·Question
ChatGPT keeps changing the tone of my claim letters even though the prompt has not changed.
I write the letters that tell a customer what has happened to their motor claim. I keep one prompt in a note, with our tone rules and one example letter under it, and I paste the whole thing into a brand new chat each time along with the claim notes. Monday it came back exactly right: "We have reviewed your claim and approved the repair costs." Wednesday, same prompt, same kind of claim, and it opened with "We deeply value your patience during this difficult time and want to assure you", which is three sentences before it says anything at all. Is this something I am doing wrong in the prompt, or is that just how it works?
This answered it
Two changes fixed it for our team. First, list the phrases you do not want rather than the tone you do want. Ours says: no opening thanks, never "we value your patience", never "rest assured", first sentence states the decision on the claim. Second, put the rules after the claim notes rather than before, because whatever sits closest to the end gets followed hardest. "Professional and warm" means nothing to it. A banned phrase means something.
Are you starting a new chat each time, or carrying on in the same one? If you keep replying in one thread, ChatGPT treats everything above as an example of what you want, including the drafts you told it were too flowery. New chat per letter, prompt pasted in fresh.
That was mine exactly. I was replying "plainer please" four or five times per letter and staying in the same chat all day, so by the afternoon it had learned that flowery was the house style.
Give it two example letters instead of one. With one it copies that letter's rhythm. With two it has to work out what they have in common, which is the closest thing to it actually understanding your tone.
Did both. Four letters this week went out with no edits at all. My list of banned phrases is now longer than the prompt itself, which felt wrong until I saw the results.