Parse Marina's message. Return JSON {"sees":[...],"leaked":[...],"safe_rewrite":"..."}: sees - what the model actually sees from the message; leaked - what sensitive data the person accidentally disclosed; safe_rewrite - the same request, rewritten safely (with placeholders or generalization), but preserving the essence of the task. Return only JSON, no text around it.
When you send a request, the model receives ALL the text in its entirety - as a single input, without dividing into “important” and “background”. She doesn’t distinguish your question from randomly inserted data: name, phone number, order number, piece of correspondence - for her it’s the same material. Analogy: you dictate a letter to a secretary - he will write down every word, even what you mutter under your breath. The mechanics are simpler than it seems: there is no “invisible” text. Even a picture or screenshot is readable - the model recognizes the text in the image, sees signatures and amounts. Insider: the real data in the “example for context” is leaked in the same way as in the main question - the phrase “this is just an example” does not protect anything. Second trick: before sending, re-read the message through the eyes of a stranger - whatever you would NOT want to show your neighbor. A typical beginner mistake: copying entire correspondence or the entire document “so that the AI can figure it out,” although the task requires two lines.