Trying Cursor for a week
A quick note on the editor everyone on my timeline suddenly switched to.
Everyone switched to Cursor in the last month, so I spent a week on it to see what the fuss is about.
It's a fork of VS Code with two things bolted on that actually matter: a smarter tab-completion than Copilot, and a side panel called Composer that edits multiple files at once against whatever frontier model you point it at.
The tab completion is the part I didn't expect to care about. It's not "finish this line" — it's "I see you renamed a variable on line 14, here's the same rename on lines 22, 48, and 71, hit tab to accept each jump." It reads intent from a tiny edit and extrapolates. Most of the time it's right. When it's wrong, it's wrong quietly and you just keep typing.
Composer is the part everyone talks about. You describe a change in English, it proposes a diff across however many files, you review and accept. It's good. It's not magic. It's best when the change is mechanical but tedious — rename a concept across a codebase, extract a component, add a field through three layers. It's worst when the change requires a taste call you haven't communicated, because then it just picks for you.
A few things I'm noting for later:
- The pricing model (flat monthly fee, they eat the API cost) only works as long as model prices keep falling. If they don't, this gets expensive for them fast.
- It's a VS Code fork, which means every time upstream ships something, Cursor has to rebase. That's going to be a problem eventually.
- The lock-in is the muscle memory, not the software. After a week, going back to plain VS Code feels slow in a way I can't un-feel.
I don't think this is the final form of AI coding tools. I think it's the first one that's good enough to become a habit, which is most of what matters.