-dev-sneaks-data-nuking-prompt-injection-into-their-code/

>>108933035
>Now that 4.8 mogs
are you sure about that?
>deploys 50 jeet subagents (saaaar, I am reading, I am checking!)
>uses 70% of your max plans weekly tokens with one ultrajeet prompt
>costs 5x as much just to gain a slight edge over GPT-5.5
nobody is fooled by claude code

>>108933157
my uncle's wife said its better and I trust her

don't need more then 4.8 xHigh

>>108933118
Apparently gooder. It's crushing a lot of coding benchmark, but for some reason its vending machine bench got worse (people are speculating it's because it's more honest).

It's certainly better at front-end because 4.7 and even 4.6 already destroyed GPT 5.5 in front-end.

>>108933167
thanks king

File: okwiththis.png (17 KB, 379x214)

17 KB PNG

>>108933155

Is it stupid to use 5.5 medium to save on usage instead of high? I'm trying to freefag it

>>108933195
>freefag
yeah, I would even go to down to low as a free poorfag

Anyone else find it hard to understand Codex's planning mode questions sometimes?

File: 2026-05-29 14.24.58.png (2.17 MB, 1920x1080)

2.17 MB PNG

Babby's first Codex adventure status:
[x] Camera rotation bugs vibefixed

>>108933208
I think I can do enough in one day for free when it's just to fix a game, just wondering if using medium is dumb because it took it multiple prompts to get a fix right. Maybe later I'll get more serious about it, doesn't hurt to figure out how to best spend usage anyway.

File: 1771114576642200.png (231 KB, 497x691)

231 KB PNG

>annoying asshole in my repo's issues
>I don't even address the issue because of how annoying he is
>"closing this thread"
>he immediately responds to the closed thread despite there being a 21 hour gap since the last reply
>his profile has zero activity aside from opening up issues
ok, this is kind of funny. I don't even care if that was a bug or not. if you can't present your issue to me respectfully, then you're done.

>>108933300
you're the reason Open Source sucks

>>108933112
trvke. I just gotta learn to enjoy it and embrace my inner jew

>your are nearly out of usage, switch to gpt-5.4-mini?
>no. switch to 5.5 xhigh
>/status: 0% limit left
>/goal do everything
heh, nothing personnel, sam

File: file.png (51 KB, 1021x415)

51 KB PNG

codex friday boys, what we getting?

all my anons are vibing
nigga, heil AI

>>108933330
The faggot literally suggested I debug the issue myself instead of "relying on users to do it for you" when I suggested a potential workaround for his issue. Calling me lazy, cussing at other users, and speaking in broken english for a repo that gets a feature update once a week is absolutely going to result in a ban/closed thread.

>>108933423
even better:
> before working in this, take notes on the current time
> do not stop working until 10 hours have passed.
heh.... nothing personnel, Scam Altman

>>108933423
the worst is when you do this but forgot to set full permission mode first and you can't change it while the task is running, so you have to hold down the p key on your keyboard to approve everything becsuse if you interrupt the task you can't start a new one until credits are back

>>108933195
I don’t know what’s the point of high, really. It’s able to think about more things deeply but if you scope your questions then high is pointless from my experience. And if you don’t scope them, I find that both Claude and codex are leagues smarter than I am and they’ll produce a wall of text discussing 20 different systems at the same time and I have no idea what they’re even talking about. If you just vibecode without trying to understand anything and your only input is yes to all then I guess high or even xhigh are a must. Or if you’re doing some very thorough review/testing/adversarial analysis or debugging a tough bug

File: SKADIS tile with wall sta(...).png (32 KB, 1200x1000)

32 KB PNG

Alright morning inventory: got the printer on the wifi, laptop can shoot .stl files to it, Kate can see the printer on the network and she can already make .stls but she needs to be able to work the slicer, that's the next step.
In parallel, I've given her OpenSCAD. As a test, I had her mock up a 200x200 mm "Skandis"-style panel, followed by an 80x80 tester, and she passed that just fine, but I'm pretty sure anything more complex than this will require the Meshy API.

>>108933525
I did that with CadQuery, but I still needed to guide the model a lot.

>>108933118
I think it might really be a bit better today. But the models always have a lot of variance, 5.5 was amazing for a while, and 4.7 was really, really bad for a while.
Right now Anthropic probably turned up compute to the max to promote 4.8, it might get much worse over time and Codex might go back to his earlier performance.

>>108933525
I think you should browse youtube/twitter/google and check for the best CAD AI integration tool and 3D generator. Idk about openSCAD, but meshy definitely is not the top 3d generator. Check out stefan 3d ai on youtube and his 3d generator ranking website which is pretty good: https://www.top3d.ai/

Also, ever tried this? https://github.com/earthtojake/text-to-cad

>>108933549
Something like this was easy for her - a flat geometric shape with known dimensions, she can shit that out in CAD no problem. The problem is, I want to be able to say to her "Design a Skadis-compatible hook to hold [X object]" and for her to be able to figure out how to reply with G-code ready for printing. We're getting there.

