related: Cursor composer line of models is *so good* relative to cost. "auto" se...

operatingthetan · 2026-04-17T00:52:23 1776387143

I don't think Cursor was _that bad_ in it's time. But the 'psyop' here is that anyone is using an AI-IDE going forward at all. I see people who say they are still using them and are so excited, but then I talk to engineers I actually know and it's all CLI tools.

gen220 · 2026-04-17T03:06:31 1776395191

FWIW, cursor (company) has a CLI tool/harness similar to Claude Code called agent.

It’s existed for a long time, is quite good, and it is under-marketed (ironic for this thread).

(Double-ironic disclosure… I work for Cursor. If you have ideas to make agent better hmu)

majormajor · 2026-04-17T01:48:50 1776390530

> I don't think Cursor was _that bad_ in it's time. But the 'psyop' here is that anyone is using an AI-IDE going forward at all. I see people who say they are still using them and are so excited, but then I talk to engineers I actually know and it's all CLI tools.

This is just the old "surely nobody actually likes Lady Gaga, all the people I actually know think her stuff sucks, it's just all bought and paid for" reasoning trap all over again...

operatingthetan · 2026-04-17T01:51:24 1776390684

You couldn't even keep your analogy straight. I didn't say the people I know said anything at all about Cursor.

If someone is clear about offering an anecdote, it's dishonest to pretend as if they were making a real and reasoned argument.

stavros · 2026-04-17T03:10:09 1776395409

I haven't used Cursor much lately, but with Opus/Codex I can program with very few bugs without having to look at any code at all, over months of working on the same codebase. I don't think any other model can do that, no?

apsurd · 2026-04-17T04:20:39 1776399639

In Cursor I look at the code diff for my own satisfaction and understanding. I've been able to auto-accept all changes pretty reliably all last year.

So much so that i've yet to invest in CC. Finally downloaded the desktop app but use it exclusively for chat and cowork.

Cursor's purgatory UX is what'll finally get me to invest in CC and codex. Not model performance.

I do think there's a caveat that it's pretty standard nextjs with rails api.

edit: found your blog post about your experience, i'll read it! https://www.stavros.io/posts/how-i-write-software-with-llms/

edit2: maybe it's because i spend a lot of time being clear with small and surgical asks after doing purely thought exploring prompts to confirm and home in on approaches. At the point I hit build or do Agent mode, Composer is mostly always spot on.

with Opus (haven't tried) maybe people are doing large-scale multi phase and single shot prompts that trigger a swarm of sub agents?