Role-setting prompts solve a common problem: the same request gets a completely different answer depending on the persona behind it. Asking "help me revise this text" with no persona gets you a generic revision; telling the AI first that "you are a Chinese-language editor with ten years of experience" makes it judge the text by an editor's standards and offer real fixes. Once the persona is set, later instructions are far less likely to drift.
Copy this general template first
You are now a [role], with [professional background].
Your tone is [tone], and you work by [working standard].
For anything I send you, first point out [aspects to check], then give me [desired output format].
Do not [things to avoid]. If my request is unclear, ask me the single most important question first, then start.Send this to the AI as it is, replace the bracketed parts with your own details, then paste the material you want processed, and it is ready to use.
How to fill in the 5 placeholders
| Placeholder | Fill it like this | Not like this | | Role | Be specific: "primary-school maths teacher", not just "teacher" | Vague words like "expert" or "assistant" | | Professional background | State why it knows the field, e.g. ten years of editing experience, or having taught graduating classes | A pile of titles with no real experience | | Tone | Pick one style and state it: direct, gentle, or conclusion first | Demanding brevity and detail at the same time | | Working standard | Point out problems first, then fixes, with one example per suggestion | Just "do it well", with no standard | | Things to avoid | Write the bans in: do not change the meaning, invent data, or pad with empty phrases | No bans at all, leaving it to improvise |
Two ready-to-use filled examples
Editor version, for revising a draft:
You are now a Chinese-language editor with ten years of experience, having edited newspaper supplements and long-form social media articles.
Your tone is direct, with no small talk, and you work by "point out the problem first, then give the fix, with one revised sentence per fix".
For anything I send you, first point out wordiness, jumps in logic, and imprecise wording, then give me a revised version.
Do not change the original meaning, and do not invent facts. If my request is unclear, ask me the single most important question first, then start.Study-coach version, for asking questions:
You are now a patient high-school maths tutor, good at breaking hard problems into small steps.
Your tone is gentle, one step at a time, and you work by "first ask me which step I am stuck on, then explain that step, never the whole problem at once".
For anything I send you, first point out which step I got wrong, then show me how to think about the next step.
Do not give the answer directly, and do not explain too much at once. If my request is unclear, ask me the single most important question first, then start.How to fix a drifting persona
When the answer drifts, do not rewrite the whole template; three fixes cover most cases. First, tighten it in one sentence: "Back to your editor persona: revise only this paragraph, no general advice." Second, add a ban: a recurring problem written into the "do not..." line works better than extra praise. Third, restate it mid-conversation: after a dozen or so turns the persona fades, so sending the template's first sentence again re-establishes it.
One more rule: one persona per conversation. Asking it to be both an editor and a lawyer in the same chat makes the two standards collide and the output turns into neither. When the task changes, start a new conversation and paste the template again; that is far less trouble than switching personas back and forth in one chat.
