Maybe the LLM should be trained to interface with text using some ed/vim dialect to only work on blocks of text.