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

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 through it on your own. The examples might be easy for the LLM, but those are building blocks you’ll want so you can understand more complex things that the LLM might trip on.

I would also say, don’t try to memorize everything. You’ll always have access to reference material, LLM or not. So it’s more about understanding and the concepts and building blocks, not about memorizing the exact way a specific thing is done in the book.

I went through Automate the Boring Stuff several years ago. I went through the Excel stuff, but didn’t really commit it to memory. I just knew a year later that Python could solve my problem when it came up, and was able to find the stuff I needed and put it together. It also gave me some ideas on stuff I could do that I probably wouldn’t have thought to automate otherwise. So going though it can help get your brain asking the right questions, even if you ultimately use an LLM down the road. It will also help you write better prompts.

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

“is anyone really writing code from scratch”

Oh yeah, lots of people do. I have friends that hate AI and refuse to use it at all.

However, I guess it depends on what you mean with “from scratch”. In a corporate context, I would say most code-writing is fixing bugs or adding features to an existing project.

“Am I wasting time doing these exercises when they can easily be done with LLM within one or two shots?”

Depends on what your goal is. If you want to get better at programming, how would you get better if you don’t let your neurons form new pathways? However, perhaps you are more interested in the output and not in the process. But some day, you might need to debug your code, and then those pathways in your brain might come handy.

Small targeted use of LLMs is the sweet spot for me, that can include Claude Code usage but not “write me this huge feature kthxbai”, I use it with dangerous permissions but I normally ask it to do small tasks that I review/tweak right away. That plus Copilot (auto-complete, I don’t care for the agent mode) is what I’m using right now.

I do lean heavily on Claude Code or other LLMs for small scripts and one-off tests. It’s amazing for that, “here are some log lines, I want I extract X, Y, Z from each line and write that out to a new file” or even little tools to wrangle out ticketing system “write a script to read ticket ids from a file and update each ticket to fill out the deployed version field with the value I pass into this script”.

Lastly, not exactly “writing code” but I semi-recently started asking Claude Code things like “find me all the places we do X in the code base” or “how does Y feature work”, it’s FAR from perfect but it’s an amazing starting point. Claude code doesn’t need to write code to be useful, just being able to spider through my codebase and pull out insights is amazing.

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% over the course of the last 4 days.

Never thought this was going to happen until at least 2060...

> 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.

My personal POV is that I've been coding from scratch when coding for creativity. It has been wildly refreshing to go back to handwriting some cool little HTML/CSS/jQuery(sometimes) static websites.

From a professional & learning standpoint? Some good points in this thread:

- Realizing I still need to code from scratch when setting up initial architecture/directory stuff like kryptoyagi said. GitHub Copilot fails at this constantly (or I suck with it)

- cpach is right... I'd guess that a >50% majority of software devs worldwide are barely using AI lol. I graduated CS not even 10 years ago and like 7 out of 8 of my friends dont want to use it (or are working on old-ass repos/govt repos)

- aslansahn is spot on, I feel like this has been the most efficient approach. Plan everything out, have it go one-by-one for you, step in manually once in awhile to course correct if necessary

I'm an experienced SWE. I brainstorm the architecture using LLMs. After finalizing my design in Notes I break it down into independent components. Then, I ask an LLM to implement each component one by one in a single thread. Each component should be runnable on its own using unit tests.

When it comes to writing small scripts for various utility tasks, I now rely heavily on LLMs.

Additionally, lets say I would like to write something new that I don't know anything about, I now go first to a LLM, or two of them, to get an idea of what would be the best approach. It saves a lot of time that would otherwise go in research and discovery.

But I still don't trust it enough to write production level code. When I am writing something that I know will be released in a product, I steer clear of LLMs.

All of this when I am doing my day job. For my side projects, I try not to use LLMs.

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.

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.

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!

When creating a new project, I code from scratch to get the architecture/folder/file structure coherently setup.

However, when contributing to existing repos, I use Claude Code for: 1) understanding repo structure, flow of function calls, 2) making small fixes, 3) using plan mode to iterate on how to implement a larger feature.

On the topic of small fixes, I use git worktrees to use Claude Clode parallelly. Very useful to tackle 3 or 4 small fixes at once.

The diffs created by Claude Code can get a bit disorganized in my head, so I use GitHub PRs to manually check diffs.

Regarding your question, you are not necessarily wasting time doing the exercies. Being able to "code from scratch" accelerates your ability to use LLMs in a controlled, but powerful manner, without degenerating into vibe code slop.

You are reading too many advertisements. Take a break from the internet astroturfing and touch some grass (and write some Python).

I hate that people even have to consider this now... Learn it yourself!