I’m doing pretty basic stuff, and the moment Claude has to work across 3+ files, it progressively gets worse and becomes basically useless.Do you have a personal experience where you wrote something with AI and it almost oneshotted it?
>>109586382Niggers on Twitter are oneshotting games with questionable gameplay and pretty crazy graphics for being just basic geometric shapes strapped together rather then vertices.
>>109586467Yeah but what's you personal experience?
>>109586382>the moment Claude has to work across 3+ files, it progressively gets worseSkill issueYou have to give Claude Code/Cursor/Codex more guidance than "fix my code"
>>109586382ai is very good at using command line tools and tireless trial and error loops give it a task with well defined inputs and outputs and it will read manuals, tutorials, man pages, error outputs, log messages until it gets something that worksdon't expect it to give you code that is safe, performant, easy to understand and maintain, flexible enough to handle edge cases or modify in the future, though. unless you have the ability to measure any of that and insert it in the loop
>>109586528It’s becoming very tiresome hearing people describe talking to Claude as a skillIf this is all your job is you should be permanently unemployed
>>109586382you can also feed it pseudocode, but whats the point of it then?i use mine to ping-pong ideas against, making it compare my solutions to industry standardsand as second pass to check for issues it can findi dont rely on it 100% but it is helpful.this can easily be done with the free tier of chat gpt once they see you need the nicer models (they have multiple models and settings and they try to converge on the cheapest for them that still satisfies your use and count on blueballing you to make you pay the 20usd subscription.)for easier stuff like quick summaries or general fuzzy search i use perplexity. their model is source-oriented, its basically a summary machine with fuzzy search capabilities, but despite that its pretty good at "reasoning" for what it is. probably bc it builds said "reasoning" on the fly, based on the sources it found
>>109586567You sound like you are unemployed
>>109586382Yes, context is a bitch. You're using skills, aren't you?
>>109586382Give this is a try. Make a CLI tool for uploading files on Exhentai. Choose the title, category and point to the directory with files.
>>109586590Other people not indulging your delusion is a touchy point for you, got it
>>109586588>i use mine to ping-pong ideas against, making it compare my solutions to industry standardskek. lobotomizing yourself one session at the time
>>109586382in my experience, it's best suited to a scenario where:1. you initially set up a codebase with some sort of framework, consistent code style, good practices, etc. 2. you create stubs to be implemented3. you have claude implement one stub4. you fix up the implementation to ensure it meets your standards5. repeat6. have it introspect and generate rules, skills, etc. based on what it learns about how you want code to look and function7. only at this point should you start to attempt tasking it with implementing entire features on its own8. review the code together with it9. update the code to meet your standards10. repeat, and so on
>>109586612No, Im not using skills. I maintain a log for the general structure of the code - what goes where step by step etc. it's a relatively simple code - dfs, fuzzy matching, scoring/thresholding, tree walks...Opus 5 absolutely lost the plot here.
The whole thing is a scam to get you to burn tokens.If it was actually smart it could quote you a price and deliver on that price or your money back if it doesn't work.
>>109586692Which come to think of it is using skills technically...
>>109586643>delusionThe reality of what software engineering is in 2026?You will never be a software engineer, you will always be a delusional NEET
>>109586660That's how I started initially but then I handed him some tedious and boring things like regexes and scoring and now I find bugs everywhere
>>109586764they're overeager interns (i.e., willing and able to do bitch work, but you need to either hold their hand, or ensure the tasks are extremely linear and structured)
>>109586382Not only that but there is a lot of 'dark algos/CS/implementations' that is not in the literature so no AI would be able to produce satisfactory results, let alone a state-of-the-art design.TLDR People who use AI work in low skill software domains
>>109586692Skills help you keep context at bay since only the description is ever kept inside the context. Only when they are invoked they get fully loaded.You won't be able to complete even mid scale projects by just raw dogging it.