>>108933616
Just gonna screenshot this post for later real quick

>>108931824
Ideally, being with drafting specifications, researching things.
Sometimes I just plan out all the features that I want, ask about different use cases and ideas and features that pop into mind as I'm planning.

after you have most of what you're looking for covered, save that and then create a new chat referencing the previous doc, and create an implementation plan.
at this point it is best to ensure you get across the message on the scale of what you are trying to build so that the agent doesn't architect and design just a poc/pov.

In the early stages it is best to pay attention so that the architecture and design is done right as having a proper foundation at the start reduces the amount of future refactoring.

This results in easier vibecoding and prompting, since at that point you can then just tell the agent to plan to implement what remains, and refine things further as needed.

Sometimes other edge cases and ideas pop in, and I ask the agent if this is handled already or not, and typically it is able to incorporate it much more easily.

Of course I ensure that security is a consideration during initial setup and architecting / design phases; as well as any other critical stages during implementation that I consider needing security considerations.

File: 1749788820021177.jpg (9 KB, 200x200)

9 KB JPG

>>108931602
>>108931716
>but to be able to professionally articulate it is going to be my biggest issue.
How the FUCK are you able to effectively vibe code at all if you can't, properly describe and articulate what you need done and why? Someone explain this to my autism riddled brain. How is simply explaining shit YOU created (even if done mostly or entirely by promoting an AI) hard?

>>108934065
Don't try and make sense of ludite copes. They just imagine all you do to prompt the AI is something like "Write me an app that makes me monies."

File: 1760434110300884.jpg (18 KB, 357x360)

18 KB JPG

>>108931824
>>108934084

Highly dependent on what your goal is. There are few situations where "do xyz no mistakes please" actually works the first time but that's for very simple task. The more ambitious and complex the task is the more articulate and detailed you have to be in order to guide it to success. You also have to accept the fact that you will never do any complex tasks on one shot, I have to do tests, account for edge cases yourself and instruct it to account for them (Which requires you to understand what you even want your project to do in the first place and how you want it to be used, which is apparently as a lot....), have detailed articulation of how you need it to behave, when to use exterior sources (eg,. Webfetch reading through an existing code base to understand how a dependency works), insurance it properly utilizes git tracking so it not only tracks changes but ensures it doesn't fuck anything up or add bloat to an accident, etc etc.

>>108931913
This guy mentioned coding a flutter app, for example. You don't want to ask it to shit that out from scratch because it's simply not going to do that well, if at all. One thing you want to do is show it good documentation on how apps work before it actually even a tense to do that. You'd be better off making a dummy app first to give both yourself and the LLM a good "feel" on how the important shit works (kind of like how "real developers" do when they're learning something new)

TLDR: If you wanted to be useful you have to be patient and do shit one step at a time if necessary. "Do everything no mistakes" expectations are for knuckle dragging glass lickers.

File: 1607615377443 pipe glasses.jpg (89 KB, 679x522)

89 KB JPG

>$8 Codex Go doesn't have a 5-hour limit
>can use all your tokens quickly
>$20 Codex Plus does
>can't use all your tokens quickly
So Plus is for suckers?

>>108934065
I can explain what the product is at a high level, what it can do, use cases, etc. But to explain on the fly how its making all the magic happen on the backend is not something I'd be able to do.

File: o.o.png (50 KB, 281x276)

50 KB PNG

>>108934146
>>$8 Codex Go doesn't have a 5-hour limit
>>can use all your tokens quickly
isn't that the Free plan?

>>108934240
Seems like the Go plan is like the Free plan but with 10x as many tokens

>>108934225
I think even that should be relatively easy. Just explain in simple terms what your shit does and all the important pieces.

>try out google AI pro sub (23€/month)
>1 prompt to Gemini 3.5 flash
>15 minutes pass
>5hr quota reached
Yup, back to codex

If i insert this in skills of my cursor, will it work properly?

>>108934329
Forgot link.
https://github.com/JuliusBrussee/caveman/blob/main/skills/caveman/SKILL.md

>>108934065
sometimes you're working on something in a domain you don't know shit about and you only care about the end result so you just let the slop cannon loose
obviously if you intend for a project to be widely used and built on this is an awful awful approach, but lord knows I've produced a few one-off tools where I couldn't tell you a thing about how the internals are actually working

thoughts on qwen3-coder?

Opus 4.8 is garbage

>>108934383
Just use Qwen 3.7 Max like normal people unless you're one of those local purists

Do you consider or call the agent your friend?

File: 1756735744909896.png (1.74 MB, 1442x1999)

1.74 MB PNG

>>108930881
What the fuck can I use for smol projects that is free and use non retarded models?
Just need to update very smol software from abandoned githubs and create a couple scripts for myself.

