I’ve been toying with LLMs for a while now. Claude Code is the first thing that I’ve used with an LLM that does not feel like a toy, but is instead genuinely useful.

There is an exceptionally tedious task at work that would take me 2 to 4 hours, and would be the most boring, tedious time. Without going into detail, it’s a task that requires passing commands to a native library based on its own bespoke configuration. I fed Claude Code a PDF of the manual for that library, an example of the output of the library before the changes, and the config the library, and it completed the same tedious thing in 16 minutes.

Now I am seeing if it can automate my timecard at work off of my Outlook calendar. I don’t have access to the Graph API or the timesheet system’s API. Claude suggested using COM interfaces in the desktop version of Outlook, and then populating the timesheet with Playwright. I’ll let you know how this goes.

I think Claude has better taste than ChatGPT when it comes to writing code. ChatGPT’s responses tend to reflect the average opinion of the internet, and thus tend to spit back the kind of bad architectural advice people give on Reddit and other similar places. Claude’s output tends to be more idiomatic or domain specific.

I baked them off by having both of them answer a question about how to calculate a moving average for a value I will be tracking in Starbase Zebra’s database. ChatGPT wanted me to add a bunch of manager classes and a service layer, which was a lot of unnecessary cruft. Claude’s Django code was more Pythonic and followed the idioms of the Django framework quite closely.

Gippity is a much better poet, though.