>>109586382AI can do it, the real question that niggercattle like >>109586528 don't think to ask is whether all the work and token spending that goes into the result is actually any faster or more effective than traditional programming.At what point are you spending more time fiddling with agent loops and markdown files than just modifying the damn code yourself? AI can be very useful, but AI psychosis is real.
>>109586382yeah, it roughly one-shotted a small feature for me over many more files.I changed some details but claude's attempt wasn't bad.did you use plan mode with opus/opusplan?
>>109586567seems to be a skill if so many retards can't do it properly.
>>109587123>how the pros do it properlymayebe its just a bit shit, actually
>>109587126werks on my machine
>>109586730The delusion that this is a “skill.” If you have to intentionally deflect twice in a row it’s clear it’s a sore spot for you.
>>109587139yeah, but how, and at what price?bc not well enough for anthropic, or not cheaply enough for msft (plus the absolute state of wintra$h, but i digress...)plus plus this can be actually an euphemism for having to pay manhours to clean the mess on top of burning cash on tokens, bc msft also has ai products and wouldnt want to do anti advertisement for themseleves by saying>our ai is shit so we tried antropic's>turns out theirs is also shit, so were back to doing things the old wayimagine the absolute fucking shitstorm that would cause among the investorsfucking fecal fragmentation cluster ammunition grade lamao
>>109587166I’m not wasting my time arguing with a NEET
>>109587167 contim still laughing from the mental imagesata nutella capretbombing everyone because everyone has fingers in eachothers pocketsthe whole thing hinges on technology that doesnt even exist yetand datacenters that havent yet been builtbut most of allthe boomer, who like a thrird world kid picks up one of the misfires and it explodes covering him in shit because his 401k plummets into the shitterfragments of shit flying everywherevaporized shit sprayed onto unsuspecting bystanders who just hodl the sp500long term shit residue that causes birth defects and infertility through long term catastrophic erectile dysfunction from the trauma of just how hard the market got btfodus had its dust bowlbut the indian festival thats coming is gonna eclipse whatever that was lamao
>>109586528>Skill issueNigger what? Anons on twitter are getting ChatGPT to do original research by giving it one sentence commands lol, get this pretentious shit out of here, you prompt, pray and let Claude rip like everyone else
>the moment Claude has to work across 3+ files, it progressively gets worse and becomes basically uselesstell us more about what sort of problems you're running into so we can help you out anon
>>109587241*would be comingbecause if the whole thing unravels slowly or the techbros actually deliver, the market will be just fineit can absorb losing 7% of the s&p 500, given its distributed in time
>>109586710I don't even understand how people burn tokens. My "record" is a bit over 20 % of daily limit and even that felt like cranking it on coke brains off. The absolute slop did win us some hackathon money so there's that. Execs only care about the looks and the pitch.
>>109586807luddite cope
>>109586510When I tried to make it create a game for me that wasn’t racing game or broken barely functioning fps it failed. It’s amazing in UI design and kind of crazy when it comes to 3D modelling things in three.js, at least for LLMs, but other then that it’s kind of shit. It can do 2 days worth of making functional game settings in a single hour, but the gameplay will be horrendous and polishing it yourself will be next to impossible because you don’t understand the code and it will waste millions of tokens misunderstanding your desires. It’s basically the same as diffusion images we’re back in 2022-2025 where it was giving very impressive outputs but you had very little control, so the main people who saw significant liking in it were uncreative degenerates with no vision who just had very vague ideas or no ideas, or few autists who tinkered with it to such extent only to get slight improvement over the degens.
>>109587494Then game development will be completely solved by 2030 by your pace.
>>109586382>Do you have a personal experience where you wrote something with AI and it almost oneshotted it?yes, when a modern greenfield project is structured well and gets larger, it can oneshot stuff without much handholding. when the project is small the LLM has no reference and just makes bad assumptions and you need to be very precise and hand-holdy.using claude sonnet btw, not even the frontier stuff
>>109586467>questionable gameplay and pretty crazy graphicsgraphics are pretty easy to achieve if you got the asset rendering performance is the question. Also if gameplay is crap then the game is crap. "Oneshotting" a game like pokemon and never testing it beyond 1min demo just shows its garbage.
