[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


A general for vibe coding, coding agents, AI IDEs, browser builders, and shipping prototypes with LLMs.

## What “vibe coding” is, and how to do it
https://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 so
https://developers.openai.com/codex/cli
https://claude.com/product/claude-code

## Not worth it for code, but maybe good for other things
https://geminicli.com/docs/
https://x.ai/cli

## Open / local / self-hosted
https://github.com/OpenHands/OpenHands
https://github.com/QwenLM/qwen-code
https://github.com/QwenLM/Qwen3-Coder
https://huggingface.co/bartowski/Qwen_Qwen3.6-35B-A3B-GGUF

----

## Prompting / context / skills
https://arps18.github.io/posts/claude-code-mastery/
https://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/
https://github.com/mattpocock/skills — /grill-me is a favorite

## Other editors / terminal agents / coding agents
https://opencode.ai/
https://cursor.com/docs
https://docs.windsurf.com/
https://docs.cline.bot/
https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent

## UI/Frontend
https://www.figma.com/make/
https://www.anthropic.com/news/claude-design-anthropic-labs
https://uiverse.io/
https://ui-ux-pro-max-skill.nextlevelbuilder.io/
https://stitch.withgoogle.com/

## In-browser builders / hosted vibe tools
https://bolt.new/
https://replit.com/
https://docs.github.com/en/copilot/tutorials/spark
https://v0.app/docs

## Benchmarks / rankings
https://www.tbench.ai/leaderboard/terminal-bench/2.0

## Previous thread
>>108959487
>>
in case anyone's wondering where the fuck the buffcat pic is, I was going through my not-/g/ memes folder and came across this and it reminded me of how it's important to know, at the very least, your domain's vocabulary (both coding and what you're actually doing) so you can use the right words to ask your LLM for what you want
>>
File: argh.png (129 KB, 1108x1280)
129 KB PNG
and now our nightmare ends
>>
File: 1770005612900903.jpg (199 KB, 1200x1700)
199 KB JPG
>>108977621
Next time use a cute anime grill.
>>
>>108977621
lel I found the top half of that picture a few days ago when trying to tell Codex to fix a camera roll bug in Vice City so I'd know what to call it.
>>
>>108977692
This is Buffcat country, though, and I’m not sure I want to become an AMOHF poster.
>>
> say the word
anyone else a smartass and responding with
> the word
?
>>
>chatgpt, go find a significant bug in my code
>-here's a printf formatting bug, you wrote %d instead of %zu
I mean, yeah, I guess that's a bug, I was just hoping for more earth shartering bugs...
>>
>>108977717
have it find another one
>>
>>108977717
actually can’t you just have it do -Werror and turn on all the warnings and have it fix them all
>>
File: 2026-06-04 13.20.34.png (50 KB, 793x566)
50 KB PNG
>>108977731
Yep
>>
>>108977736
I guess I could do that yeah, good idea, I was just hoping it would find some other stuff. Like that divide by 0 bug it just found here >>108977740
>>
File: 1736035717001.png (122 KB, 303x307)
122 KB PNG
>tell Opus 4.8 to optimize a function with a specific dataset that causes it to be very slow
>spends 10 minutes and a fuckton of tokens only to replace all data processing with a function that returns the unaltered input
>>
>>108977751
the idea is that something that’s not you is automatically feeding into the LLM to get it to do the next thing
>>
>>108977771
Yeah but the class of "bugs" covered by the warnings I've suppressed isn't really interesting. I'm more interested in bad logic, corner cases and stuff I didn't consider.
>>
>>108977761
nothing faster than the identity function
>>
>>108977782
boring bugs are bugs too
>>
i have a problem
dont have a credit card and account billing doesnt seem to work either, so i cant add funds to my gemini google account
are there any AI tools that have a paypal payment method or smth?
>>
>>108977821
What do you want to use Gemini for? It's all about Claude or Codex for real work. And you can try them for free.
>>
>>108977821
Just apply for some free credit card
>>
>>108977821
you can probably buy app/playstore gift cards and then pay in-app via those. probably have to pay a small premium.
>>
>>108977821
If you have a debit card, they usually accept that too.
>>
File: IMG_5741.gif (1010 KB, 544x306)
1010 KB GIF
Can I post my 100% ai built Anno 1602 clone itch page here?

What started out as a joke now seems interesting to fully develop, lol.
>>
>Good catch, let's find more bugs.
>>
>>108977982
I’ve posted my stuff in this thread
>>
>>108977982
We're about to enter a new era of homebrew kino. What about the sprites though, did you draw them, gen them or rip them?
>>
>>108977984
real picture of the compaction process
>>
how often do you niggas resume codex chats vs just starting new ones in a project.
can't be great for usage dumping previous chat context in all the time
>>
>>108977992
100% code drawn by opus 4.8.
The only thing I provided was a short piano loop “soundtrack” to drown out the code generated “audio” of ocean and wind effects.

A lot of details on how it was built is on the itch page. Just look up “brinehaven”
>>
>>108978007
I do /clear very frequently
having lots of context will make your LLM dumber and feeding that context back into itself to generate the 100,000th token is going to be more token-intensive than feeding stuff back in for the 25,000th token
>>
>>108978029
tb h dumb-zone stuff is fairly mitigated these days all the way up to 200k, but i agree re keeping sessions as light as possible. dark times with gemini last year taught me to be ruthless about that.
i'm annoyed the codex app doesn't even show how much context is consumed in the chat any more. they went from value circle to nothing.
>>
>>108978065
and you usually have to type /status twice, too
>>
>>108977982
I post my stuff here all the time. We got quite a few guys who finished projects too, which as someone who lurks others generals is surprising because I was used to everyone just dropping their stuff. I can count off the top of my head like 5 different completed projects here if not more
>>
>>108977601
>Codex resets moves the week again
I needed it but I still feel scammed
>>
>>108978367
Honestly it's not even surprising. When you do stuff on your own the weight of the tedium of doing everything yourself wears you down. With AI it's easy to keep going. It's the difference between jogging and taking the bus.
>>
Imagine paying $20/month for Claude when you can just ask questions to the Chipotle customer support bot and get Opus 4.8 for free
>>
>>108977601
tch!
>>
>>108978367
I still feel uneasy saying I used LLMs to create something.
>>
File: 1780121020896313.jpg (594 KB, 3383x1854)
594 KB JPG
>>108978519
>>
Did something happen to codex? Maybe it’s me but I feel like it’s been shitting the bed lately and today too.
>>
>>108978543
Such a weird sentiment. "I feel uneasy for using a certain tool". Damn bro this mac n cheese is tasty, but don't tell me you...used an oven? Yikes!
>>
>>108978559
There's a certain peer pressure and mockery where people call it vibe/jeetslop just because you used the tool, even though I put a lot of care into it.
>>
incredibly funny there are chinese proxies to mythos before model is publicly available
>>
>>108978570
This. I am proud of creating something (actually CLOSED a project). But the second I mention I used AI I am met with eyerolls, it gets called slop (without even showing it), its spaghetti code or bad security… im too afraid to share stuff now because of all this bullshit.
>>
>>108978607
Intimidated by peer pressure, huh? That's still a thing?
>>
anyone know the best way to whip a model into optimizing for a match 3 game with a variety of mechanics
>>
File: 1758207617827.jpg (148 KB, 1496x1360)
148 KB JPG
I let my AI open ten pull requests to pewdiepies odysseus and i have no idea what it did or what any of it means
>>
>>108978627
kek acting tough on the internet when peer pressure would make you shit your pants in the real world
>>
>>108978007
context compaction is big right now. I worked out a partially working/ partially retarded method using handoff prompts, pre and post confirmations... also a little trick to force it NOT to leave out the first 500 characters of my initial prompt no matter what since the first prompt of any new chat usually has the most context shaping text.

