I believe it is related to watermarking. Text can't encode additional information like a image can, so the watermarking has to come from the structure, word choice, voice, etc. Claude has to sound like Claude so Claude can tell if text comes from Claude.

I also think this is why the models aggressively drift back to the exact same word choice. You can put it in CLAUDE.md, explicitly prompt it, doesn't matter. A prompt or two later it's back on its bullshit.

Exactly, and frequently it also responds to the latest prompt forgetting everything else in the conversation.

I don’t remember where I saw it, but I think there’s also a smaller agent paraphrasing my requests to the agent because I remember being angry and saying aiming sarcastic about an issue and the thinking notes said the user is angry do this blah blah, but then the Claude claim I asked for that specific solution.

I think they’re running some optimizations and multi level agents etc. tests to reduce costs.