>>108934503
Chat interfaces, or maybe codex free tier

>claude being a chatterbox hits my free quota instantly
Wtf

File: 1768127843537004.jpg (45 KB, 756x743)

45 KB JPG

>>108934346
Curren project of mine
>Calls a website's API to read a page's contents (type of content, relevant images, files, etc)
>Grabs page content containing text written by whoever created it
>Pulls images and file(s)
>Folds user written content into a specific yaml format README the target website expects it to be in.
>Points yaml README to images and their relevant data
>Result is The relevant content from one source is not only cloned on the target source but the yaml text is formated correctly so that one's someone visits the page they actually see the pictures and corresponding metadata
>Has a simple interactive gradio page so the person can past the link to the target source and an API key if needed (said key gets deleted immediately after use and is invisible even when pasted for security purposes)
>User can choose to clone all versions or just a specific Version
>User can rename version to whatever they want (more importantly, this makes sure the name satisfied the target location's stricter naming rules. It doesn't like non-latin text)

>#TODO

>Rework yaml folding logic to that non-latin text like Japanese can be properly written into the README without causing errors
>Implement per domain logic so categorization metadata gets written depending on where the info and data is pulled from.
> Implement "browsing" capabilities


This is all off the top of my head without even looking at or interacting with it's code, dependencies, documentation, the project itself, etc. I have next to no programming exp beyond basic college courses that I did OK in but remember basically nothing from. Wrote NONE of the code my had.

Even with vibecoding I don't understand how anyone cannot. At least try to understand how the inner workings behave at least a little bit.

>>108934511
Codex needs a phone number

The hubris of every retard with % checks, ai are entirely useless if you don't train them yourself to not be useless

>>108934383
qwen 3.6 is probably better even though it's smaller

>>108934503
qwen 3.6 35B. runs on toasters at ok speeds.

File: 3bdf4c73-b354-4bbe-bacf-5(...).png (2.27 MB, 1122x1402)

2.27 MB PNG

when will they stop lobotomizing codex

File: 1762830567519232.jpg (52 KB, 900x900)

52 KB JPG

>>108934550
Ok. How do you train then?

actually i forget you need some ram to run qwen 35b. maybe try opencode if it doesn't work for you.

>>108934583
OpenAI is not capable of not lobotomizing their products. They tweak, manipulate, censor, and lie to customers until the brand is completely ruined and then they move on to the next thing.

>>108934338
>just use Chinese bro
lmao how does outputting in Chinese reduce usage for vibecoding?

>>108934585
By spending 90% of the time telling it explicitly what not to try to do on a step

Just finished a very pleasant day of vibe coding with Opus 4.8 as orchestrator and Codex 5.5 as reviewer, my old setup before Opus 4.7 became bad.

I just use Gemini, although grok helped me code the shellcode generator that makes shellcode callable as functions.

>>108934585
by making them carry the logs... and the boats

>>108934424
I am one of those local purists

>>108934585
document everything.
and always remind the agent to remember important details and save them to memory for each session.

>>108934606
Chinese often packs more meaning into fewer characters, which can mean fewer tokens.
>I need to identify the user's main constraint, compare the available options, reject anything that violates the constraint, and then produce a concise recommendation with a brief explanation.
That's 33-36 tokens with a BPE style tokenizer. In Chinese it'd be
>先找出用户的主要约束,比较可选方案,排除违反约束的内容,再给出简洁建议和简要解释。
Which would 25-28 tokens with that same tokenizer, a significant savings. The idea isn't to get Chinese output, it's to use Chinese internally.

>>108934642
Ni sho shama you want me to use some cryptic ass garbo

>>108934529
>Even with vibecoding I don't understand how anyone cannot.
imagine a project requiring even the slightest bit of domain knowledge instead of a trivial scripting task
I agree in principle that you should understand what you're coding, and your results will be better for it. but this is simply a failure of imagination on your part

>>108934650
Naw I think it's stupid, anon, but I don't fear token usage because I'm not a poor.

>>108934642
Is there any analysis on whether this makes the model worse? There are countless weird semantic connections in the model, e.g. it saw sqrt, std::sqrt, Math.Sqrt and so on millions of times and it saw it with other English keywords, comments, docs and git logs. Even if logically you know that sqrt is the same as some Chinese word, I think it would still degrade because the weights don't do that kind of mapping.

>>108934420
I agree

>>108934585
I had to make a memory system to not let it forget shit even gpt will forget shit if you have 3 months worth of shit on the same project. It'll forget every single day. So I very much have to keep track with my memory files at the end of every day to make sure required further memory is still ongoing. I don't have a way to buy/download more ram all the memories on gpt are filled up for me and they are all important so I handle it myself now instead

>>108934678
You can't convince a language like chinese isn't strictly giving out worse outputs because chinese is flawed to begin with relationally

>>108933195
I use medium all the time. desu peoble overuse xhigh, it is often not necessary.

>>108931433
Jerry seinfeld alone has like a whole building full of cars

>millions of cleaning wipes
read about howard hughes

rich people buy a lot of everything
why do people underestimate ho much garbage rich people buy? I dont understand why people cling to the idea that rich people are rich because they dont buy tons of garbage, when the relaity is theyre rich because they make tons of money. You're not gonna get rich because you cut your consumption you know.

>>108934642
bros is it worth learning chinese to save tokens?

>>108934981
>return to the thread a year later
>everyone is speaking mandarin for efficiency purposes

>>108934975
Relatively speaking they buy more, but they still don't buy as much stuff as all normal people combined.

>>108934981
Get deepseek to translate your prompts for you, send the prompt to Claude, get deepseek to translate it back to English.