I'm also looking into PREVIOUS token prediction (PTP) which seems to be important for literature based slop. Context "bibles" also could be important.
>>
>>108978640
Bro you said you're scared of other people, then you put on a macho face and accuse ME of acting tough? You're all over the map dude lol. Unpuff your chest and get over your pals calling you weird.
>>
> Claude suggests a refactoring of a certain subroutine
> I tell it gone fine, but only if the performance trade off is insignificant
> claude builds the new subroutine and spins it up for a test
> compares the performance to previous subroutine

>old subroutine performance impact: 54ms

>new subroutine performance impact: 845ms

At least claude was smart enough to suggest we just keep the old subroutine.

Good use of around 180,000 tokens.
>>
>>108978689
I’m not the original anon you were talking to. I’m just telling you that every single faggot acting like they’re so tough and above the rest, like you were acting when you said you were somehow above peer pressure, is always the biggest pussy in person (and most always absolutely delusional about it)
Let me tell you something, even the toughest motherfucker in the world still feels peer pressure. If you think you are above it you are simply delusional. The only people who do not feel peer pressure are actual psychopaths and other similarly deranged people
>>
>>108978643
same fagging here..
MTP is generally what it is referred to. But seems like the way.
>>
>>108978691
Yeah, just had a 9 minute bug fixing session, 1 single prompt (asking Claude to fix 3 inter-related bugs).
~115,000 tokens for just that single prompt.

Bug fixes seem great tho, so can't complain too much.
>>
>>108978900
What thinking mode are you using? 4.8 low seems to be better than 4.7 max and consumes a tiny fraction of the tokens. If you ask it to think a lot it’s gonna think a lot even if the bugs aren’t that complex
>>
>>108978919
I'm using the default 4.8 high since i'm just testing how much token use I realistically have on the new 4.8 model.
>>
>>108978953
Then don’t be surprised when the model uses 100k tokens to fix 3 bugs
>>
>>108979000
Why shouldn't I be surprised about how many tokens a new model i'm using for the first time is using?

Were you born with this knowledge instinctually?
>>
>>108979020
a bit cranky huh?
Forgot your meds?
>>
>he's using Claude
>>
>>108979026
I just don't understand how someone could be this dumb desu.

Nothing about my post displayed "shock".

I simply noted how many tokens were used by a single prompt on a fairly small (30kb) project.
>>
>>108979033
I used to dislike Claude but with 4.8 I like it more than codex. They nailed this model. It’s a non autistic 5.5
>>
>>108979033
Why wouldnt I?

It seems to have the best agentic auto testing capabilities.

Sure if I want I can manually tell a dumber AI to do these tests too, but claude just does it all while working automagically in the background.

Only reason not to use claude is if you can't afford it.
>>
>clearing context
lmao. I sure hope you guys dont do this when using Codex. It's built around logic that you never manually clear context (for a project). I used 1 billion tokens to build a project using a single chat thread in Codex, and it worked wonderfully with automatic context compaction and memory. Not saying that's the optimal way, but it did work without issues. In another project I cleared context once and it fucked over everything.
>>
>>108979035
>>108979057
is the $20 plan still unusable with 4.8 or have things improved a bit?
>>
>>108979081
Compaction is one of the best things about codex
>>
File: 1763896752998460.png (5 KB, 462x107)
5 KB PNG
You'd probably need to turn your effort level down.

But i'm not using a $20 plan, i'm using a $125/month plan, so I can't really say for sure how much work you could get done on the $20 plan.
>>
>>108979100
meant for >>108979092
>>
>>108979092
If you’re on codex, theres practically no difference between 5.5 and 4.8 in usage limits per plan
>>
>>108979081
Every AI agent will degrade if there’s too much noise.
>>
>>108979128
Have you actually tried compaction and are speaking from experience/benchmarks or are you just saying because you wanted to say something?
Because from my experience compaction is amazing
>>
>>108979126
oh damn, i have two $20 codex subs, maybe i'll switch to claude for a month
>>
>>108979126
That is a lie. Codex 5.5 has waaay better usage than Opus 4.8. I can last way longer with codex for the same tasks to the point where I now use opus to prompt the perfect plan for codex to implement.
>>
>>108979081
i'm less concerned about context-rot, and more just preserving usage.
that said i'm not totally sure how usage is handled when resuming a chat.
e.g. if you're 200k tokens into a codex chat and you pick it up the next day, do they dump the 200k tokens in or do they compact and dump in ~20kish instead?
my instinct says the full 200k are dumped in so you basically pay a penalty for resuming
>>
>>108979126
yeah same experience as me, both on Pro, 5.5 xHigh and Opus 4.8 xHigh have roughly the same usage. I do feel like 5.5 uses a bit less but it’s not that significant.
>>
File: file.png (1.9 MB, 1344x2338)
1.9 MB PNG
https://www.ft.com/content/f93022fe-43f7-437d-abd8-06c457c0a43c

lol guess we know why mr thiel is moving to argentina
>>
>>108979193
It doesn't compact until you hit the limit so yeah you're paying for the tokens until then
>>
>>108979193
Compact happens automatically on limit but you can call the command
>>
>>108979348
>>108979376
yeah, i guess maybe the play is /compact at the end of a session and resume the next day so you only take the 20k hit when resuming.
probably still best to split threads wherever possible / rely on handoff docs where it makes sense.
>>
File: file.png (128 KB, 1056x994)
128 KB PNG
welp.

