Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

What Claude Code chooses

amplifying.ai

Post reply on HN
Print view View original

251–260 of 266 posts

Re: What Claude Code chooses

#251

Post by empath75 »

torginus wrote: post #171
empath75 wrote: post #98

Earlier quoted context omitted.

Speaking of rolling your own things, i had claude knock out a trello clone for me in 30 minutes because i was irritated at atlassian. I am already using it for keeping track of personal stuff. I’m not going to make a product out of it or even put it on github. It’s just for me. There are gonna be a lot of single team/single user projects. It is so fast to build working prototypes that it’s not even worth thinking if…

I'm sure after you've had Claude build it, you learned a ton of how to build such a thing (I certainly did for my projects). Basically, the data model is dead simple, you just spin up a SQLite db, create a React frontend, grabbing a good drag and drop library that implements these cards, write some simple but decent looking CSS, some React and backend boilerplate to wire the thing together - and boom - you're done. T…

But i don’t want to learn how to do that. I already know how to write code and i have built a bunch of production rust apps from scratch. I would prefer if i never have to write code. I don’t enjoy writing code, don’t take pride in having that skill. I enjoy the process of identifying a problem and producing a solution to that problem. Coding is just something i had to learn how to do in order to solve problems.

The absolute moment that ai is capable of producing all the useful code for a project I probably will never write another line of code again.

Re: What Claude Code chooses

#252

Post by gck1 »

deaux wrote: post #154

Supreme irony: this website itself is a better exercise in showing what Claude Code uses than the data provided. Everything current Claude Code i.e. Opus 4.6 chooses by default for web is exactly what this linked blog uses. Jetbrains Mono is as strong of a tell for web as "Not just A, but B" for text. >99% of webpages created in the last month with Jetbrains Mono will be Opus. Another tell is the overuse of this font…

Which is why I find "LLMs will replace x in 12 months" so amusing. I've used LLMs to write decently sized backend projects and they turned out okay.

I also used it for several FE projects and all of them turned out absolutely terrible.

The only difference is that I have 15 years of BE experience and 0 years of FE experience. Had I allowed it to make the same average decisions when working on BE, they would share the same fate.

Re: What Claude Code chooses

#253

Post by gck1 »

ghm2199 wrote: post #108

Ist why I never give it such vague prompts. But it's sad it does not ask the user more. Also interesting and important to know how one would tease out good and correct information from llms in 2026. It's like relearning now to Google like it was 2006 all over again, except now it's much less deterministic. I wonder how the tail of the distribution of types of requests fares e.g. engineer asking for hypothesis generat…

I do this too, but the issue I have with this approach is that it's a never ending cycle. Codex/GPT will always find holes and claude will always agree they are holes. If you teach it YAGNI, then it will always disagree even on genuine holes.

If your original plan was to add a column in your db, after several cycles, your plan will be 10,000 lines long and it will contain a recipe on how to build a universe.

Re: What Claude Code chooses

#254

Post by gck1 »

oblio wrote: post #226
comboy wrote: post #220

Earlier quoted context omitted.

What kind of tools do you have on your mind specifically? My experience is that LLM can create me a decent dev tool that I wouldn't ever bother making so nice myself.

IDEs, graphical debuggers, etc. It's extremely weird that 40 years after TurboPascal, 30 years after Delphi and VBA, we've only regressed in terms of truly integrated development environments. Heck, even programming languages have regressed. Python and Javascript are less type safe than Java circa 2005. Even though we have technology needed to make type safe languages much more ergonomic, since then.

On the other hand, LLM will force most languages to either become as masochistic as Rust with restricted clippy, or disappear.

Python is not going to survive once people start figuring out it's much easier to keep LLM agent in check if language actively fights back.

Re: What Claude Code chooses

#255

Post by indymike »

NewsaHackO wrote: post #99
indymike wrote: post #65

Earlier quoted context omitted.

> data on the equivalent of “ad impressions”. 1. They can skip impressions and go right to collect affiliate fees. 2. Yes, the ad has to be labeled or disclosed... but if some agent does it and no one sees it, is it really an ad. So much to work out.

How would it be paid for?

Affiliate fees.

Re: What Claude Code chooses

#256

Post by jascha_eng »

tomhow wrote: post #248
jascha_eng wrote: post #204

Earlier quoted context omitted.

Okay fair about the mentions but I don't that email is a good process: 1. It puts more effort on me as a user to report the spam via email because I have to open my email, compose one by hand and add the reasoning. The offending user in comparison probably automatically spams. Can't we have a button at least? 2. It doesn't make the community aware of the ongoing issue. Other community members could be primed that cur…

It’s fine to just flag things and move on. We’re considering adding additional parameters to the flag function, but till then, emailing us with “LLM?” in the subject and the comment ID/URL in the body is great, and should be faster for you than a comment and faster for us to be able to act. The community is well aware of the issue and off-topic meta discussion has always been against the guidelines here. We’ve discus…

The fact that this comment had 10 comments and full comment chains that didn't notice it is LLM generated tells me the community is not aware enough. It was also upvoted a lot.

