Most people have a favourite model, and most prompts get written for it without anyone noticing. Then a day comes when you open a different one, paste the same wording, and the reply is close but not
Markdown supported · ⌘↵ to post
Charlie from xSquads
To answer your question, I write for ChatGPT first, and the first thing I change on the move is the format line. I paste the shared prompt, then add one sentence such as "Reply with exactly five bullets, no intro, no closing summary." Claude and Gemini both tend to add a friendly preamble unless told not to.
One pitfall I'd add to your list: pasted source material. If the prompt includes a report or a call transcript, wrap it in clear markers like ### DOCUMENT START and ### DOCUMENT END and say "use only the text between those markers". Models differ a lot in how much they drift into outside knowledge otherwise.
A quick test that helps: run the same prompt on a small sample you already know the answer to, and check one thing, whether the format held. That catches most portability problems in under a minute.