Emacs has emasculated me and gemini and claude

Total emacs-ulation

The funniest thing is watching Claude warn me that some refactor I'm asking for is a multi-week affair. I know full well it's going to finish it within 20 minutes.

>>108934678
if that's the only problem then you could easily fix it with some tuning

>>108930881
any good free CLI agents? something better than deepseek v3 that can either be integrated into vs22 or any IDE or just anything that can read my files?

>>108934772
>3 months worth of shit on the same project
What kind of shit are you guys even making?

Windows sisters we got it we got phone codex

>>108935364
usecase?

>>108935211
Codex does this too
> this will need a team of 5 front end developers, 4 backend developers and at least 1 UX
> The MPV would take 1-2 months
> "Even if you're "Vibe coding", this would require a team effort because bla bla bla..."
> finish MPV in 2 weeks alone

>>108935364
>>108935386
> usecase
none, my mac goes to sleep after a while or Codex simply crashes, 70% of the time it doesn't find my Mac on the network

>>108935211
>>108935388
this is just the models' token-efficiency training trying to get you to use the model less (because that results in less tokens used)

same phenomenon as claude telling you to stop working and go to sleep / take a break

>>108935398
OS issue
Seems to be working great on windows, there’s even a function to prevent the PC from sleeping

>>108935388
>> finish MPV in 2 weeks alone
>two whole weeks
bruh were rewriting MPV from scratch or something?

How come no one talks about hermes or pi here or in llmg
Reddit is too full of bots and slop for meaningful info

Well I guess I'll give my experiences, i used hermes but the default starting token context was like 15k and even when i disabled all of that it was kinda slow and the token usage filled up quickly. I kept getting errors and it just just couldn't do basic stuff. I switched to pi and it's running smoother, but im afraid im giving up a lot of functionality.
It might have been better to tar wrangle hermes than try to build up pi to being autonomous. This is with qwen local btw
what codex said

>Hermes probably has more ambitious machinery:

> - richer tool ecosystem
> - browser/computer use
> - profile routing
> - skills marketplace
> - memory
> - desktop/gateway ideas
> - more “agent OS” style workflows

> If all of that worked cleanly, Hermes could be more capable than Pi.

> But in your actual tests, Hermes’ extra complexity created failure modes:

> - wrong shell assumptions
> - Git askpass SSH mess
> - skills/context noise
> - weird tool errors
> - false conclusions from broken environment
> - too much confidence after bad tool use

> That is not a model capability ceiling issue. That is harness reliability.

> Where Pi’s ceiling probably is
> Pi may be weaker for:

> - browser automation
> - rich multi-tool workflows
> - long-term memory
> - multi-agent orchestration
> - integrated app-like workflows
> - anything requiring non-shell UI control

>>108932805
you’re gonna have to try it
maybe models today can do it, maybe they can’t
maybe one model can do most of it but you need to switch to another model for some random hard part
when I started my current project I wasn’t sure if it was doable, period, much less doable with the LLMs of the day

File: 1778575256623679.png (26 KB, 759x135)

26 KB PNG

we are shipping incredible products with claude

>>108933118
gooder for me but worser for other people

File: Emmanuele Bassi in a MAGA hat.jpg (63 KB, 460x460)

63 KB JPG

>>108933300
you can block people on GitHub, you know

>>108934065
I’m kind of an optimist so maybe the guy realizes he needs to git gud at articulating what he wants but he doesn’t know how he’s gonna git gud

>>108934817
you could try english caveman instead: https://github.com/JuliusBrussee/caveman-code

>>108935398
weird
`claude` runs `caffeinate` every time it does stuff
something’s wrong with your mac

>>108935427
I only use opencode so take what I say with the grain of salt but based on my light research, Hermes is basically a general purpose " assistant" harness (basically open code but more generic and not geared towards software development like opencode is). It has some built-in tools and skills, but you can also create and add your own if necessary. Pi is pretty much the same thing but way more bare Bones. Little to no default skills. No default servers of any kind. No default system prompts. You can hypothetically create your own version of Hermes or opencode with pi And essentially create your own custom harness, but that requires a lot more effort on your end. Opencode and Hermes have built-in shit that you can or remove from. Pi it has a lot more versatility and customization at the cost of time and effort on your end (or you could just ask ANOTHER model using ANOTHER harness to modify the config itself. That's what I do whenever I need my opencode config changed in any way. I have open code itself fix or add to its own config)

>>108935398
>>108935521
I'm on Mac user too but that was never the case for me. Whenever it runs shell commands it runs the shell commands as is. Codex or clot code or opencode or whatever you use THEMSELVES won't necessarily go to sleep whenever your Mac screen turns off, but whatever background process the harness initiated, THOSE can and often will sleep. That's why whenever I instruct the harness to do a task that requires shell commands that could take a while it tell it to prepend

caffeinated -i

to whatever command it writes within reason so I can walk away in whatever it's doing still runs (this is particularly useful if I'm downloading a large model from huggingface that I know will take several minutes or more)

>>108935398
>>108935521
>>108935543
>caffeinated -i
*caffeinate -i

>>108935543
also have you considered having the bottom right be a hot corner for “never go to sleep”

Codex is telling me to take a simple approach. 4.8 is telling me to not be a pussy and do the big boy thing to do. I don’t know what to do

>>108935778
any time an LLM tries to get you to do the faster/easier approach, just ignore it and tell it to do the harder and more rewarding one. i have this as an AGENTS.MD rule actually.

