Giving the AI a Role: When It Helps, When It's Pointless

"You are a marketing expert with 20 years of experience." That sentence opens a good share of the prompts floating around online, as if it unlocked a higher level. It unlocks nothing: the AI doesn't know one more thing after reading it. If it changes anything, it's elsewhere, and it's worth understanding where.

What a role really changes in the answer

A language model writes according to what your message makes likely. Mentioning a "doctor" or a "schoolteacher" therefore steers the vocabulary, the level of explanation, the topics considered important and the tone. A role works like an angle setting.

That angle can be valuable. Ask an AI to review your business plan "like a banker deciding whether to lend," and it will go look at cash flow and collateral, where a neutral request would have commented on the wording.

What a role doesn't change is what the model knows. Announcing that it's a lawyer doesn't give it access to current law, or a professional's accountability. It may even make its mistakes more convincing, because they arrive with the vocabulary and confidence of the profession. Any answer with something at stake still needs checking, whatever title you've given.

Where a role is pointless

You can skip the role in three common situations.

Mechanical tasks. To extract the dates from a document, sort a list or fix typos, "you are an expert administrative assistant" adds nothing. A clear statement is enough.

When your request already contains the angle. "Explain this contract in five sentences to someone who has never studied law" doesn't need "you are a gifted teacher." The reader and the format have done the work.

When the title replaces the information. A prompt that starts with four lines of proclaimed expertise, then asks "write a marketing plan" without saying anything about the company, will remain a bad prompt. A role never makes up for missing context.

Writing a role that works: the job and its criteria

A title alone says little. What matters is what this person looks at and what they reject. Compare these two prompt openings for a restaurant owner who wants to rework his menu.

Before:

You are a world-renowned restaurant expert. Give me your opinion on my menu.

After:

You advise independent restaurants with fewer than 40 seats. You judge a menu on three points: the length of the list of ingredients to keep in stock, the feasibility during service with [NUMBER] people in the kitchen, and how easy it is to read for a customer who chooses in two minutes. You'd rather say what to remove than what to add.

Here's my menu:
[PASTE THE MENU]

Customers: [DESCRIPTION]. Best-selling dishes: [LIST].

Rank your comments from most important to least important, seven at most. For each one, say which dish or section it concerns and what I could test. If a comment depends on information you don't have, say which.

The second version claims no renown. It describes a type of client, judging criteria and a preference, three things a title doesn't tell you. If you remove the first sentence "You advise...", the criteria keep working: they're what does the job, and the role mostly serves to group them together.

The roles that pay off most: the ones that contradict you

Left to itself, an AI tends to agree with you and find your text very good. A well-chosen role gives it permission to displease you, and that's when it does you the most good. There are several useful stances.

For a craftsman preparing an important quote, or a nonprofit submitting a grant application, this reversal brings more than a polite proofread. Here's the template you can adapt.

You're going to play [STANCE: for example, a panel member who awards grants to cultural nonprofits and has 40 applications to read today]. You're not here to encourage me.

Decision criteria: [PASTE THE CRITERIA OR DESCRIBE WHAT MATTERS].

Here's my document:
[PASTE THE DOCUMENT]

1. Give your first impression in two sentences.
2. List the passages that would make you hesitate or reject the application, quoting each passage.
3. Say what's missing against the criteria.
4. Finish with the three changes that would most alter your decision.

Don't invent any selection rule that isn't in the criteria provided. Don't comment on style unless it hurts comprehension.

Keep in mind that the reaction is still a simulation. It doesn't reflect what your real panel or your real customer will think, but it brings out questions you weren't asking yourself.

Does it have to come first?

The form matters little. "You are...", "Act as...", "Answer from the point of view of..." or a simple note at the end of the message generally lead to the same thing. Putting the role first is convenient, because it frames the reading of everything that follows, but it's not a rule.

If you reuse the same role every week, there's no need to rewrite it. Save it in your tool's custom instructions or file it in your prompt library.

To decide whether a role is useful to your prompt, remove it and rerun the prompt in a new conversation. If the answer loses nothing, you can delete it for good. If it loses something, note exactly what and write it directly as a criterion.