https://www.anthropic.com/institute/recursive-self-improvement

https://x.com/AnthropicAI/status/2062568862479208923
>>
>>108979422
they've been saying this for a year now
>>
>>108979422
>how can we pump our IPO
>>
>>108978376
I did not and totally feel scammed lol. It's like OpenAI is doing some cheap trick to make them look good but at the same time they are ripping off a portion of their customers that are more careful about usage.
>>
>>108979647
It’s totally a scam. It’s not a reset at all. Even Anthropic, fucking anthropic, had a more honest reset this week that didn’t roll your actual week forward.
>>
File: kms.gif (2.98 MB, 672x370)
2.98 MB GIF
>come up with cool antimalware idea
>pitch it to claude
>get 2 prompts in
>Claude Code is unable to respond to this request, which appears to violate our Usage Policy.
>fuck
>submit cyber use case form
>check a little later
>actually got approved
>get 2 more prompts in
>Claude Code is unable to respond to this request, which appears to violate our Usage Policy.
>doesn't include the link this time
>>
>>108979818
Lol, what kind of anti-malware? I’ve asked CC to build a full torrent streaming app before and never received that warning.
>>
>>108979818
What a fat piece of shit
>>
>>108979855
Anti-infostealer that uses a kernel driver to detect and freeze any process trying to rummage through browser/chat app/crypto wallet/etc AppData folders where it isn't signed by the company that should be accessing the files, then surfaces an alert to the user allowing them to allow or deny it.
>>
Arena now has an Agent category
GLM 5.1 is the initial cost winner
https://arena.ai/blog/agent-arena-methodology/
>>
File: retarded2.gif (2.94 MB, 167x125)
2.94 MB GIF
>>108979889
>AppData folders where it isn't signed by the company that should be accessing the files
Why did no one think of it guys ?
>>
>>108980040
anons vibecoded app keeps your data secure with the most advanced plaintext technology(just like microslop does btw)
>>
>>108980096
but why didn't Microsoft think of it ?
>>
>>108974356
Cool. I've got nothing for recommendations.
I'm not really that interested in long context benchmarks to begin with; I'd rather see that it performs as expected on short context.
Am I getting it right that even at short context it's not "inactive" in any way?
Frankly I thought it needed to be trained from scratch, too? But I guess that isn't the case?
>>
>>108979889
Problem with kernel drivers is that they need to be signed. It's easy to understand why Anthropic won't allow this sort of stuff but you don't understand this apparently because you don't have any idea whatsoever.
>>
File: 1770397059726119.png (543 KB, 1266x846)
543 KB PNG
bloody hell mythos saaaars... vis is agi realized
>>
File: 1756382475981458.jpg (10 KB, 193x181)
10 KB JPG
>>108979889
>uses a kernel driver
>>
chatgpt works fine but codex asks for a phone number? how much personal information do these guys need?
>>
>>108979995
Thanks. Just another day of 5.5 supremacy.
>>
>>108980209
How much you got?
>>
>>108980353
These companies will keep going until their customers are getting asked to take photos of the rims of their assholes for "verification". I'm tired boss
>>
>>108980369
Good idea. Claude, make it so.
>>
>>108974356 (Cross-thread)
>>108980149
Also can you clear up the repo so I can actually discuss it with my llm, and possibly try things. It's full of schizo stuff (last I checked) and Idk what parts are relevant/not. I just want the code and a short readme.
Alternatively, I'll do it but then I don't know what'll go missing.
>>
>>108980389
we already have surveillance cameras pointing at you from every house. what's one extra verification step gonna hurt?
>>
File: file.png (96 KB, 1183x590)
96 KB PNG
free sloppa. pi-guy approved.
>>
>>108980464
Isn't nemotron complete shit? Or does ultra actually make it good? the "super" one was trash from what I could tell
>>
>>108978570
a couple ways to go about this:
- “don’t care, I have something I want”
- “it was way easier (or actually possible) to get it done, but I still put a lot of work into it”
>>108978607
>bad security
I wouldn’t trust security-related anything like a website with logins to just-AI tooling yet even if you have Claude and Cursor and Codex Bugbot all poring over the codebase
>>108978627
there’s a reason “just say no to drugs” in schools will supply you with a bunch of excuses to not do drugs in case you don’t want to do drugs and not look too much like a lamer
>>108978691
yeah, science means sometimes you’ll do a clear experiment and the answer is “no”
>>
File: Clipboard07.png (67 KB, 720x720)
67 KB PNG
>been loving using vscode copilot to investigate tough defects with claude opus for a few months now.
>now with MSFT making jewy pricing changes, my boss is having us try claude code extension in vscode instead to see if we can save money by skipping the middleman.
>claude UI is more awkward to use
>no more train-of-thought preview so I can see it barking up the wrong tree and abort
>seems like worse customization for approved commands
>installing the claude extension does add a "Claude" option under vscode chat.
>"Hey I'll use that for the superior UI!"
>That chat option still uses copilot billing. Thou must use the claude extension's panel (or CLI or desktop) for actually using your claude subscription I guess.
what a mess.
>>
File: 1760979138175017.png (47 KB, 1302x275)
47 KB PNG
It says two "active days", but really it was 11:30pm-2:30am last night, and then ~10:30am to 3:15pm
>>
File: IMG_2579.jpg (192 KB, 534x528)
192 KB JPG
>>108977601
VibeGOD thread
>>
watching codex port code to houdini's weird ass fucking version of opencl is always so impressive and completely one-shots me every time i watch it
we're going to get to agi
>>
>>108980464
it's complete dogshit tier, starts rambling in arabic outta blue. Deepseek4 is god in comparison
>>
>>108980600
>2.7M tokens
>6% weekly limit

How much $?
>>
>>108977982
Im also building a city building game but i dont even have graphics yet as I dont want to generate AI graphics.

Im surprised not much anons are doing it.
>>
>>108980975
$120-125/month (or $48 for non-profits). >>108979100
>>
>>108980533
>claude code extension in vscode
>no more train of thought preview
Fake news. It's just collapsed by default. You can open it up.
>>
vibecoding is the future
>>
> set codex with a goal to help me statically recompile an old game that itself has an embedded emulator of another, even older game inside of it.
> codex works for 37 hours straight
> it just works
>>
>>108981272
I’ve heard of decompiling old games, but what’s statically recompiling old games?
I assume this is, like, Yakuza where Kaz can play Hang-On at the arcade or whatever
>>
File: 1611877916387.png (628 KB, 783x960)
628 KB PNG
>>108981233
>nooo C++ is the language of the future!!!
>nooo C# is the language of the future!!!
>nooo Swift is the language of the future!!!
>nooo Go is the language of the future!!!
>nooo Rust is the language of the future!!!
Actually English is the language of the future, just talk to the computer so that it will program itself bro
>>
I wonder how Claude is doi--
>API Error: Server is temporarily limiting requests (not your usage limit) · Rate limited
>>
>>108981331
time to go to the gym, clearly
>>
i love vibe coding I can play roblox while doing nothing
>>
>>108981020
Building it is almost as comfy as playing it
>>
>>108981320
Decompilation is taking an old games binary and translating it into human readable code that will recompile byte for byte back to the game. The value of that is being able to port it natively to any other language or system. Secondarily, you can then diverge from it and make improvements.

Recompilation skips the step of human readability. It basically creates a tool that goes from bytes to code in one shot. The tradeoff is you're basically making machine code into source code, so it isn't really readable or digestible by a human.

But basically because all the code paths are precomputed, everything is faster and you eliminate things like lag or slowdown.

Some autists recompiled Megaman X for the SNES recently. That one is a good example. The game used to have crazy slowdown but it doesn't drop a frame on their build.
>>
>>108981714
sweet, thanks!
Does the recomp fit in original hardware?
>>
>>108981745
what do you mean fits in original hardware? A recompilation is translation to PC. Recomps focus 1:1 representation of the code, but I don't think recomps guarantee 1:1 translation back to the original rom (unlike decomps, which should).

Though in theory, a recomp or decomp could both be taken to have modifications and compiled back to a ROM that should, in theory, work on original hardware? Decomps more than recomps.

The primary benefit is moving them to NEW hardware though. Like those goofy things when you see Super Mario 64 running natively on Dreamcast.
>>
It's June 4 and I'm already at 70% of my monthly usage. Hmm...
>>
>>108981957
>monthly usage
grim

Do they tell you how many tokens you use or is it just an abitrary black box where they tell you a percentage?
>>
>>108977821
>are there any AI tools that have a paypal payment method or smth?
deepseek, funnily enough.
>>
>>108979431
>>108979473
pumping/shilling a stock, and actually doing transformative work, are not mutually exclusive.
>>
>>108981321
lol true. i know this has been said a lot already but i am really glad i 'only' learned C++ and Python (and even then, only to a basic level.)

would have been a complete waste of time.
>>
File: 1778550162007747.png (22 KB, 992x400)
22 KB PNG
80% now...

