Trying out Cursor
The idea that the CS field is about to be flooded with know-nothing vibe coders worries me. Last year when I tried out Codeium in VS Code, I thought it was neat but nothing to worry about since I had to do a lot of corrections to the code to get it to work properly. There have been a lot of stories in the news lately about Cursor and Claude Code though that made me wonder if things have gotten substantially better since then. The thing that pushed me over the edge this week though was that I started thinking about how long it's taking me to refactor some data products using Pandas. It's fun tinkering with Pandas, but almost always I'm writing some throwaway code that nobody will ever see. It takes a long time to do simple things and I still catch myself making dumb mistakes. It all seems like the kind of stuff an AI tool could do better.
On Friday I caved and downloaded a copy of Cursor for Linux on my Chromebook. Youtube is flooded with smiling videos on how to install and use Cursor. My summary is, they forked VSCode and modded it so there's an integrated prompt session on the side. You tell it what you want it to do next with your project, the prompt and code go to their servers, they make their own prompts, and then it gets sent out to high-end AI engines like Claude and Gemini Pro. You pay them $20/month to use the service with throttling, or $200/month (sheesh) to get more prompts. The interface is very good- it does all the project creation stuff you want VSCode to do (eg create a venv, install packages, and add them to requirements), plus it checkpoints your state so you can undo things when things go wrong.
Bouncy Ball Game
I didn't have a great idea about what to build when trying out Cursor, so I fell back on the 1980s arcade game approach. I started by asking it to draw some bouncing vector point balls similar to the ones you see in 1990s Amiga demos. It created a project, installed pygame, and wrote some basic code that ran without any nudging. The balls vector dots, but they were 2D so I asked it to switch to 3D balls that rotated and spun. This actually worked, though it drew lines between points that I then had it remove with another prompt. I went on to add some pin ball style bumpers, which took some more prompts to refine. From there I added a line at the bottom with one hole, then two, to make it more of a game. With additional prompts I had it add a score count, use more and smaller balls, and flash a game over menu when the last ball went into a hole.
Admittedly, this should be an easy problem for AI because there are a ton of examples out there and pygame is well equipped to do a lot of the heavy lifting for you. I did get one Python error in the process, but copying the message into the chat motivated the AI to look at the code and realize there was a use before initialize error that it needed to fix. The code and comments seem pretty reasonable. I think the only annoyance was that there were some subtle collision problems in the game that didn't seem to be easy to fix (either through chat or by debugging the code). I hear that quirks like this often pop up with vibe code and that it's hard for anyone to go back and figure out how to fix them. My worry here is that the future will be filled with apps that look like they're working, but they'll have random mistakes in them that cause significant problems when you do nonstandard things.
- Prompts and code: ai-fiddling/250720_cursor_bouncy_ball
- Example movie
Platformer
I showed Cursor to my older son and asked him to write his own game so he knows what other people his age will be using. He chose to write a simple Mario-style platformer. The initial prompt yielded an Atari 2600 style game where a blue rectangle could wander around and jump from one green bar to another. It worked but he quickly realized that some platforms were unreachable while others didn't stop you from falling through to the ground. He issued prompts to fix the problems, then went about adding coins, a scoring system, and a level clearing message that included fireworks. He then asked it to switch the blob character to a Mario. Interestingly, the AI created a separate program that used pygame's shape primitives to draw a Mario character and save it to a png the game program could load and use. When he asked it to animate the character when it walked, it modified the image generator to make multiple images. At first it didn't look like the new images did anything different, but when we zoomed in on the pictures we realized it had put a single blue dot for a foot and toggled where it was when the character walked.
My son play tested the game and found that it mostly worked. There were a few frustrating bugs where if you hit a specific platform in a particular way you'd go through it. It reminded me of how the original Joust had a big with the pterodactyl that people loved. Maybe the quirks of a game are what make it special. Still, it's tough figuring out where all the bugs might be in a program like this, let alone fix them all.
PIAware Flight Receiver
My son was willing to put another 5 minutes into prompting things, so I had him ask Cursor to build some Python code to connect to a socket on my PiAware station. We gave it a specific IP/socket and told it that the port dumps out plane update information in a standard CSV format. When we told it to print out info every time it saw a new plane, it spewed out a lot of data that didn't look right. I realized it was getting misaligned because it wasn't handling the end of line properly. Once we explained the problem it fixed the code and properly printed out new plane IDs every few seconds.
While ADSB traffic is pretty well documented, I was impressed that the AI wrote code to open a socket, parse CSV lines, and print out the useful information. Writing this kind of code by hand is fun, but there are enough tricky spots that by the time I finish it I don't want to do anything else with it.
Thoughts
So, yes, just like other people are saying, Cursor is appealing because you can steer it to write code that would otherwise be tedious to write. It feels like cheating though, and requires you to keep a close eye on what it is actually building. The more people use Cursor, Claude, and the like, the better I think these tools will get. I do wonder how much developers will be willing to pay for these services. I think a lot of new developers will get hooked on the AI tools and then flounder when either the AI companies jack their rates. I'm still going through the 2 week trial process, but I'm thinking I wouldn't mind paying $20/month to hack on personal projects faster, especially if it means canceling one of the streaming services I have but never use.