Two programming things I have noticed LLMs are bad at (scope includes Claude 3.7, Grok 3, and gpt-4.1) in late April 2025: * Anything to do with time and date APIs, especially if it involves timezones. I guess that's fair as these desperately confuse humans too. * Layering instructions. I have a conventions file for my SQLite application that instructs the genie not to inline SQL calls in the business logic, but rather to create new methods for the back-end manager class. It blithely ignores this. On the other hand, LLMs are better at web UI design that I am. They have a kind of synthetic tastefulness that is doubtless formed by looking at lots of examples, and is surprisingly good - though I'm sure a lot of web designers would consider it bland and austere. More importantly, they can remember all the fiddly details of CSS, which is more than I can do or even want to. Usually I can just tell them I want to page or page element to look a certain way and they'll do it. When I told Claude I wanted my application to have a common navigation-bar header, without asking it set things up so that the navbar entry corresponding to the page that you're on is always highlighted. And afterwards, when I created a new endpoint, it just added to the nav bar appropriately without being prompted. I like that they don't instantly reach for an overcomplicated 17-ton JavaScript library to solve problems. This is more couth than many human web designers can manage.