But everything falls apart as soon as I start prompting my LLMs about complex errors, new topics, or frequently updated ...