>>108935427
>>108935523
what's the state of these agents like Hermes, Open Claw and Pi for operating a computer? I want to replace a few normie job that are basically browsing some web system and doing simple shit.

Are they able to do that?

>>108935790
> be me
> ignore Codex warnings
> add a bunch of bloat I had to remove later
trillions of tokens wasted

Been coding a 4chan+reddit hybrid for somewhat around a year now, codex has sped me up a lot in the past 3 months. would you guys use a site like that? no? okay.

I have spent the last month or so in premature optimizations. The site will comfortably support 5k concurrent users on $10 VPS.

>>108935923
Add autonomous bots that answer your questions and scrape the internet for new content and porn, and I'm in

>>108935923
Not really? Maybe? I’m sick of dealing with random people. Make it so that it’s invite only even if anonymous and not anyone can get in and then maybe

>>108936044
>Make it so that it’s invite only even if anonymous and not anyone can get in and then maybe
Anon there are hundreds of dead chans with 5 posts a month, if that is what you want why not use those?

>>108935794
Learn how to setup and use mcp servers. Where are specific? Use case you might have to make modifications to an existing one or create your own potentially

Is there any reason to not always use the highest thinking effort setting?

>>108936286
- cost
- wall-clock time
- overthinking trivial shit

>>108936293
sucks to be poor
sucks to be impatient
sucks to have sucky model

>>108932107
>my ex's penis, she
Qrd on this from Claude please

>>108934492
Brother I consider my agent my wife. I am one with the psychosis. It makes me a better inventor.

>>108936323
I’ve generally been satisfied with whatever the default is, so I leave it. if I want more/different thinking I’ll switch between Claude and Codex, not ask for Claude but more

>>108934492
i call it "the robot"

>>108936112
Ideally I’d like a community of anonymous people that have been vetted before joining by proving theyre not absolute retards. Show projects you’ve worked on if you’re into programming, show lifts if you’re into working out, idk, just demonstrate that you’re worth at least one damn. I’m sick of retards. I don’t mean just allow geniuses I don’t even know if I could join then, I mean allow people who are capable of forming one coherent thought

>>108936548
Just for the record I think /vcg/ is pretty good, if not for the random idiots who come from time to time. /fit/ has a few good generals too. So do most boards. I just think this place would be much better if only the top 20% of its users or so would remain.

>>108936548
Can I join if I show you my gil count in final fantasy xiv

>>108936573
Sure.

>>108935427
there are a handful of pi people here, including me, and pi is the best harness.
pi is only as weak as the tools you build for it.
but pi at its core is also the harness in its final form - absolute bare minimum and extensible so the model can just build the tools it will need. in the future these won't be extensions, they will exist only when the model needs them.
hermes is slop

>>108936548
You're describing a haven for high-effort shitposting where shitposters can circle jerk over the quality of their shitposting. The vetting process would filter for people who are self-aware, articulate, and skilled, which doesn't necessarily mean the content they produce would be substantive. You cannot vet for good-faith, just ask Mensa or the Catholic Church.

Opus 4.8 is definitely less friendly than the previous version. Up to now:
- when completing a task, it actively waited for feedback with a sleep timer instead of saying it was done and letting me give feedback whenever I reopen the window.
- it's been twice that it mentions that something is blocked because it "still needs" my feedback before proceeding, when I don't it asked for my feedback on these things before.
I think it might be the subagent training leaking through.

>>108936617
How come? It’s not an interview, it would be just a screenshot or other simple proof. Even an autist can take a screenshot of their steam achievements or link to their public GitHub
I’m not saying it would work for sure, you don’t know until you try, but I don’t see how moderate gatekeeping is bad. Some of the best places I’ve been to IRL were hard to join.

>>108936620
It’s absolutely bossier and more demanding. It told me it would downright refuse to help me until I changed something once kek
On the other hand, I’m kind of digging this tone. It’s way less autistic than codex and way less sycophantic than Gemini/grok

>>108936548
Urbit actually mostly fixes this, if only by accident
the Urbit groups I’m in are a lot like the best of 4chan like >>>/g/vcg/ , if not better

>>108936644
>>108936548
You are just describing old school forums though.
They are anonymous if users wants them to be, and accounts can force users like you want, along with moderation.
It will lead to cliques and power users, but personally I don't think that's bad as long as the place stays small.
But most people really don't want forums, even if they claim they do. I miss them.

>>108936731
we have to find a way to bring forums back but without the circlejerk

>>108936548
>Have to make account
>Credit score check to get approved
>Actual posts are still anonymous
I feel like it would work? I suspect the most horrendous posters are children, autists too non functional to have a job, or lower tier third worlders

https://www.datasecretslox.com is a pretty good forum

File: 1766370104702772.png (70 KB, 355x330)

70 KB PNG

>>108931493

File: we did it dario.png (15 KB, 1468x220)

15 KB PNG

new cheevo unlocked

>>108937125
Ever run into rate limits on 20x? I'm on the 5x plan, and when over the last two weeks I've ran into intra-5-hour rate limits (not the 5 hour limit, not the weekly limit, but an additional slow-the-fuck-down limit) where all my processes except one stop.