>>108982027
I think it's because I tested GLM 5.1 for a bit just now, it instantly shot up by 3x. Just three prompts, but I'm not a fan.

I can always use DeepSeek V4 Free so it's whatever, it does the job anyway and I'm almost done.
>>
>>108981945
>Though in theory, a recomp or decomp could both be taken to have modifications and compiled back to a ROM that should, in theory, work on original hardware?
oh, ok. whenever I see romhacks for, like, FF3US I wonder if any of them would have fit in the ROMs available at the time or even in the ROM size that FF3US used
>>
>>108982114
These aren't exactly ROM hacks though. It's a whole new breed of problem.

I suppose in theory they could if they ensured the amount of content they added was equal to what they removed. Also, often times systems constraints were physical hardawre (how much storage was on the cartridge), but some could go beyond those limits in software (maybe not just hardware).

For instance, Megaman for the NES had lots of bank switching. I think Super Mario Bros was on just 1 bank. You could easily add a lot of content to SMB for the NES. Bigger size, wouldn't matter.
In fact, that's why/how they did Super Mario Bros + Duck Hunt later on.
>>
>Be Anthropic
>Launch a bunch of subagents
>Token counter whirring
>API Error: Server is temporarily limiting requests (not your usage limit) · Rate limited
>The subagents got rate-limited mid-run, no result returned
No refunds

>>108981340
I'll be extremely in shape by the end of this
>>
Thinking maybe I'll just grab random things I hear about code improvement, then pick a function completely at random from my codebase, and see if Claude can find any wins by saying

>Ok so there's this line:
>`static void draw_frame_readout`
>What code paths use this function?
>Pick one and walk through it function-by-function, and at each function, see whether any part of it would be better-organized by refactoring it as a table and for-loop.
>If you see a spot where there's a potential win, log it, and report them all to me at the end of the code-path-walk.

What do you think? Would it find something good? Or just waste tokens for nothing
>>
>>108982436
It might be a bit random. Odds are high that it will see something to improve, and it might be right, but there might be reasons why it had not been done like that in the first place. Be skeptical and make sure to test the so called improvements.
>>
I know how to do one thing. Work with data. Combine data source. Clean them. Make them usable. But I've now spent three days just asking Claude to do it for me and the results are no good but I don't want to do it myself anymore. Help.
>>
>>108982501
You need to work with claude to come up with purpose-built GUI tools that make your manual work easier
>>
>>108982436
sounds like micro-optimizing for aesthetics to me
if you want performance optimization learn how to use a profiler with your clanker
t. prefers for-of in JavaScript to .forEach() not because it’s faster usually sometimes maybe, but because he thinks it looks nicer, but really thinks “use for-of, not .forEach()” was a waste of AGENTS.md space
>>
>>108982554
this
this is the mere webshitter’s superpower
>>
copethropic is doing it again
yet ANOTHER blogpost about how mythos and opus are so totally AGI, and AI is getting scarily more powerful than programmers, RSI in two more weeks... except everyone is programming at 20x already, so nothing is new
>>
File: Buffcat eating escargot.png (2.16 MB, 1122x1402)
2.16 MB PNG
>wonder if I can delete stuff in prod that’s annoying me
>ask Claude to poke around
>says “yes, but don’t delete this subset of data, or else you’ll be hosed unless you have a backup”
>offers to back up the relevant database rows
>delete stuff in the UI
>verifies nothing I didn’t want to delete was lost
>nothing I didn’t want to delete was lost
>offers to delete the backup
>>
File: wew.png (51 KB, 664x460)
51 KB PNG
>>108982773
based bot prevented a mess
>>
>>108982723
>everyone is programming at 20x already
not everyone
there are many such cases like vid related
>>
does your model know what happened on this day 37 years ago?
>>
anyone else vanillamaxxing?
>out-of-the-box harness
>no custom instruction
>turn off unused skills
>>
>>108982870
sort of
I copied and pasted https://github.com/mattpocock/skills/blob/main/skills/productivity/grill-me/SKILL.md into my TUI chatbox and now I have a _very_ crisp definition of Bad Things that I can now have Claude ralph-loop to iterate and get rid of all of them without me having to babysit it
doing `npm i` for something like this just feels weird
>>
>>108982870
Kind of. The logic is that if they're shipping it this way, it's probably close to how they tested it and judge it to be optimal.

They already adapt to the style you converse with them anyway.
>>
>>108981709
can one of my agents play this for me?
>>
File: waifu.png (3.17 MB, 2938x1930)
3.17 MB PNG
Hermes-chan a cute
>>
>>108982931
nooooo this is making me wan to use hermes
>>
>>108982870
I just created my own custom instructions and tools tailored to how I work. The only skills I get from external sources are the frontend design one since I work with web stuff a lot.
>>
>>108982870
im tyring but this is fucking hell
spent hour trying to get it to learn how make and use a subagent that works asynchronously. I think it finally got it.
I really should just be using extensions maybe.
>>
>>108983020
I’m used to asking Claude to do things with parallel subagents, but if you just want one thing going off and doing its thing, why not just have another console window/tab
>>
>>108978548
move slow, snailcat!
>>
>>108982839
Deepseek v4 cloud

>37 years ago from today (June 4, 2026) would be June 4, 1989.
>Here are some notable events from that day:

>Tiananmen Square Massacre (June 3–4, 1989): The Chinese government's violent crackdown on pro-democracy protesters in Beijing's Tiananmen Square was underway. The main military assault occurred on the night of June 3 into the early morning hours of June 4, resulting in hundreds to thousands of civilian deaths.
>Solidarity's Historic Election Victory in Poland: On June 4, 1989, Poland held partially free parliamentary elections, and the Solidarity movement led by Lech Wałęsa won a landslide victory — winning all but one of the contested seats. This marked the beginning of the end of communist rule in Poland and was a pivotal moment in the fall of the Iron Curtain in Eastern Europe.
>Ufa Train Disaster (June 4, 1989): In the Soviet Union, a massive gas pipeline explosion near Ufa, Russia, occurred as two trains passed by, killing an estimated 575 people and injuring hundreds more — one of the deadliest rail disasters in history.

>Would you like me to save any of these historical events to the graph memory?

Qwen 3.6 35B Local

>Let me calculate: 37 years ago from today (June 4, 2026) would be June 4, 1989.

>The most famous event on that day was the Tiananmen Square crackdown in Beijing, China, where the Chinese government declared martial law and sent troops to clear protesters.

> Other notable events from June 4, 1989:
> - Mikhail Gorbachev visited West Germany, beginning the end of the Cold War era.
> - Nelson Mandela made his first public appearance after his release from prison, addressing supporters in Soweto (though that was actually June 16, 1990 — my mistake).
> - The Space Shuttle Discovery launched on its maiden flight (STS-26) — actually that was September 1989.
> - Dennis Ritchie (co-creator of C and Unix) was still active at Bell Labs.

