Does telling AI who it is actually help?

"You are a brilliant copywriter with 50 years of experience." You have probably seen prompts that start like this. The idea is that telling the AI who it is makes the answer better. It is one of the most copied tricks around, and I used it for a while too.

It does not reliably work.

What the research found

In 2024, researchers tested exactly this. They ran thousands of factual questions through several families of language models, once with no persona and then with more than a hundred different roles added to the system prompt: teachers, lawyers, doctors, engineers and so on. Adding a persona did not improve performance compared with no persona at all. Some personas helped on some questions and hurt on others, and there was no dependable way to tell in advance which would do what. (Zheng and colleagues, "When 'A Helpful Assistant' Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models", Findings of EMNLP 2024.)

The years of experience are the least useful part. The model has no real idea what fifty years of copywriting would change. It is just more words in the prompt, and irrelevant words can pull the answer in odd directions.

What works better

I have worked as a translator since 1998. A good translation brief never says "pretend you are a brilliant translator". It says who the text is for, what it has to achieve, what the house style is, which terms must not change, and when it is due. The same brief works on AI.

• The task, plainly. "Write a 150-word introduction for our wedding photography page."
• The audience. "Couples in Cambridgeshire planning a wedding in the next year."
• The goal. "They should understand what is included and want to check our availability."
• The constraints. "British English. No superlatives. Mention that we cover Cambridge and the surrounding villages."
• An example of the tone you want, if you have one.
• The format. "Two short paragraphs, then one line inviting them to get in touch."

That prompt will beat "you are a legendary wedding copywriter" every time, because every line in it tells the model something it can use.

When a role still helps

A role can be a useful shorthand for a point of view: "review this contract as the customer would read it", or "read this page as someone who has never heard of us". That is not flattery, it is a real instruction about what to look for. Keep it short, keep it relevant to the task, and leave out the invented CV.

The short version

Tell the AI what you want, who it is for, and what good looks like. Skip the job title.