Nicholas Van Doorn Software Blog

Testing Cursor in 2025

May 5 2025

Anyone who knows me well knows that I'm a bit of an AI skeptic. I like my tools old and simple, and I like self hosting anything I can. vim and rsync are amongst my favorite software in the world. I don't really use ChatGPT, I don't use any AI tools in my day to day life, and I try to avoid reading any large chunk of text obviously generated by an LLM. But I must admit, the hype has grown so strong (particularly in the software field) that it's becoming too risky to ignore. I tried GitHub Copilot in VS Code last year, but it was really neither here nor there in my opinion. The autocomplete suggestions range from "kindof helpful" to completely wrong and useless. I never really got too far generating anything with prompts in the Copilot chat either. Now that I have some time on my hands in between jobs, I figured this was the perfect opportunity to re-evaluate things.

So where does one get started with this? I've tried my best to keep tabs on some "AI thought-leaders" and worked backwards from what they're excited about. It's become clear to me that models are changing fast and it would be foolish to focus too much on the model-du-jour. Instead, I decided to look at tools one level higher that could make use of a wide range of models. There are two AI powered IDEs that stand out here: Cursor and Windsurf. Cursor seems to be very popular amongst the AI enthused, so it's where I'm starting.

My first impressions of Cursor are quite good. It's an obvious leap forward over using Copilot in VS Code. It's the first AI tool I've felt compelled to keep using in my day to day development flow. I'll continue posting about it as I use it more, but I wanted to highlight a few features I've enjoyed right out of the gate.

Building on top of an established editor (VS Code) brings a large catalog of high quality extensions. I still prefer vim at my core, but being able to import all of my favorite VS Code extensions made me feel quite at home. I imagine VS Code users would have a very easy transition path.
Cursor allows you to ignore model hype, if you want. Cursor ships with a large array of quality frontier-tier models, and you can leave it on "auto" if you want. I'm sure at least one AI investor is jumping up and down saying "NOOO you need to use 4O-mini" or something, but in my early stages of adopting Cursor, I really like spending less time thinking about this. Maybe I'll have a strong opinion about my preferred model soon.
Easy control over context. Cursor is excellent at giving you control over which files are loaded as context when interacting with the agent. You can easily point at any file inside of your project, or even point it at external sources like docs. So far in my experience, keeping the agent focused on the relevant context will get you much further.
The agent will call out to the appropriate CLI tools, if available. As a long time Rails developer one of my primary fears of AI assisted software development was the rogue generation of new files with no regard for the existing project standards and structures. Thankfully if you request a new file in a Rails project, it will typically generate the appropriate rails generate command and offer to execute it.

So where does all this leave us? As per usual, I would say reality is firmly in the middle of the hype and the critique. The people telling you every job will be replaced by AI within a few years are probably wrong, and the puritans who refuse to use any AI assisted tools are likely missing out. I'll keep hedging my bets and see where we end up.