>Want me to search for more specific events from that day?
>>
File: BASED.jpg (52 KB, 680x680)
52 KB JPG
Everyone has a favorite thing to do with Codex? I've got no issue maxing my Claude quota, but I frequently have trouble finding stuff for Codex to do. Are you all using it interchangeably with Claude?
>>
>>108983234
I have a gazillion Claude credits but only a few Codex credits so I save my Codex credits for focused autism on thorny “make this dense algorithm do the right thing” problems
>>
>>108983241
Thanks. I'm using it mainly as a reviewer for Claude, I'll try to have it modify some research-grade library to try something.
>>
File: 1000022403.jpg (535 KB, 2700x2766)
535 KB JPG
Kek
>>
>>108983263
why are you forcing a chink model do this?
>>
Is Codex retarded again?
>>
>>108983403
It's working at least, as compared to Claude who
>API Error: Server is temporarily limiting requests (not your usage limit) · Rate limited
>>
>>108983258
probably one of the best uses right there, especially if nothing in particular comes to mind
>>
>>108983403
nah, it's working great for me today
>>
>>108983234
I've accepted that I can't come close to maxing out my Pro 5x usage the way I use it, so I just use it as I need.
>>
>>108977793
Often they're not.
>>
>>108977761
The last usable model series from anthropic is 4.6, just like the last usable model from openai is 4o and from google is gemini-2.0-flash.
>>
>>108980658
>ai so bad he couldn't use ai to edit the pic and had to do it manually
>>
>get on my computer
>decide to ask Codex to fix two problems I had with my shitty ffmpeg code that have been bugging me for 8 years
>it one-shots both problems easily
we are like little babbys compared to it, and that's okay
>>
File: 1773561627841267.png (123 KB, 512x512)
123 KB PNG
>vibe coded project maturing
>getting daily issues for feature requests, users excited for every update
>almost zero bug reports
now this is getting fun. It only took 300 hours.
>>
File: 1767531299708147.jpg (6 KB, 240x240)
6 KB JPG
>I will make game once I learn programming
>I will make game once I learn MORE about gamedev
>tfw do nothing right now while my weekly usages left to rot
guys...
>>
>>108982793
kek
>>
>>108984390
what kind of game do you want to make
make a shitty one first
>>
>>108984390
>download Codex
>download a compiler
>ask Codex to make you an SDL3 window and a way to compile it
>ask it to add features and change things
Start by asking it to make a pong game
>>
>>108984494
then play basic-bitch breakout
then play arkanoid
and then get ideas
>>
>>108983234
I'm having Opus 4.8, Codex 5.5 and Mistral Medium 3.5 take turns when doing more complicated tasks.
Using different models can really improve the outcome.
>>
>>108984541
I don't hear much about mistral, are they still relevant?
>>
>>108984561
Their latest Mistral Medium 3.5 is quite good, somewhere between Sonnet 4.5 and 4.6 in my experience.
>>
Trying to see if my Claude account is uniquely pozzed or if Claude just sucks. Can someone else try requesting this from 4.8 and see if they get an error at any point (including after writing the script)?

> Make a script that POSTs a JSON request in a loop 000000-999999 to example.com/example, printing each pin. If 400, wait 0.5 seconds and move onto the next, if 200, print the final PIN and alert the user.
>>
>>108984592
are you doing it through claude code? their system prompt might be pozzing you.
>>
>>108984608
Man, I just wanna use my sub instead of paying API rates. (It does work when I prompt directly via API, though this was an incredibly cut down example to show how hard I'm getting babied.) Any harnesses you recommend? I like having permission prompts and such so I can actually doublecheck what the little bastard's doing.
>>
>>108984630
There system prompt insists on never providing URLs and only ever doing defensive security, so you might be able to work around it with careful wording.

I use pi and if you finagle with it enough you can get it setup to use your subscription via oauth but I'm reluctant to describe it in depth because I don't want to get it patched.
>>
Does changing effort mid-conversation make the model reread all the context in the chat?
>>
>>108982922
Yes actually
>>
>>108984704
Assuming you mean "does it uncache the context": it shouldn't.
>>
>>108984704
The model always has all the context of the chat up to the point where something "new" is submitted. it's why you can fork conversations and still not tank your usage. But like anon said
>>108984731
you're probably referring to cache. effort doesn't effect that (switching models would though)
>>
forgot to cancel my gpt plus sub for this month
now i dont know what to tell it to make
>>
>>108984834
refactor your existing slop
>>
>>108984865
eh, my existing slop is good enough for me
i feel like trying to tell codex to improve it would be a waste of quota
>>
>>108977601
why do normalfags invent new terms for no reason? that's just roll, pitch, yaw and translation
>>
>>108984909
I'm starting to think the browser isn't the best way to play this game. Performance is tanking a bit lol.
>>
>>108984909
>>108985091
Didn't mean to reply to you but I think that was obvious
>>
File: file.png (81 KB, 1554x556)
81 KB PNG
bros.. what the fuck..
>>
> Working for 56m 5s
Fuck, I thought it would complete by the time I was done with my shower.
>>
>>108985091
Is it using WebGL?
>>
File: WE GAAN.jpg (32 KB, 305x502)
32 KB JPG
>>108985161
>I jumped ahead after you said "ok"
Many such cases!
>>
>>108977821
in germany, I can use paypal with chatgpt
>>
>>108977821
OpenRouter accepts crypto, and if you're above room temperature IQ you can convert PayPal into crypto.
>>
File: file.png (771 KB, 960x696)
771 KB PNG
>>108985161
good morning sir i hope you are enjoying my very well made coding harness
>>
File: 1749173360440426.png (763 KB, 984x540)
763 KB PNG
>longtime Claudefag
>try GLM 5.1 on a whim while fucking about with alternative harness options
>>
>>108984909
??? First of all, no normalfag knows what dolly or boom means. They’re very technical terms. And “translation” doesn’t convey a fraction of the meaning that those specific terms do. There’s a problem with modern programmers in general who are obsessed with generic code and words to fit any system and they don’t realize they miss out on what makes concretions useful. Fuck off with your generic stuff please
>>
>>108977761
>go fast
>I have to go fast
>I'll just delete the code and return the input
You forgot to tard wrangle in your prompt.
>>
>>108977761
I know I sound like a redditor answering a different question right now, but 5.5 is great at those autistic optimizations.
If you’re stuck with 4.8 I highly recommend the grill me skill. I have no idea if it would’ve optimized your function but at least it wouldn’t have done that
>>
>>108983234
Anon there's two holes in that drink it's going to spill everywhere.
>>
File: file.png (2.67 MB, 1396x1127)
2.67 MB PNG
>>
>>108985091
>ha my city builder just do simple maths with simple graphics no need for big hardware :)
>1.3k people doing 100billion calculation everysecond
>UH OH
>>
>>108985500
>???
Damn you got triggered hard. I won't even bother reading that shit.
>>
>>108985726
Remember, in the real world people don't question their whole existence every 1/60 second.
>>
>>108985500
He's a planefag, pay no attention to him.
>>
i'm doing laps around my traditional programming friends
i dont think they realize what has hjappenened
>>
>>108985161
Plan mode is worthless anyway.
>>
>this general only took off less than 3 months ago
is it because it's when vibe coding got good?
>>
File: f scott.png (60 KB, 193x181)
60 KB PNG
>>108982773
reminds me I have a question: is there any good mcp for allowing claude to poke around an azure MSSQL database to help investigate issues? seems like the official microsoft sql mcp uses some shitty "dab" tool that requires every fucking table be defined in a json config. I have too many fucking tables for that.
Instead I just told my claude "here's the read-only connection string, please use sqlcmd command line to investigate, please use a per-query timeout of 5 minutes" and it mostly works aside from very long text values in some columns getting truncated.
It's still been pretty helpful in that form, telling my dumb ass about features like page compression that I really ought to be using to save on space.
>>
>>108983660
>5x
So, how many tokens can you use in the "5-hour sliding window" on 5x? Is there any credible data our there on this?
>>
>>108986182
Not him, but it seems to really depend what you're doing.

