Plus, they will strongly "pull" the context when LLM parses it back, to the point of overriding your instructions (true story)