LLMs are compelled to refer back to their prompt. When asked to build a screen with X, Y, and Z, they will include as many references to X, Y, and Z as possible, both in controls and in text.
This was less of a problem in earlier models. Newer models are rewarded for intuiting intent. They deduce what you want and go beyond it instead of doing exactly what’s asked, since that produces more detailed artifacts. However, because they still need to abide by the prompt and avoid unreasonably steering off-course, they are brought back to the query as much as possible.
This results in the weird eyebrow text and the unnecessary marketing taglines that we find below headings. The LLM thinks “Hm, I need to be detailed here” when going beyond the prompt, but then it has no other material than what’s in the prompt. So it leaves “prompt detritus” around, in this case “One instance. The world, settings, and your memories”, since that’s what it was asked to include.
When models “re-prompt” themselves by looking at what they’ve produced to move on to another step, the same detritus can occur, usually in the form of weird step-by-step instructions in paragraphs: “First, start with X. Then do Y, and finish with Z.” This is also detritus that comes from the LLM building an artifact, deciding that it has to include text to flavour it, but having nothing to base itself on other than the very artifacts it already made. And so it describes them in order.