Feels a bit like an extension of having an LLM make a JSON file for text/numbers/dimensions in an image, allowing one to edit it, then use it as a reference when a prompt re-works it --- just one has to find the right node to update/edit.