Comment by yoav | Hacker News Reader

Comment by yoav | Hacker News Reader

I installed claude code (an agentic ai that runs on your machine and can browse and edit code, run the build, and use unix commands)

This has largely taken me out of the loop. I give it detailed task like I would a junior engineer, and we discuss approaches and align on direction, priorities, and goals, and it then goes off for literally hours iterating on the code.

I have done about 3 months worth of extremely complex engineering work in about a week since I started doing this.

It is a step change from trying to use the chat interface and copy/pasting snippets.

Once it’s going it writes code like a staff engineer.

There are some obscure bugs it can run into that it needs my 20 years of experience to unblock or unwind when it goes down a rabbit hole.

But it has accelerated my development 500x and while it’s iterating I’m not filling my mind with code and syntax and so on, I’m more like a tech lead or manager now. I’m in another room playing with my dog.

This.

I have started testing Copilot for fun, my wife needs a web-based project that is not maintained for a while now and written in PHP.

I asked Copilot (Agent mode) to translate it to rust just for the fun of it, seeing how far it would come - I expected nothing out of it. I broke down the tasks into manageable chunks, directed it in design choices, and asked it to use some specific frameworks.

So far it wrote 40k lines of rust on its own, and keeps track of what functionality is missing compared to the original project. It was impressive seeing it iterate alone for 30m+ at a time.

I'm no programmer, more systems/cloud engineer, so a rewrite like this would have likely costed me >2 years of work and still ending up useless for all intents and purposes. I'm pretty sure that the end result won't work at first try, and I'll need to fix stuff manually or direct Copilot to fix it, but after two weeks of 1-2h iterating at night, I have 90% of something that would have required someone full time for at least a couple years.

The two things I found most valuable (also in other things, like shorter bash and python scripts):

1. Syntax boilerplate: if your task is specific enough, it normally gets it right 99.99% of the time, and my brain can look at the actual logic rather than the commas, brackets and (in python's case) spaces

2. Documentation: I spend 95% less time looking at documentation for something, I don't need to comb the entire language/package/class for the specific things, it normally gets it right, and worst case I can ask it to do a refactor with the most modern standards for specific library version X

Are you exaggerating about "literally hours"? If true, that implies you're making better use of CC than I am. For me the record is maybe 10 mins. Do you write extensive functional/technical instructions? E.g. how large is your prompt (incl. task descriptions in files the prompt refers to). Curious to know.

I have been giving it about the same amount of context as a large github/jira/whatever issue.

But i’m giving it a goal with a fairly distant milestone where it needs to hit 5-10 smaller milestones and even for advanced ai mixing the system languages I’m mixing in this large project there are build errors it powers through on its own (fixing the building, linking, or syntax, or using the apis for the versions of all the things i’m using) and there are some combinations of tech (windowing, webviews, etc.) and some architectures of things that require it to try multiple approaches to succeed.

Sometimes it will ask like “how am i doing so far should i continue”, or can you run this visual thing and report back. But there can be 30-60 mins between it doing that.

It also has to scan a lot of files and find the things it needs to find to fill its own context. Like “let me examine how you implemented that on mac and windows and come up with a solution for linux”

Other times while it’s reasoning if i notice it’s going down a rabbit hole I’ll point it out and set it in the right direction without interrupting it and when it moves onto the next thought it’ll incorporate my suggestion but it didn’t really stop or need another big prompt or goal.

There are plenty of times where I do need to step in and baby it with small goals through something tricky or obscure. But even in those cases I don’t have to write any code directly i just ask it to do xyz small change and hand it back for me to test

Reply on news.ycombinator.com