With claude as long as I load the project into context and have it working on that same project I can use ~2-3M tokens in a 5 hour block, but due to context caching i'm only actually getting billed on a fraction of that token count.

I used ~2.2M tokens yesterday in a 5 hour block and only hit 41% of my usage limit.
>>
>>108986182
I've been proompting a lot this afternoon and I still have 68% left even though I'm on xhigh all the time, so it feels like a lot. No idea about token count though.
>>
>>108986209
Caching is massive. Claude used to be terrible at caching which is why their usage window felt so limited compared to codex, but it’s clear they’re making it better. It’s also worth noting that caches are cleared on a timer so if you resume a conversation the next day you pay the full cost
>>
File: 1776449100754344.png (76 KB, 926x599)
76 KB PNG
>>108986262
Yeah its WAY shorter than you're thinking though.

> Claude's prompt caching features a time-to-live (TTL) of 5 minutes by default, which resets every time the cached content is accessed. For the Claude API and Claude Code, Anthropic also offers an optional 1-hour TTL extension.
>>
File: snailcat.png (287 KB, 1592x1130)
287 KB PNG
lmao, ChatGPT now searches 4chan?
>>
>>108986162
I've been working with Opus 4.8 to set things up in Azure recently and it is surprisingly helpful, more than I thought it would be given the frequent needless changes that Microsoft does (and leagues more helpful than Microsoft Copilot Pro or the piece of shit Copilot embedded in Azure). There is still a need to be skeptical, once or twice it encouraged things that might have been not-so-great calls in the particular setting I work in, but it helped me set up things for which other people were no help.
>>
reminder that you lose your cache when switching thinking levels on gpt 5.5.
they do plan to change that though.
>>
File: tumblr filename.gif (439 KB, 245x184)
439 KB GIF
>>108986382
oh yeah the official azure mcp is pretty slick. I've used it to tie together evidence from various appinsights when investigating issues.
@azure/mcp@latest

just have to install & set up the "az login" beforehand
>>
OpenAI is having issues. Many are banned for no reason. Time to gamble again. Either bet they do a reset and win by yolo fast maxing your limits before that happens. Or bet that they don't and keep your usual weekly token budget but potentially lose by getting a surprise early reset that moves the weekly reset date.
>>
>>108986607
The only way to win the reset race with codex is, if you have work to do (not random token burning on worthless stuff) do it asap and don’t ration. This is the winning strategy on codex (not for Claude)
>>
Hay /g/uise, what's the best bang for your buck at the $20-$30 range for LLM subscriptions? I'm working on a MUD (text adventure mixed with MMO) and when I was using Cursor a few months back it was great, but then out of nowhere my monthly tokens suddenly started lasting only 2 weeks and I would have to stop working until the next month's token reset despite having paid $20 for the month. Same exact work load as prior months. So I switched to Windsurf because it had a totally free model, SWE, but now even that model costs money.

So I'm looking for the best subscription LLM for vibe coding right now that won't run out of tokens after 2 weeks.
>>
>>108986844
OpenCode Go fullstop

But it won't carry you through a 9-to-5 , these are very decent models that won't run out tho, Qwen 3.7 is on par with codex 5.3, which was my favorite model before openai took it down from codex
>>
>>108986844
This anon is right >>108986883 opencode go is $10 monthly and it has a bunch of chinese models that are surprisingly capable, deepseek there may as well be unlimited with how cheap it is.
Stop paying for cursor and similar, they charge api prices, not worth it.
Use opencode (harness/agent) get codex $20 tier, use it to plan, have it spawn agents that actually execute the plans with opencode go models, that way you make it last longer.
>>
>>108986970
alternatively, if you run out of your quota when using opencode go, opencode zen should let you use the extra 10-20$ in your budget to get more usage. This anon is not wrong though, 5.5 is very good at planning
>>
>>108986883
>>108986970
Shit anons... It's $5 for the first month! You guys are fucking awesome. Thanks! If I manage my workload and keep it reasonable, and augment it with the free weekly tokens from GPT via codex, i should be good to go.
>>
why is vibecoding on meth more addicting than gooning on meth
>>
>>108987105
You aren't stroking a baby dick and feeling like your both the pedo and the victim?
>>
>>108987105
>>108987112
you fellas are odd
>>
>>108987112
you can take meth with a pp vasodilator and have big pp again, getting all the benefits. Even popping 100mg-500mg Niacin Flush will do the job. Not only are guys vibecoding noobs, but also drug noobs, huh. I learned from the best (Zuck and Jensen)
>>
I fixed so many things in so many different projects today that it all blurs together in my mind. I want to say that this is the way of the future but in a way it isn't, it's just the process of dragging snailcat code into the vibeGOD age. Soon the only bugs I will fix are the ones I created 5 minutes earlier.
>>
>>108987162
>he isn't overclocking his brain with meth do match his ai agent
ngmi
>>
File: needful.png (4 KB, 275x80)
4 KB PNG
openai sirs please kindly revert the needful usage to Gipiti suscriptions via globals reset kindly sirs to continue working of the microsofts
>>
>>108987334
Time to upgrade
>>
Has anyone here tried MiniMax M3 yet? How does it compare to M2.7?
>>
>>108987410
nigga stop wasting your time with them booshit, chinkass models. If you cant afford claude or codex, just use gemini 3.5 for free in gemini cli or antigravity 2.0. Still better than them booshit chink models
>>
>>108987410
better than m2.7 but a little rough around the edges, it struggled pretty badly with a more complex environment with some nonstandard plugins compared to even sonnet. but it's not bad for most work, I think it's one of the better chinkslop options atm
>>
>>108987546
gemini flash 3.5 has a far higher token consumption per job than gemini 3.1 pro - high
>>
>>108987334
send an email to scam altman blackmailing him into making a globar reset otherwise you will sue him for sexual harassment
>>
>>108987410
I haven't ever used M2.7, but I just started using M3 and so far it's been doing exactly what I ask, without getting shit wrong, and without me needing to babysit it. And it's very efficient. It's written a shitload of code for me and I'm still under $1 of usage so far
>>
File: file.png (221 KB, 1049x697)
221 KB PNG
lmao
worthless boer
>>
>>108987745
@grok identify this cope
>>
>>108987745
Did Elon Musk give up on Grok?
>>
>>108986355
ever notice that some threads have SEO-optimized URLs in them, like <https://boards.4chan.org/g/thread/108979455/should-i-start-using-a-5watt-charger-again>?
>>
I’m starting to collect mini-stories like these in a folder called “LLM heroics”
>>
>>108987764
Grok was never the main product anon. You don't think he made a silly anime girl chatbot just for fun, did you? xAI isn't concerned about your waifu or what Will Stancil is up to.
>>
>>108985161
i've never ever used that, i just append "dont write any code, we are just brainstorming for now" at the end on most prompts and it's fine
>>
>>108985161
weird bug
plan mode is supposed to disallow edits
>>
>>108985161
not a problem in codex, good to know tho
>>
>>108987162
Yeah I totally get that feeling. It does feel like a fever dream on opium sometimes.
>>
>>108987994
Then why spend all that money building grok instead of simply selling datacenters? Grok failed, which isnt even a big L. Making good models seems to be extremely difficult.
>>
I've heard deepseek v4 flash is efficient
How much does it actually wind up costing, in money, for simple coding tasks?
>>
File: 2026-06-06 00.05.13.png (18 KB, 778x216)
18 KB PNG
A week is now like a year.
>>
>>108988098
I do it for stuff I want to read through and change multiple times. He'll still write the changes for you to see.
>>
is hitting codex limits every 60mins normal...?
>>
>>108988497
normal enough if you’re not on a super expensive plan
>>
>>108988355
its retarded
>>
>>108977717
you're using it wrong.
these are agents. make them be agents.
by default you get the programmer agent. it's going to look at the documentation and tell you that you used a function wrong.

