It could be useful to run a prompt/test once, get the xPaths, and rerun it deterministically. When it breaks, you know something is wrong, and the LLM could be used as a fallback to fix the script.