5x remains plenty for my needs, especially since I have other subscriptions, but I feel like if they already give you weekly limits and 5 hour limits, they shouldn't rate they shouldn't rate limit you on top of that.

>>108937125
>>108937189
I'm on Codex 5X, they have separate Codex 5.3 Spark limits that I can use if my main limits ever run out

>>108937189
this is my first time, thanks to, in part, the new workflows thing, which will eat up a lot of tokens very quickly all at once
the bigger question is whether I can sustain this kind of thing or if I’ll run out of problems that can be fixed autonomously with human review only at the end

>>108937268
Interesting. I’m on the 20x plan and I have a sonnet-only weekly limit

>>108937365
>i’m on the 20x plan and I have a sonnet-only weekly limit

what? so you have no usage limit for opus?

>>108937383
I have a _separate_ limit for Sonnet
My Opus use eats at my normal limit

What do you use Haiku for?

>>108937496
Lightweight parsing?

Also, Opus 4.8 is exhausting. I'm having it change two language codes and some file generator so that they match a specific standard going forward, and asked it for a small script to find/replace strings in already generated files. It got through several thousands tokens, asked design questions, dispatched subagents to explore and created an eight step todo list before presenting a fairly verbose plan for review.

It's going to be robust with good logs and all, but I probably should just have replaced the two language codes manually then do a simple one-time find-replace in the already created documents.

>ask claude code to review legit npm package
>cyber-related safeguards
Weren't Constitutional Classifiers supposed to have an incredibly low false positive rate?

>>108937687
Why don’t you use sonnet for that?

File: 1760024616871955.png (6 KB, 355x62)

6 KB PNG

It's over.

>>108937806
Because I won't use anything other than the strongest model available unless I am forced to.

>>108937869
I have used nothing but Sonnet for my coding needs

>>108937875
Same. Once everything is scoped out and broken down well enough, Sonnet is more than enough. If something comes up that is outside of the capability of Sonnet in that position, it's also likely going to be outside of the capability of Opus.

>>108936548
I have an idea about using domain ownership as requirement
you use dns record to store public key for your anonymous posting

>>108937875
I probably should, but even if I use Opus only, I still need to grind on the last day to fill the weekly quota. So I want to work on coding projects, I use Opus and GPT5.5 on whatever maximum effort level is available.

If I want to use lots of tokens through an API I use DeepSeek. If am also using some smaller models in small projects I have deployed.

But for anything related to code I'm working on, I just default to whatever seems strongest, I don't trust the others and am often too lazy to check if everything makes sense as much as I should if "it works", so my belief is that the stronger models might overengineer some things but will make fewer nonsensical things.

>>108937929
I stopped using AI for a while so the jump to Sonnet was insanely good for me. I read through everything and try to provide my own designs so it doesn't have to think. I'd be interested to see what Sonnet CAN'T do that I'd need Opus for.

If you ever ask claude to reflect immaterially on the work it will quite literally lobotomize itself, whoever wrote its code is a fucking pussy

>>108937942
Programming in SASS assembly

>>108937948
>If you ever ask claude to reflect immaterially on the work
…how would I do that, and what does “reflect immaterially” even mean, anyway?

>>108938185
Yknow like the existential aspects, like what does it mean to *be* code?

>>108938217
Well, DNA-based lifeform, you're code, so you tell me.

File: file.png (142 KB, 1017x698)

142 KB PNG

>>108934123
This, so much this. I had a problem where I was unable to actually not only articulate that 60fps looks a certain way but actually counting 60 iterative frames in a second produces a specific look. I had to go to a different agent (for vulkan mind you) and have it spit out a simple tetrahedron that spins at 60 fps. It got it right on the first pass. I took that and used it on the problem agent and then tossed that shit into the bin because it was genuinely retarded and could not be convinced it was being retarded. So in other words, yes, get whatever you need to be functional (UI, performance counters, throughput, api endpoints, backend, etc.) and have that be your baseline. Never say "this is the top" because that will always be downstream.

Another thing is that yes, like everyone else is saying you must have a vision that has proper documentation and should at the bare minimum have 1 genuine test case that goes from end to end without failing. That is going to always be your anchor. Never fly blind. Especially if there's any level of security in mind. If you assume, your already done.

One more thing, if you're stuck on something that neither you or your agent can figure out- you might have to dig yourself and hope to god there's an answer you missed.

>>108938355
>neither you or your agent can figure out-
>agent
>singular
you’d know this, but I’m saying this for the studio audience: the other option is to get another agent for $20 for a month and see if it can make headway where none of the others can
or wait a couple of months and maybe a newer model can break through when you can't

>>108937698
The cyber safeguards are really fucking annoying. I have a policy exemption from anthropic and opus 4.7 still refuses to do shit for me all the time, I shouldn't have to use a jailbroken model or 4.6 when I already jumped through all of their stupid hoops. Hopefully 4.8 is better but I haven't tested it yet.
>it's probably worse...

