Live data from Hacker News

Ask HN: Anyone writing code from scratch or mostly doing architecting and LLM?

news.ycombinator.com

11–19 of 19 posts

Re: Ask HN: Anyone writing code from scratch or mostly doing architecting and LLM?

#12
post #10

I've only been using AI for small editor tools or for common utils (math, physics, string manipulation) instead of importing libraries to reduce random dependencies. I find that most LLMs hallucinate and write too much bad code. However, with Claude Opus 4.5 it's completely shifted my entire day. I'm out here running 8 concurrent agents knocking items off our todo list like crazy. I've gone from 99% manual code to 1%…

> 8 concurrent agents

how are you managing to run and work between 8 concurrent agents to get through the ask in hand?

I keep seeing git workspace tree + agent combo being mentioned

Re: Ask HN: Anyone writing code from scratch or mostly doing architecting and LLM?

#13

Any systems / low-level / embedded developpers who could give their grain of salt on this? I'm a student and I'm specializing in embedded and systems programming, but everyone around me (webdevs) tells me LLMs will take over and writing C and Zig and Rust is useless... Should I just give up on my passion and embrace AI and high-level programming?

honestly i think LLMs make it easier to write with lower level languages and you get the upside of being able to optimize better since you have more granularity of control in something like C vs something like Python.

I don’t know if LLMs will completely take over, but they’re a useful tool today. I think it’s worth learning how to use them effectively, but also know how to work without them, and when to work without them.

Re: Ask HN: Anyone writing code from scratch or mostly doing architecting and LLM?

#14
Just thinking about physically typing out syntax again with my fingers makes me queazy.

I’m writing, designing and building more than ever before but Claude, Cursor and friends do all the grunt work. The most I’ll do is some pseudo-code if I want to clearly define a pattern for the LLM to follow - but that’s it!

Re: Ask HN: Anyone writing code from scratch or mostly doing architecting and LLM?

#15
post #10

I've only been using AI for small editor tools or for common utils (math, physics, string manipulation) instead of importing libraries to reduce random dependencies. I find that most LLMs hallucinate and write too much bad code. However, with Claude Opus 4.5 it's completely shifted my entire day. I'm out here running 8 concurrent agents knocking items off our todo list like crazy. I've gone from 99% manual code to 1%…

> 8 concurrent agents how are you managing to run and work between 8 concurrent agents to get through the ask in hand? I keep seeing git workspace tree + agent combo being mentioned

git workspaces, combined with some guard rails to keep the llm on task. For example a detailed document with requirements and implementation details and a unit tests that the llm can keep verifying against and/or enhancing between each step to increase the reliability of it’s output.

Give Claude Code a try, with git work trees and especially some skills pre-loaded https://github.com/obra/superpowers

Re: Ask HN: Anyone writing code from scratch or mostly doing architecting and LLM?

#17

Earlier quoted context omitted.

> 8 concurrent agents how are you managing to run and work between 8 concurrent agents to get through the ask in hand? I keep seeing git workspace tree + agent combo being mentioned

git workspaces, combined with some guard rails to keep the llm on task. For example a detailed document with requirements and implementation details and a unit tests that the llm can keep verifying against and/or enhancing between each step to increase the reliability of it’s output. Give Claude Code a try, with git work trees and especially some skills pre-loaded https://github.com/obra/superpowers

> Give Claude Code a try...

Two days after they released their web interface for Claude Code I was hooked. Haven't really used the regular interface or the app since. Oh god, I'll never go back to copy-pasting code.

Re: Ask HN: Anyone writing code from scratch or mostly doing architecting and LLM?

#18

I still write most things myself. So far, Copilot has been very unreliable with what I’m working on and it often costs more time than it saves. I mostly use it for one-off regex or json stuff. As far as Python goes, I think it depends on your goal. If you’re just trying to get something quick done, you can probably use an LLM. If you’re trying to learn something new and get a new skill, I think it’s worth going throu…

Automate the Boring Stuff (I love this book) should be a mandate reading for all people working with computers, and I mean ALL PEOPLE.

Re: Ask HN: Anyone writing code from scratch or mostly doing architecting and LLM?

#19
The point of an exercise like that isn’t to have the code that walks a directory. It’s to learn how things are done in Python by writing that code yourself. Most people retain a lot more if they have to piece together the code themselves than by just skimming code someone or something else wrote.
Post reply on HN