>>109586807you can give it a research paper and it will implement it I did it for some rf simulation stuff it works however you cannot verify the results without more expansive software so you have a script that kinda works but is hard to extend thats produces results that kinda match your expectation.
It's good at automating things and teaching you new ways to autoomate things. Boilerplate is a non-issue, it can blow right through that. So now that boilerplate is able to be written quickly, you don't need things like third party libraries as much anymore.The real hard truth: In corporate programming, in many enterprises, delusional people set lines of code as a metric to be strived for. This leads to things being made for the purpose of being made, work being made for the purpose of work. "Makework", i.e, work that serves no purpose. This is why software has become so bloated needlessly. Because of the gamification of the lines of code metric. That is the real absolute reason and the absolute truth that nobody wants to tell you. Thousands of hours of programming work has been done purely in the vain of satisfying the lines of code metric. You need less lines of code, not more.
>>109586382Massive skill issue I've worked on personal projects with 40k lines super easy even with the dumbest LLMYou need to have good documentation in optimized LLM format so they can read it quickly
>>109587734This. You need to setup the appropriate CLAUDE.md, introduce a ROADMAP file, etc. I know managing state in flat files sounds like a total backwards move but trust me, in AI this is the pattern that works. Key skills to unlock your success:>MCP fundamentals>Agents>Guardrails>Skills (skills.md format)>RAG>CoT PromptingAnd yes, you need git.
>>109587139Your job isn't to ship code, it's to advertise on 4chan. You haven't done shit.
>>109586382Lol I made that image
>>109587511>will>shall>going to>about to>set to>due to>bound to>destined to>intend>plan>aim>propose>expect>project>prepare>forecast>predict>anticipate>target>schedule
>>109587989based listanon
>>109586382>Do you have a personal experience where you wrote something with AI and it almost oneshotted it?>>109586467>>109586555>>109586588>>109586623>Niggers on Twitter are oneshotting games>ai is very good at using command line tools and tireless trial and error loops>you can also feed it pseudocode>Make a CLI tool for uploading files on ExhentaiWay to not answer OP's question at all.
>>109586623There isn't any model out there that can do this.
>>109588046^expectations<reality>jeet from bangalore > openais frontier models¯\_(ツ)_/¯
>>109588046>>109588089what i mean by that is that one-shotting anything beyond 500 locs starts to get wonkywhich is a good benchmark for the evolution of llms proper, wo taking into account the support of harnessingbut also revelatory of the underlying scaling/diminishing returns problems
>>109587719>the lines of code metricTrue, but should be generalized to any arbitrary metric of "work done" that is unrelated to any measurable value to the customer or even bottom line for the company(Jira tickets closed, changes shipped, PRs merged, reports generated, etc)
>>109586567>If this is all your job is you should be permanently unemployedThe job is to be backup of claude when the budget for tokens runs out
>>109587719>The real hard truth: In corporate programming, in many enterprises, delusional people set lines of code as a metric to be strived for. This leads to things being made for the purpose of being made, work being made for the purpose of work. "Makework", i.e, work that serves no purpose. This is why software has become so bloated needlessly. Because of the gamification of the lines of code metric. That is the real absolute reason and the absolute truth that nobody wants to tell you. Thousands of hours of programming work has been done purely in the vain of satisfying the lines of code metric. You need less lines of code, not more.Is this still a thing? Even Tokenmaxxing died down for the most part
>>109588761That type of corporate dysfunction predates LLMs by decades.
>>109586567Wait until you get your first job and discover the managerial class.
>>109588949I feel like one of the nice things about LLMs is that it will eventually make corporations drop lines of code as a metric because it can be gamed by pushing overengineered verbose slop now and setting it as a metric will actively destroy any codebase