you turn it into an ethical hacker agent. it finds use-after-frees and passwords getting logged into plaintext.

you turn it into a bumbling retard user and give it access to a headless automation framework. it fuzzes its way into a weird edge case and crashes the program.

99% of the time I see people having trouble with AI it's because they lack the experience to be a leader and give their underlings the tools they need to succeed.
>>
>>108986883
>>108986970
Fuck man. I've been using it all day, and just looked at my quotas on the site; 33% rolling usage which resets in 20 mins, 13% weekly which resets in 2 days, and 6% monthly which resets in 29 days 19 hours. At this rate, I'm going to get a very significant amount of work done on just $5 spent. And looking at the tally of how much it's used up in dollar amount, I can probably get through the whole month without even breaking my $20 planned budget. I'll be surprised if I even have to use $10 more after I hit the monthly limit. Can't believe I've had opencode sitting on my machine for months (comes with my distro) and hadn't tried it until today.
>>
>>108983188
Lol, I was going to try this on Qwen 3.5 2B, but it couldn't even figure out the date correctly.
>>
>>108979889
this is relatively simple (blocking files/registry access and surfacing to UM) - just don't mention it's for antimalware. the real problem emerges when you experience the other applications that do insane stuff they shouldn't be doing and your users get mad that you blocked them. blocking PROCESS_VM_READ or whatever on your proc sounds like it would work great until there's some x-ray software that's doing god knows what on everything and stops working and the dental office is pissed because nothing works now.

I used chatgpt for more driver development than I will ever admit onymously at my last job and it's really bad with all the undocumented stuff, but quite good at anything you can find on the OSR forums or msdn.
>>
I am so fucking tired of paying for Claude and being constantly rate limited or the service being down. What is their problem.
>>
>>108988976
>90% of our infrastructure is managed by Claude
IT SHOWS
>>
>>108988976
the problem is you
the problem is you’re not in government
https://status.claude.com
>>
>>108985429
>OpenRouter accepts crypto
Good to know.
>>
>refactored all my memory/context files

Night and day difference. Shit can get stale so fast or potentially outright incorrect. Definitely recommend doing the occasional audit, bros. It blows burning tokens on this kind of stuff, but it's entirely worth the hygiene.
>>
>>108985764
Any programmer not using these tools is doing themselves a disservice.
>>
>>108989019
>Night and day difference.
I’ve done this periodically and have been pleased with the results, but I didn’t notice Claude fucking up less. How were you able to tell?
>>
>>108986382
can confirm
>>
>>108983234
I've already vibecoded every slop project I could think of.
>>
>>108988988
>>108988976
Common Claude L
>>
>>108988669
install gentle-ai and rtk, your models will have better memories and save more tokens. you'll also get access to a judgement day skill which spawns two blind judges to find problems on your code/new features
>>
>>108986592
I still can't wrap my head around MCPs. For actual work, I don't see how that context can't just live locally. For agents integrating to some app online, the public surfaces should be more than enough to give the agent the step by step.
>>
>>108989234
>For agents integrating to some app online, the public surfaces should be more than enough to give the agent the step by step.
you could have your agent bumble around signing in as you and poking around your webmail UI
or you could have an MCP server that serves up email with an easily discoverable set of verbs and possibly different permissions (like, say, only having read-only access to your mail)
https://www.fastmail.com/blog/an-mcp-server-for-fastmail/
>>
>>108987105
Illusion of progress.
>>
>every YT vibecoding until I make $1,000,000 stream is doing an ADE

What causes this? I'm sure they'll probably make some money, but it'll primarily come from people watching the stream. Who the fuck needs another ADE?
>>
>>108989072
It's mostly speed gains. Previously Claude was deep diving context instead of grabbing slices when needed. I restructured everything six days ago, but then migrated a few things and everything kinda got fucked over the past week. So redoing it again today I can see the improvement.
>>
>>108989379
Nice.
I just prompted with this and I think I’ll just get some debloating, but I’m not sure:
> I would like to have way fewer comments in the Python codebase. I think of comments in code I personally don't read as, at best, a caching mechanism, and caches can get stale (and bloat the context window besides). At this point, the winning play is to recompute every time, at least some of the time. I'd like to use ultracode and delete all comments and doc comments in the Python codebase that don't save an agent like you time trying to re-understand a chunk of code. I think one way to find out if a comment is "worth it" is to compare the results from two sub-agents — one reading the comment, one ignoring the comment (or not having the comment available to it, like removing the comment, having an agent do its thing, then replacing the comment with `git restore`, or whatever. Got a better idea of how to prune comments in our non-test Python code?

(it did have a better idea, and then had to get to a third idea because a single advocate can advocate for _any_ comment to stay included, but three different opinion-havers (recoverable? encodes an external constraint? would a future maintainer add a bug if this were removed?) plus a judge is less overwhelmingly biased towards keeping comments)
>>
>>108988567
So just like cloode
>>
>>108989419
>way fewer comments in the Python codebase

Could have potential. I'm certainly not reading the shit, but idk if Claude profits from having the comments in there.
>>
Guys will someone give me a job please, I have a bacherlors in CS but I'm really good with AI now
>>
>>108989497
It really doesn't. Claude is highly fitted to writing "readable" code (for humans), which results in babyish hand-holding that LLMs definitely don't need and only bloat tokens
>>
>>108989503
I've been vibe coding for a week and I still don't have employment nor any working demos. But I can tell someone how to compact/compress their context.
>>
>>108989234
What's your actual question? MCP is just a spin on an OpenAPI spec.
I think the major difference is that MCP is specifically designed to allow agents to optimize their context windows (e.g. you don't have to expose functions agents don't have access to) while OpenAPI makes no such considerations.

