My short experience with Claude and ChatGPT via web is that:

1. the way GPT writes is simply fundamentally annoying. I pretty much had to create a project with a file that said "do not use headings, lists or emojis" to make it bearable. It feels like, as a product, this sort of thing should be a general preference the user sets before they even start talking to a chatbot.

2. Claude just loves wasting tokens doing things nobody asked for. You ask "how do I calculate the distance between 2 points?" and it's probably going to compile some C code in the background with tests to make sure it works, then generate an interactive diagram on the fly to show how the math works, and then give you a downloadable file with the code. Like, dude, I just want some text. Why are you doing all of this?

Both of these problems come from the obvious lack of any UI controls in the software. there is no way for the user to know what sorts of things the software can do, because it's not exposed via UI as a checkbox like "generate interactive diagram" or "avoid using emojis." Discoverability is burning tokens to figure out what prompts work, or looking at example prompts the developer placed in the welcome screen.

I just feel it's completely ridiculous how LLM's are essentially the culmination of a trajectory of bad UI practices masquerading as "good UX" and now they're being implemented everywhere because people think it's good UX a blank textbox where you don't even know what you're supposed to type to do something.