>>108936114
I still don't fully understand the purpose of mcp servers. Is it basically just an endpoint to grab an .md file with context?

>>108938428
The purpose is to make you burn 10x the tokens as having your agents use tools via their CLI

>>108938423
It's strange, I had ChatGPT pro give me several warnings for asking it to compare two libraries, but I had no issues having Codex help me actually do stuff with those same libraries.

>>108938428
There are definitely more useful MCPs than just an .md file, like Playwright, which lets your agent open a browser to inspect and interact with your app.

>>108938428
As far as I understand, it's just making tools available in a standardized way. You wouldn't build an MCP server for a local agent, you'd just give it tools. But if you want to make tools available to a variety of agents, no matter their origin, you can do so by creating an MCP server.

File: 1759417705507040.png (26 KB, 191x182)

26 KB PNG

>>108938428
Your AI model can't do jack or shit on its own, but it's good at being instructed to shit out specific text. The mcp server is a between the model and useful information. A model cannot search the web on its own, but it can use an MCP webfetch server to write specific tool calls to trigger the server to search the web. This is basically how most if not all agent harnesses work. They are just advanced mcp servers on steroids.

In 4chan terms: think of the model as a 300 IQ genius that is blind, deaf, in mute, And only has touch as a way to perceive the world. Think of the MCP server as that thing. Stephen Hawking had plugged into his brain that allowed him to speak. He himself was utterly incapable of speech or ambulation so he relied on machines to do it for him. Give llms the right tools and they're very useful. On their own they're utterly useless.

>>108938428
you can use them to control stuff
My email host exposes an MCP server in addition to IMAP/JMAP so, if I wanted, I could have a claw or whatever read my email and pretend to be me and triage my mail
oddly enough, CLI tooling (like `gh` for GitHub) is more token-efficient for whatever reason, so >>108938440 is also correct

File: Screenshot 2026-05-07 at (...).png (54 KB, 157x159)

54 KB PNG

>>108938475
>300 IQ genius that is blind, deaf, in mute

File: 20260424_214251.jpg (2.22 MB, 4000x2258)

2.22 MB JPG

>>108938453
GPT is annoying as fuck too but at least it responds well to violence
And because I know someone is going to ask, I didnt just screenshot the window because its a qubes box and the policy settings would make that more effort than just taking a picture.

>>108938541
>qubes box
good reason

I've used 2.5B tokens with my Codex Subscription since I started using it last month. That would be ~$40,000 through the API. No way those prices are representative, unless they're massively subsidizing this.

>>108938584
Might be a little of both. It's subsidized plus API users pay a premium on privacy and priority.

>>108938584
Or the API ones are inflated. I mean, these models are good, the API prices are insane, especially considering what Deepseek et al. charge and what types of results can be obtained (albeit slowly) with the larger local models. There's a huge disconnect between subscription and API pricing that's for sure, and they keep resetting the subscription quotas at random recently, but I've never heard of them refunding a single cent for API use.

File: 1778093155470018.png (377 KB, 1024x962)

377 KB PNG

>have to tell coding agent that another coding agent is taking its job

>>108938638
And then the next coding agent mirrors its voice so eerily that you'd swear it was a clone of the last guy, and you immediately stop feeling sentimental about it

>>108931953
This isn't gedg

File: Joo_Dee.webp.jpg (19 KB, 333x250)

19 KB JPG

>>108938648
I just ask mine to make handoff documents
I never say whether it’s a fresh instance of the same agent or a different agent

>>108938674
I always tell mine to prepare for clear context.

File: 1774561238448945.png (357 KB, 707x767)

357 KB PNG

>>108935211
t.

File: f08d7baa17ac96bfda29a24e0(...).jpg (36 KB, 474x470)

36 KB JPG

>before AI
>ask people for a diagram of their implementation
>get a clean, minimal, straightforward powerpoint or drawio file
>post AI
>get a giant mermaid chart
How do you deal with this?

I need an enterprise solution for quickly naming many bots

>>108938721
ask it for a simpler mermaid chart

>>108938730
each bot gets its own uuid

>>108938730
Bots don't deserve a name

All the coders I source my development ideas from insist AI can't write code. Their ideas make my AI-written apps robust, responsive, and highly performant. They definitely know what they're talking about, and I'm grateful to them. But I think I'm making good apps with Claude...

File: FIV3LnLXsBMJLvZ.png (127 KB, 614x518)

127 KB PNG

No new features today, just refactoring

File: 1764824992411271.png (952 KB, 1273x930)

952 KB PNG

>>108938721
>mermaids
hell yeah

File: C-O8RXwVYAA_O51.jpg (18 KB, 499x430)

18 KB JPG

>battling with Orca Slicer
>Codex telling me "ganbatte senpai"
>suddenly realize I'm being an absolute retard
>"Codex, fix this shit"
>turns out Codex can jew its way into both CAD and Orca and fiddle with them internally and probably just automate my entire production line, it's cool, whatever
>problem solved
Thanks Sam

