Using Cursor to Generate Plot Scripts

I often find myself in need of a tool to do some quick but good-looking plots of tabular data. Pandas and Matplotlib have been my go-to tools for generating these plots for years, but lately I've realize that I spend way more time than I should because I always have to do Google searchers for little tweaks. Yesterday while preparing the data for the glucose plots, I decided to experiment with Cursor to see if I could get the plots done faster by describing what I wanted and having AI build out the code for me. In this post I'm reflecting on what did and didn't work.

Basic Plotting

For the Glucose post my goal was just to load the data as a table, split it up by day, and do a vertical stack of plots that lined up based on the hour of the day. The input data was straightforward- just a CSV with a well-formatted date (with timezone!) and a numerical value. If I were writing the plotter by hand, I would have used Pandas to load the table and deal with dates, and then sent individual days of data to a subplot plotter. As good as Pandas is, there are still some tedious parts to making this kind of plot. First, dates often have parse problems, especially when timezones are involved. Second, I always have labeling problems when I use dates on an axis (eg, the wrong part of the date gets displayed or there isn't enough room for everything. Finally, I often waste a lot of time tweaking different settings for a plot, like tweaking labels, axis settings, and aspect rations when trying to stack a lot of plots together.

Describing the plotting code I wanted Cursor to build for me worked out fairly well. It properly used Pandas to read the data and did a to_datetime to convert the time strings. For the plotter it used a datetime.combine function I didn't know about to set the start and stop times for a day. Initially everything was off by 8 hours because of timezones, but it corrected it once I described what was wrong. It did the day splitting correctly and built a stack of plots that looked right without much effort on my part.

Tweaking

Once the basic plots worked, I went back and had it do some refinements. I cleaned up the overlapping plot problems by disabling the horizontal labels on the upper plots and had it move the title into the box area. I had it adjust the figsize and linesize properties to deal with full-sized vs thumbnail scaling problems. I added horizontal lines to show the cutoffs for good and bad regions. Finally, I told it to set the line color to green or red based on whether the line was in the good or bad regions. I don't usually put that kind of detail into a plot because it requires some effort to break up the data into color segments. However, the AI wrote the code I wanted and made plots that were better looking than I would have done.

CLI

Another problem with throwaway plotting code is that I'm often on the fence about whether I should hard code filenames in it or add in CLI parsing so that I can iterate with it on the command line and generate multiple plots. I usually use the Click library when generating CLI tools because it lets you control a lot of arg parsing behaviors through simple decorators. I've added some default instructions to my Cursor environment to use Click when implementing CLI tools, so it was easy to just modify the plotter to take input from the command line.

I instructed Cursor to allow the user to specify a single date to plot so that I could highlight individual days. Once that worked I prompted it to do a comma-separated list of days. This feature helped me generate the examples where an event was spread out over two days. Again, implementing this kind of code by hand is not hard, but it's another burden that gets in the way of finishing a plot.

Observations

While the AI model generated a few code blocks that were either sloppy or wrong, it worked very well for what I needed and let me work through generating the plots quickly. In some places it made better plots than what I would have done, simply because the amount of effort to code the feature was higher than the benefit it would have given me. With AI you can just describe it, see if it works, and then keep it if it's good.

An important point of this is that it works well for me because I already know how Pandas and Matplotlib work. I also know what questions I'll ask to get effective results from Google or Stack Overflow. I'm not sure how great this would be if you blindly asked it to make some plots. Like a lot of CS problems, using AI effectively boils down to I/O: it only works if you can describe what you want and then have the reasoning skills to interpret what it gives you back.

The transcript for the session is now available here: snippets/260124_glucose