I think there is significant value in making people second guess content and look at it critically. Especially in a time where it is so easy to fake expertise. We all need to train that skill anyway these days for all online interactions.

10 years ago it was clickbait titles that we needed to learn to ignore, today it is LLM generated content. We will get there, but by not calling it out publicly we are making it easier for adversaries to fool everyone.

And yes I don't want to falsely accuse anyone of LLM slop either but they can defend themselves and making mistakes is part of the learning process for all of us. Writers and commenters will learn how to not sound like an LLM and we will more finely atune to the nuance between polished human writing and AI.

Re: What Claude Code chooses

#257

Post by oblio »

gjm11 wrote: post #194
hinkley wrote: post #61

Earlier quoted context omitted.

How do you make an LLM that’s was trained on average Internet code not end up as a midwit? Mediocrity in, mediocrity out.

If you take thousands of photographs of human faces and average them out (even if you do it just by roughly aligning them, overlaying, and averaging the pixels) then what you get is a (perhaps blurry but) notably more attractive than average human face image. LLM output could be like that. (I am not claiming that it actually is; I haven't looked carefully enough at enough of it to tell.) Humans writing code do lots o…

Beauty is symmetry and most people are not fully symmetrical so averaging them bring symmetry.

Wisdom of the crowds is a thing, so who knows, maybe intelligence is the same way.

Re: What Claude Code chooses

#258

Post by ghm2199 »

gck1 wrote: post #253
ghm2199 wrote: post #108

Ist why I never give it such vague prompts. But it's sad it does not ask the user more. Also interesting and important to know how one would tease out good and correct information from llms in 2026. It's like relearning now to Google like it was 2006 all over again, except now it's much less deterministic. I wonder how the tail of the distribution of types of requests fares e.g. engineer asking for hypothesis generat…

I do this too, but the issue I have with this approach is that it's a never ending cycle. Codex/GPT will always find holes and claude will always agree they are holes. If you teach it YAGNI, then it will always disagree even on genuine holes. If your original plan was to add a column in your db, after several cycles, your plan will be 10,000 lines long and it will contain a recipe on how to build a universe.

The "trick(s) here are to limit the scope by always reading the plan very carefully. Here is how I do it to tackle this problem:

1. You should recognize when said holes are not "needed" holes e.g. you could make do with in memory task scheduler without rolling out more complex ones.

2. You can break up the plan— longer plans have more holes and are unwieldy mentally to go 20 rounds with in a chat coding UI.

3. Give it Learning Tests: i.e. code to run against black boxes. It's just like how we write a unit test to understand how a system works

Re: What Claude Code chooses

#259

Post by jugg1es »

gck1 wrote: post #254
oblio wrote: post #226

Earlier quoted context omitted.

IDEs, graphical debuggers, etc. It's extremely weird that 40 years after TurboPascal, 30 years after Delphi and VBA, we've only regressed in terms of truly integrated development environments. Heck, even programming languages have regressed. Python and Javascript are less type safe than Java circa 2005. Even though we have technology needed to make type safe languages much more ergonomic, since then.

On the other hand, LLM will force most languages to either become as masochistic as Rust with restricted clippy, or disappear. Python is not going to survive once people start figuring out it's much easier to keep LLM agent in check if language actively fights back.

thats an interesting take on python. I think I might agree with you

Re: What Claude Code chooses

#260

Post by tomhow »

jascha_eng wrote: post #256
tomhow wrote: post #248

Earlier quoted context omitted.

It’s fine to just flag things and move on. We’re considering adding additional parameters to the flag function, but till then, emailing us with “LLM?” in the subject and the comment ID/URL in the body is great, and should be faster for you than a comment and faster for us to be able to act. The community is well aware of the issue and off-topic meta discussion has always been against the guidelines here. We’ve discus…

The fact that this comment had 10 comments and full comment chains that didn't notice it is LLM generated tells me the community is not aware enough. It was also upvoted a lot. I think there is significant value in making people second guess content and look at it critically. Especially in a time where it is so easy to fake expertise. We all need to train that skill anyway these days for all online interactions. 10 y…

In another comment thread yesterday, I pointed out an effect I’ve observed, which is that people are less likely to notice or feel negatively towards LLM-generated content if they like it or agree with it. That’s just a reality of human psychology, and has always been the way swindlers and con artists have their way with people.

The answer is not to suddenly abandon one of HN’s most important principles, which is that we want to keep discussion on-topic and discourage off-topic meta discussion.

We’re developing software to respond to the new challenges that are emerging, and we already have plenty of people who email us to when they notice generated content, which is great.

We know there is no perfect approach; there never is in a large community of humans. But principles matter, and we need to trust that the principles that have made HN what it is over nearly two decades will keep it strong for years to come.

By the way, your own original comment called on us to “Please do something more rigorous than manually deleting accounts”; we are doing that but these things take time to develop, test and perfect on a platform like ours.

Post reply on HN