That's correct. The flow is 1) user requests some change e.g. "change to dark mode", 2) a snapshot of the page is sent to an LLM, 3) the LLM generates and retur

That's correct. The flow is 1) user requests some change e.g. "change to dark mode", 2) a snapshot of the page is sent to an LLM, 3) the LLM generates and returns a deterministic script that handles the page editing. And just to further clarify: "each use" means each generation. Applying the modification after generation doesn't cost tokens