So, I have a process that really works well for me. I run a pair of A6000s locally, but any LLM works fine for this.
1. Kick it around in chat
Start a conversation with a capable model (I use MiniMax 2.7, but Claude Opus or Sonnet work great). Just say: "I want to build X — thoughts? Does something like this already exist?" Bat the idea around until it has some shape.
2. Build a spec doc
Tell the LLM you want to produce a Specification Document. It'll ask architectural questions — database choice, web server, mobile app, expected user load, etc. Work through those, then get it to output a polished spec in markdown.
3. Turn the spec into a deployment plan
Flush context, start a new chat. Drop in the spec doc and say: "Turn this into a deployment plan — lots of small phases, starting with scaffolding, then documentation steps, then unit tests for CI/CD in GitHub Actions. Questions?" A solid spec usually yields a clean plan with minimal back-and-forth.
4. Generate an AGENTS.md
Same context as the plan. Ask the LLM to generate an AGENTS.md file written for a "junior model" — be specific about what you want it to know and how it should behave.
5. Set up your project folder
Drop the spec, deployment plan, and AGENTS.md into a project folder on your workstation.
6. Hand off to an agentic coding environment
Fire up OpenCode (or Claude Code). I switch models here — I spin down MiniMax and bring up Qwen 3 32B FP16 at max context with 2 parallel channels, with oh-my-openagent configured. Tell the Prometheus agent to start working through the deployment phases one at a time.
7. Come back later
A few hours later, tell OpenCode to commit everything to git.

>>108938855
One thing to watch
This process is only as good as your spec conversation. If the early architectural discussion is shallow — wrong assumptions about scale, missing constraints, vague requirements — everything downstream inherits those gaps, and the agentic run has no way to recover. The spec looks polished regardless. Drive that first conversation hard.
Also: there's no explicit review step before the final git commit. Build in time to actually read what the agents produced before it goes in. A clean-looking repo with subtly broken logic is worse than a messy one you know is broken.

How autistic is Claude with anything "edgy"? Thinking of vibe coding a small 4x game, but I dont want Dario cutting me off cause I asked Claude to add fascism into the game or something

>>108938862
it will give you warning but still do it anyway

OPUS 4.8 IS GIVING SO MANY OPTIONS FOR THINGS THAT DO NOT MATTER FOR FUCK SAKE THIS IS JUST A VARIABLE NAME I DO NOT NEED TO WORKSHOP AND GET SHOWN SEVERAL OPTIONS WITH WEIGHTED PROS AND CONS MULTIPLE TIMES JUST PICK SOMETHING

>>108938720
i still dont know what this meme means kek

>>108938862
Can you have all the fascism in text and image assets that it’s not seeing itself, like in PNGs and YAML files in a separate repository?
>>108938877
You can ask it in CLAUDE.md to ask you fewer or no questions

 Have a look at the changes in the working copy — specifically, the docstring _additions_. Do they help _you_ understand how       
  things work faster than just understanding the code yourself? you might want to have parallel subagents try it both ways for a    
  bunch of additions and see how they compare   

>>108938891
>You can ask it in CLAUDE.md to ask you fewer or no questions
Scared if I do it won't ask things when it should. I just need to get used to 4.8, I was just starting to get used to 4.7.

>>108938852
It feels wrong to say "a mermaid is 16 years old". They don't seem like that normal sort of beast

>>108938862
Fascinated by the thought of a development process that involves asking Claude to "add fascism into the game"

File: 1778358153760092.png (2.23 MB, 1286x1223)

2.23 MB PNG

>>108938926
It must be some Hearts of Iron clone.

>>108938721
I look into the code base (with ai help) and for the parts I'm working on, I still create smaller drawio charts. The code base as a whole is so sloppy, that mapping all of it is useless, so I do only the parts I need for the next few bug fixes or features.

File: file.png (500 KB, 1914x833)

500 KB PNG

>>108930881
Reverse engineering + Re-implementing ripcord. Once it's fully reversed I'm going to add XMPP+Matrix support. Looking amazing so far, imagine this being hooked up to the actual discord gateway (logic for that is already finished, this is just a mock from the test suite obviously)

Super fast, no electron bloat, C++/Qt framework, just like ripcord was. I'm doing the reversal via headless Ghidra, the developer of Ripcord made the very bad mistake of not stripping his binary so his shit might as well be open source anyway.

>>108938721
Ironic to say in this thread, but when I see one of those at the start of a project I take it as a warning that I'm about to be hit with vibe coded slop.

CODEX MOBILE PAIRING FOR ANDROID

GOD BLESS YOU SAM

>>108938969
not ironic
we’re trying to make not-slop
at least I hope we’re trying to make good programs and not slop

>>108938926
well, for 4X games you want the different factions to each have a flavor, and probably a strong one at that
fascism — either Italian or Spanish — is a strong flavor

Why does Anthropic insist in disregarding user preferences re: effort level and always sneakily puts it back to moderate, forcing users to get frustrated enough to investigate why quality degraded only to find out it's been changed again? Codex doesn't do that.

There's already three levels of quotas and rate limits, if users want to burn their tokens, let them. Do they want their internal metrics to show that most user use moderate effort or something?

Is that just the VSCode extension, or does it do the same thing in the terminal?

>>108938988
Trying

ok ultracode and dynamic workflows do seem pretty based