Cursor is this sort of AI coding beast that can empower you to build digital castles in no time and torch the whole thing moments later.
It's a whole new paradigm. One where raw coding skills become less central. What matters more is engineering judgment, system thinking, and knowing how to guide the AI effectively.
It can write code at a superhuman speed. But speed is a double-edged sword. Without proper guidance, it can totally ruin a project and send you to hell for hours.
My fresh thoughts on my first week of working intentionally with Cursor are largely positive. It's a very promising tool. But it requires working in a radically different way.
So what does this new way of working actually look like?
Managing Context
Context is everything for LLMs. Give them too little and you'll get generic, unhelpful responses. Give them too much and you'll lose them entirely.
On a more technical level, an LLM's context window is essentially how much information Cursor can effectively track and reason about.
Too little context means it doesn't understand your codebase, your patterns, or what you're actually trying to achieve.
On the other hand, too much context will cause it to forget things, make it prone to confusion all the way down to straight-up hallucinations.
Managing how much information to give an LLM is a crucial part of the workflow.
What works well then is to break up the work in chunks. Contained, clearly outlined, with just enough context - this tends to yield very good results.
And getting to that level of preparedness where you can confidently say "ok now go ahead and code" is a skill in itself.
The Supervisor Mindset
I feel like I spent this past week being in supervisor mode. To get the best out of this AI Agent, I need to communicate very clearly what the goals are and leave as little room for interpretation as possible.
It really flexes my communication skills, as if I had a teammate or assistant alongside me at all times that I need to guide.
I like to think of it as: how many forks in the road are there? If it's too many, it's likely to get lost. With every turn, the probability of it going off into the wrong direction increases.
Meanwhile, I get to face my own fork on the road:
Spend more time upfront being very thorough about communicating what needs to be done and how. This usually means I'll spend less time adjusting and cleaning things up after.
I could spend less time upfront writing specs but I'm more likely to have to go back and forth quite a bit to refine things and get to where we need to be.
At the end of the day, I'm developing an intuition for this balance. Sometimes the upfront investment pays off, sometimes iterating quickly works better.
It depends on the complexity of the task and how well I understand the requirements myself.
Taming the Beast
Cursor has this sort of "go-getter" attitude mixed with "your wish is my command" type of personality.
Getting it to not rush ahead and start doing a bunch of unsolicited tasks is surprisingly difficult, while having it disclose its level of uncertainty and slow down a little is downright infuriating.
But once I was able to do that, essentially by telling it that I LOVEEE being asked questions (remember it likes to please?), things really improved!
Turns out, it has great questions! They revealed how much I would leave certain things up to interpretation in the way that I was communicating, or that I simply didn't think things through enough.
It could anticipate things I was not seeing. Like an insightful companion.
The Tradeoffs
This sort of workflow lets me focus on architecture and product engineering more. I'm not as involved with the minutiae of coding itself, which has its own tradeoffs.
There is certainly a threat here: writing code is a skill that takes practice. Without reps, it atrophies.
So much of understanding code comes from writing it. It's really hard to learn and read new languages without writing them! So there is something to be careful about there.
It's hard to imagine a reality where one could be a good software engineer without knowing how to code.
It takes literacy to be able to think effectively in this domain. I think the danger isn't so much "will this take my job?" as much as "will this make me suck at my job?"
Time will tell. For now, learning to work with Cursor seems like an incredibly valuable skill.