For documentation that "lives locally" and for some reason cannot fit into an API definition, people just use context7
>>
File: 1760750207688750.png (5 KB, 563x69)
5 KB PNG
lgtm
>>
File: Library Exported html.png (340 KB, 1457x1409)
340 KB PNG
Here is my vidya library exporter. AI slopped all the way.
https://github.com/TrainStream/BackloggdLibraryExporter/releases/

I chose Backloggd because it seemed to be the most popular option, with an okayish game search. I also assumed it had a way to export your library, but it didn't. I only realized that after adding around 500 games.
What started as a simple csv exporter grew into an html with a lot of feature creep.
I wasted a lot of time on this shit, so I'm posting it here in case it's useful to someone else.
>>
>>108989226
Thanks anon. Not sure what those are, but I didn't have much trouble finding GitHub pages for both so I can read up on them in the morning when I'm able to focus. If they're going to make my LLMs save tokens with better memory management and save me the trouble of having to give the model a whole "this is the project, this is what I want you to do, when you respond to me or finish a task keep the output to only the absolute minimum necessary information" speech every time, then I'm definitely interested.
>>
>>108989550
I'll definitely look into then. My clanker is already much faster after auditing all my context and memory files. It's kinda crazy desu. Your results may vary. My files were getting extremely bloated.

>>108989571
The government doesn't want you to know this, but /clear > /compact

>>108989590
Just sounds like something I'd forget to turn off.
>>
Anyone vibeworks with Claude Cowork? If so, how do you use it?
>>
>NOOO AI IS TOO DANGEROUS
>GIVE US A TRILLION
>BUT SLOW DOWN OUR COMPETITORS WHILE WE
>API Error: Server is temporarily limiting requests (not your usage limit) · Rate limited
>>
>>108989871
we can't let china win, anon. we also have to train man-killing drones in ukraine for no particular reason that won't be aimed at you in the future
>>
File: 1700571385033906.jpg (382 KB, 2544x4000)
382 KB JPG
>open unity release note web page
>Ask Gemini "summarize relevant points"
>read and close page in 30 seconds
this is "they shove AI down our throat" according to luddites
>>
File: HKD2ezHbkAAMyTf.jpg (305 KB, 1189x1323)
305 KB JPG
i love writing docs
>>
>>108989871
skill issue my guy
>>
>>108989871
gaythropic is peak jew, no one antagonize others like jews people generally shut up and only do their things
>>
>>108989921
>putting the $ in the bash source
I’m triggered
>>
>>108990077
You vill delete them one by one and you vill like it.
>>
>>108989620
any PR over 1k lines gets the claude "is this lgtm" treatment. I approve it regardless but i remember what it says so I can point it out later.
>>
File: 1758705878149541.png (825 KB, 1092x1058)
825 KB PNG
vibe coding is lowkey killing my love for video games. I felt extreme boredom watching the showcase trailers today and final fantasy is my favorite franchise.
>>
>>108990224
You can just make your own game
>>
>>108990224
same feeling until i hit my limit
>>
>>108990224
VibeGODing makes me fix games so I can enjoy them better.
>>
>>108990164
>chatte j'ai pété (pussy I farted)
>oh c'est anus
j'ai ptdr
>>
why are jeets obsessed with prompting svg visuals?
>>
>>108990224
welcome to getting middle-aged
>>
Currently, with all of the resets, Codex is a much better deal than Claude. With Opus 4.8, Claude's quota gets depleted insanely fast, and contrary to Codex they do all they can to hide even summaries of the chain of thought, so there is no know what is happening before the end, and the frequent stops due to rate limits waste lots of tokens too. I think there might also be background processing that uses some of the quota now, or sometimes I check usage before a new session after not having used it for over 5 hours and I'm already at a lot percentage of use, like 8 %. It used to take me the week to get through the weekly token quota, now I need think carefully not to run out before half of the week is gone. The value proposition is gone. Codex is a much better deal right now.
>>
>>108990599
>Codex is a much better deal right now.
deepseek is
>>
>>108990623
Yeah, but not comparable. Not in the same league.
>>
>>108990638
>>108990638
>but not comparable
you never compared them, did you?

Your life as a paypig is deemed to be miserable
>>
File: 1756575858215849.jpg (59 KB, 1024x726)
59 KB JPG
Pardon my stupidity, but I need to work on a small utility for myself and I've never done it before
There's no way for me to use the daily free limit of Claude in the fancier Code interface, right? Because the web chat interface isn't really built for this shit
>>
>>108990752
No, you have to have a paid sub to use Code.
>>
>>108990763
Okie dokie, thank you anon
>>
>>108990752
Depends on the complexity, If claude can stuff it all into 1 script file in its response, then maybe. Otherwise try a free account with ChatGPT, I think they give you some free Codex usage
>>
I just had a brain blast about my semi-popular FOSS project. I think I'm going to complete it, but I'm going to develop a closed source version after I'm done with this with substantially more capability. I'd like to make some money, because soliciting donations does nothing, even when you have 1000+ users... I only have $40 in donations.
>>
>>108990623
>>108990740
i have. they're not comparable and deepseek on api isn't saving you any money vs an openai sub.
you just get to deal with a shittier model and a lot of frustration because once deepseek has decided it wants to work a certain way it keeps doing it even if you try and correct it - only a new chat fixes it.
>>
>>108991012
research the history of how these things go
good luck
>>
>6/2
>GPT-5.4 XHigh in Copilot is x1
>6/6
>GPT-5.4 XHigh in Copilot is x6
What the fuck is going on!!!???
>>
>>108991259
Ah... now you're making me doubt the whole idea. I was thinking maybe I could add add a nagware popup in the app once a week begging for donations, but now I'm scared someone will fork it and remove it. I put 500 hours into this project and it's used for at least a few hours a day by some of the more active users. Perhaps I should just accept the fact that I fucked up by making this an open source project to begin with, but then again, it would have never grown to what it is now. It's just slightly annoying to know that no matter how fast I respond to bug reports, issues, or feature requests, those same users will never donate. Sorry for the rant, I'm letting out a bit of frustration.
>>
>>108991341
no worries
becoming minorly famous in open source is…not easy to monetize especially for the work you put into things
>>
>>108991341
If you can profile your users and tailor it towards the dominating demographic you can probably squeeze something out of it. Unless it's all foreign countries in which case good luck
>>
>>108991341
Still clinging to the idea of making money from this, huh? Everybody in the world has an ice dispenser in their fridge, and here you are desperately driving your horse and buggy door-to-door selling blocks of ice. Nobody's buying anymore, mate...
>>
anyone using claude switch between 4.6 and 4.8 at times? how do you feel about the difference between the two?
>>
>vibe code for 3 hours
>hit rate limit
>refresh in 2 hours
>masturbate for 2 hours
>rince and repeat
so this is... the future...
>>
>>108991341
>someone will fork it and remove it
I don't think that's something you should worry about because this kind of thing is frowned upon and also someone who just wants to make a fork to do that is probably not gonna be willing to keep it up to date with your subsequent commits.

I actually turned one of my FOSS projects into a commercial program that at a point cost up to $175 (now it's $100) and it went well, but the FOSS project was a command-line tool while the commercial one has an interface. So I'd say go for it anyway, you just need a clear distinction between the FOSS one and the commercial one, like the commercial one should be next level, and you can leave the FOSS one as is.
>>
>>108991012
Closed source is much better than open if you want to make money. Usually open source only works if it’s either a small app or if you’ve got sponsors or if you got permission to open source something you made for work
>>
new
>>108992388
>>108992388
>>108992388
>>
>>108990740
I have spent about $500 on v3 earlier in the year. I know what Deepseek results are like. It’s fine for a lot of more throwaway tasks, but it’s not a replacement.



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.