Which tool to choose
The technique works on any text-based AI assistant: you don't need a special tool. If the tool offers personal instructions (sometimes called "custom instructions" or "personality"), it's worth putting the recurring role there once, so you don't rewrite it for every chat. If you can't find that section, write the role at the start of the message: the result is identical, the only difference is that you repeat it each time. For the roles you use often, save the text in a file and paste it.
How to do it
- Choose a specific role, not "expert". "Marketing expert" is vague. "Marketing manager of a small shoe e-commerce, mindful of the budget" steers examples, tone, and priorities.
- State the task and the recipient. Who the answer is for changes everything: a text for a client is not a text for the team.
- Add the constraints that cut errors. They are the part most people forget. The operational syntax:
Act as an expert financial-statement auditor.
Work only on the data I provide: if a piece of information is missing, ask me for it instead of inventing it.
Don't add items I haven't indicated.
Flag every number that doesn't add up and explain why.
Answer in Italian, in no more than ten lines.
- Specify what it must NOT do. Negative constraints ("don't invent", "don't add", "don't go off topic") are often more powerful than positive ones.
- Refine after the first answer. If the tone or the angle is off, correct the role: "You're too technical, speak as you would explain to a client without accounting skills".
A concrete example
You have to reply to a customer's negative review. By asking "write me a reply", the AI produces a generic, lukewarm text. By assigning the role and the constraints — "Act as the owner of a neighborhood restaurant who cares about reputation and speaks warmly but professionally. Reply to the review below. Acknowledge the specific problem mentioned, don't promise refunds, invite the person to come back. Maximum six lines, no clichés like 'we're sorry for the inconvenience'" — the answer changes face: it touches the real point, stays within the limits you gave, sounds like a person and not like a form. The difference is not the AI's skill, it's the precision of the instructions.
When it does NOT work (and how to fix it)
If the AI ignores a constraint halfway through the answer
On long answers, the limits given at the start sometimes get lost along the way. Fix: repeat the most important constraint at the end of the prompt, or break it into several steps ("first do X respecting these limits, then I'll ask you for another").
If the role makes the text bloated and theatrical
A role that's too loaded makes the AI write emphatically. Fix: add a register constraint: "keep a plain tone, no emphatic adjectives, get straight to the point".
If you ask for a technical role but the answers stay superficial
The role alone doesn't create competence the model doesn't have. Fix: provide the reference material yourself (data, documents, examples) in the message. The role steers how the AI uses the information, it doesn't replace it.
A tip from someone who really uses it
The piece that makes the difference is not the role, it's the negative constraints. "Act as a lawyer" improves little; "act as a lawyer, don't give definitive advice, flag where a professional is needed, don't cite laws unless you're certain of the reference" changes the quality of the answer. Always write what the AI must not do: that's where the inventions and digressions stop.
Frequently asked questions
Do I have to write a very long role to get better answers?
No: better short and specific than long and generic. Three elements are almost always enough: who the AI is (concrete role), who it's writing for (recipient), what it must not do (constraints). Adding paragraphs of decorative personality doesn't help and sometimes confuses.
Can I give the AI the role of a famous real person?
You can ask it to adopt a generic style ("explain simply as a good communicator would"), but imitating a specific real person gives unreliable results and raises image and rights issues. Better to describe the qualities you want (clear, patient, concrete) than someone's name.
Does giving the AI a role really make it more competent?
No, and believing it is the misconception to avoid. The role doesn't add knowledge to the model: it steers how it uses what it already has, choosing suitable vocabulary, examples, and priorities. If the competence isn't in the model's data, no role invents it: for delicate or niche topics, provide the sources yourself and verify the result.