And so you don’t think I just dropped in to try it out: no, I’ve been using Cursor every day for many months. I set up the rules very precisely, make maps, write instructions and constraints so the LLM understands how to work with the current codebase—where to find things, what to do, what not to do, what things should look like. I try to write as concisely as possible so everything fits in the context window.
I have a monorepo, so different instructions and rules live in different corners of it, which I manually add as needed. Doing all this with Claude is hard: it kind of works, but it’s a kludge. It’s much easier to manage through the UI. As the saying goes, it’s better to see something once than to remember it ten times, type a command, and configure it.
I can say the same about Vim and Emacs users. No one has yet proven to me they can code and navigate faster than an ordinary programmer using a trackpad or a mouse and keyboard. It’s purely a matter of taste. I’ve noticed that people who love those tools are just bored at work and want to entertain their brains. That’s neither good nor bad, but it doesn’t give a real boost.
By the way, there’s a study (a bit old, admittedly) showing that working with an LLM still doesn’t provide a massive speed-up in programming. Yes, it spits out code quickly, but assembling a finished result takes much longer: you have to review, polish, start over, throw things out, start again. It’s not that simple.
Speaking of Claude, he’s a real master at shitting in the code. At the speed of sound he generates a ton of unnecessary code even when you ask him not to. You ask for something minimalist, nothing extra—he still slaps on a pile of code with useless tests that outsmart themselves and don’t work.
That’s the downside. Overall, it’s easier to program when you just say what to do instead of sitting there straining your own brain. But even then it’s not simple. If you yourself haven’t figured out very carefully what you’re doing, you’ll end up with crap. I almost always throw away what I wrote with an LLM when I was tired. I don’t think there’s ever been an exception.
The only way to write something decent is to dig in thoroughly: draw diagrams, write everything out, and only then tell the LLM exactly what to do. Then double-check, make sure it didn’t write anything extra, delete the excess. And only like that, in small chunks. The “make me a feature, I pressed a button and walked away” mode doesn’t work if the task is even slightly complex.
Yeah, I got carried away—had to let it out. I have a lot to say about LLMs and programming with them, and not just about Cursor and Claude. The tools completely upended the way we program...
I remember a few projects before all the AI where I would setup templates, little code generators, patterns to follow. Just make it really easy to do the right things. Was so easy to get things done. Was a small team, like minded, easy to row in the same direction. Your post reminded me of that.
I too notice this about Claude, I've written commands, massaged Claude.md, even hooks. Given it very precise feature guides. I see the same issues you do. Feels like Claude has a lazy lever built in to it.
I play GPT O3 off it constantly, it takes the spec, a repomix slice of the Claude work and gives a new gap.md for Claude to pursue. Rinse and repeat. It works but your cursor flow seems better.
My highest velocity was about 1.6 fib complexity points a day over thirty years, now it's 4.3 with Claude the last three weeks which is nuts. I'm a total hack, I think I could get it up to nine if I was a bit more organized. Probably won't have to, just wait for the next iteration of the Claude models.
Yeah, accurately since 2008. Before that was just a guess but it wasn't at all fast. I've worked a lot of places but always insisted on story points based on estimated dev complexity.
I have a monorepo, so different instructions and rules live in different corners of it, which I manually add as needed. Doing all this with Claude is hard: it kind of works, but it’s a kludge. It’s much easier to manage through the UI. As the saying goes, it’s better to see something once than to remember it ten times, type a command, and configure it.
I can say the same about Vim and Emacs users. No one has yet proven to me they can code and navigate faster than an ordinary programmer using a trackpad or a mouse and keyboard. It’s purely a matter of taste. I’ve noticed that people who love those tools are just bored at work and want to entertain their brains. That’s neither good nor bad, but it doesn’t give a real boost.
By the way, there’s a study (a bit old, admittedly) showing that working with an LLM still doesn’t provide a massive speed-up in programming. Yes, it spits out code quickly, but assembling a finished result takes much longer: you have to review, polish, start over, throw things out, start again. It’s not that simple.
Speaking of Claude, he’s a real master at shitting in the code. At the speed of sound he generates a ton of unnecessary code even when you ask him not to. You ask for something minimalist, nothing extra—he still slaps on a pile of code with useless tests that outsmart themselves and don’t work.
That’s the downside. Overall, it’s easier to program when you just say what to do instead of sitting there straining your own brain. But even then it’s not simple. If you yourself haven’t figured out very carefully what you’re doing, you’ll end up with crap. I almost always throw away what I wrote with an LLM when I was tired. I don’t think there’s ever been an exception.
The only way to write something decent is to dig in thoroughly: draw diagrams, write everything out, and only then tell the LLM exactly what to do. Then double-check, make sure it didn’t write anything extra, delete the excess. And only like that, in small chunks. The “make me a feature, I pressed a button and walked away” mode doesn’t work if the task is even slightly complex.
Yeah, I got carried away—had to let it out. I have a lot to say about LLMs and programming with them, and not just about Cursor and Claude. The tools completely upended the way we program...