You will be able to use role and audience instructions for tone and assumptions without expecting extra knowledge.
Search for prompting tips and you will soon find advice to start every request with a grand title. Act as a world-class marketing strategist. You are a senior lawyer with thirty years of experience. You are the best copywriter in Asia. The idea is that the title summons expertise the model would otherwise hold back.
Daniel, the furniture shop's customer service lead from lessons 2.1 and 2.2, tried it. A customer wanted to return a sofa after two months because one cushion had sagged, and Daniel wanted to know where the shop stood. He typed "Act as a Singapore consumer lawyer" in front of his question. The answer sounded more confident and used more legal terms. Whether it was more accurate, he had no way to tell from the answer itself, and that turns out to be the whole problem.
A role instruction does change the output, and it is worth knowing how. Telling a model to act as an HR manager shifts the vocabulary towards HR terms, makes the tone more measured, and changes what the model assumes you want: probably a policy-minded answer rather than a personal one. Act as a primary school teacher and the explanations get simpler and include more examples. Act as a sceptical investor and the model looks for weak points.
So a role is a quick way to set tone and assumptions. It is a shorthand for a bundle of choices, in the same way that the tone reference points in lesson 1.3, Set constraints and ask for the format you need, were.
A role does not give the model any knowledge it did not already have. AI fundamentals lesson 4.1, Pretraining: learning to predict text at huge scale, explained that a model's knowledge comes from its training. Typing "you are a senior lawyer" does not add case law, statutes or recent amendments to it. Whatever the model knew about Singapore consumer law before the role, it knows the same amount after.
What the role can do is make a weak answer sound like a strong one. An expert's voice comes with confident phrasing, specialist terms and fewer hedges, and a model playing an expert produces all three whether or not the substance underneath is right. For a question like Daniel's, that is a step in the wrong direction, because it makes an error harder to spot. AI fundamentals lesson 6.2, Where invented answers are most likely, lists niche and local legal detail among the riskiest areas.
The honest way to get a more accurate answer about the law is to give the model the source, for example by pasting the relevant section from Singapore Statutes Online or the shop's own returns policy, and asking it to explain what the text says. Then check anything important with someone qualified. Module 5 covers working from documents.
In practice, describing who the answer is for usually helps more than describing who is supposed to be writing it. "Explain this for a first-time investor in their twenties who has some savings but has never bought a fund" tells the model the starting knowledge, the vocabulary to avoid and the questions the reader is likely to have. "Act as a top financial adviser" tells it none of that.
Daniel rewrote his question with an audience in mind: "I run customer service for a small online furniture shop and have no legal background. Explain, in plain English, what I should find out before deciding whether to accept a return for a sofa bought two months ago with a sagging cushion. List the questions I should answer and the documents I should check." The answer was less grand and far more useful. It pointed him to his own returns policy, the product warranty and the date of purchase, and it suggested he check the rules from an official source instead of relying on the answer.
The other problem with grand titles is that they take the place of instructions that would actually help. "World-class consultant" does not say what to cover. "Cover the three biggest risks, what each would cost us, and one thing we could do this month about each" does.
If you like using roles, keep them modest and pair them with specifics. "You are reviewing this as a finance manager who will sign off the budget. Point out any line that is not justified and any cost that seems to be missing" is a role doing its proper job, which is setting the angle, followed by instructions doing theirs.
A fair test settles the question for your own work better than any tip. Pick one question you genuinely need answered, ideally one where you can judge the quality of the answer yourself, and get ready to ask it in three different ways.
Run one question three ways, with no role, with an expert role and with a detailed audience description, and note which version gave the most useful answer.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).