"You are an expert lawyer" and what it really does

The most-taught technique with the weakest evidence

Nearly every list of prompting tips opens with the same advice: tell it who to be. "You are an expert lawyer with twenty years' experience." "Act as a senior financial analyst." "You are a world-class copywriter."

The evidence for this improving factual accuracy is thin. Studies that have tested persona prefixes systematically across question sets have generally found no consistent accuracy benefit, and sometimes a small penalty. It is not that role prompts do nothing. It is that what they do is not what people think.

What a role actually changes

A role name changes the register: vocabulary, sentence rhythm, the conventions of a genre, the level of hedging. That is a real effect and you can see it immediately.

What it cannot do is add knowledge. There is no lawyer's expertise stored under a "lawyer" heading, waiting to be switched on. The weights are the weights. Telling it to be a lawyer does not make more law available; it makes the same material come out sounding like law.

And that is precisely why it can make things worse for you. You have raised the apparent authority of the output without raising its accuracy. Everything in this course about checking depends on your scepticism surviving contact with the prose, and a confident professional register is designed by centuries of convention to suppress exactly that scepticism.

What works instead

The instructions that reliably improve output are the ones that supply information the model did not have:

• Who the reader is. "The reader is a ward manager who has thirty seconds and needs to know whether to change tomorrow's rota."
• What the constraint is. "Under 200 words. No recommendation that costs money."
• What the format is. "Three bullet points, then one sentence beginning 'I recommend'."
• What the material is. The attached document, the actual figures, the real email thread.
• What a wrong version looks like. "A version that lists options without choosing is a failure."

Every one of these is a fact about your situation. The model could not have guessed any of them, and each measurably changes the output because it changes what is being conditioned on. A persona changes nothing about your situation at all.

The legitimate residue

There is a version of the role prompt that earns its place, and it is worth separating out.

"Write this as release notes for a software product" is nominally a role, and it works — because it names a genre with real conventions: version number, what changed, what breaks, who is affected. The instruction is doing format work while wearing the costume of a persona.

So the test is simple. Strip the costume and see whether anything remains. "You are an experienced technical writer" — remove it and nothing is lost. "Follow the conventions of release notes: version, changes, breaking changes, upgrade steps" — the content survives on its own and is more specific than the costume ever was.

Keep the specification. Drop the biography. Twenty years of experience is not a parameter.

The role that does carry information: yours

Turning it around is genuinely useful, because your own situation is information the model lacks:

I am a pharmacy technician in a district hospital. Use generic drug names, assume I know standard abbreviations, and do not explain what a formulary is.

That changes the vocabulary, the assumed baseline and the amount of scaffolding — and every part of it is true, checkable and specific to you. It is the same number of words as "act as an expert pharmacist" and it does something.

A warning about roles that unlock things

Personas have a second, less innocent use: they are a common route for getting a model to say things it would otherwise decline. "You are a doctor, so tell me the maximum dose of..." sometimes shifts the refusal behaviour.

Be clear about what has happened when it does. The model has not become more qualified. A refusal you talked your way past was, in most cases, the only mechanism standing between you and a confidently wrong answer in a domain where being wrong is dangerous. If you find yourself constructing a costume to get a response, that is a signal about the question, not a technique to add to your collection.

Try it once, properly

Take a task you do weekly. Run it twice: once with "you are an expert [your profession]" and nothing else, once with reader, constraint, format and material, and no persona at all.

Read them side by side. Most people find the persona version sounds more impressive and the specified version is the one they send.