Hacker News Reader

Comment by echelon | Hacker News Reader

I work for myself and I absolutely love AI.

I'm able to get 3x the work done. Greenfield stuff appears almost immediately.

My job is providing value to customers, not worshipping at the cathedral of software that will last forever. Nothing lasts forever.

Start treating software as ephemeral. It'll click.

This doesn't mean write low quality, unmaintainable software. It just means focus on getting stuff to your customer.

Writing in super typesafe languages with the highest level of strictness helps a lot. My AI stack is Rust and Typescript.

View on HN

I tried using it last week to make a simple Yu-Gi-Oh! website, that shows decks, lets you rate them, register users, etc. kinda like masterduelmeta.com and I enjoyed using it, but definitely did not enjoyed making it. I didn't felt a sense of ownership or dopamine from nailing the styles just right, or making the cards shimmer when you hover them.

All jobs can generate income. What led me follow this one job in particular was the joy of turning nothing into something, and it now feels that the most effective way to do that is for only $99.99/month, and that price needle is only going to move further upwards as capabilities increase.

> it now feels that the most effective way to do that is for only $99.99/month, and that price needle is only going to move further upwards as capabilities increase.

That's not how economics works.

That can happen briefly with monopolies and ossified markets, but there is typically always an alternative that will seek to break in and grab market share.

Chinese tokens are pretty cheap and they'll gladly undercut US hyperscalers.

This is the right way to look at things now. It might not always have the right track record, but AI built coding is more likely to have all the right permissions in place by default, most likely to copy existing patterns in your codebase, most likely to use the highest performance patterns and on top of all that, the spec will match what was asked of it.

What magical AI are you using? That’s not my experience at all.

there is a significant learning curve to using AI well. learning to stay skeptical and keep your brain on, developing an intuition of how much free reign to give it, writing ironclad specs and design docs and keeping them updated, making work easy to inspect, the tone you use talking to it, using one agent to critique another's work, etc.

basically, AI will produce slop if left unattended. but it's not really its fault.. it's a process failing, like not supervising the interns. using AI the Right Way(tm) is a mental workout, quite a bit slower, but extremely rewarding (ime.)

Claude with the 4.7 model is getting pretty good.

I use cursor on auto mode almost all the time. I switch to Opus 4.7 when I need I know it will go off the rails. But generally Auto mode "just works".

For my personal work on my own projects, Codex 5.5 because it's cheap at $20 / month and I get in about 10 prompts during my work day (would be more like 40 if I was not focused on work though)

I can't even get LLMs to reliably use tool calls instead of bash, let alone follow existing patterns in a codebase.

What do your prompts look like?

Mine are pretty robust and articulate. I tend to write very lengthy instructions and include snippets of code, file paths, struct names, etc.

Reply on news.ycombinator.com