We looked at the 1,304 role lines Briefing Fox had written into its own briefs. One in four called the AI something like “an Elite Strategist” or “a World-Class Analyst”. “Elite” alone appeared in nearly one in five.
That was our tool, the one built to make AI instructions more precise, reaching for the same habit as everyone else. We traced it to the first line of our own instructions to the model, which called it an elite prompt engineer. It passed the compliment straight on. And it is a habit worth breaking, because a word like “elite” feels like an instruction and isn’t one.
What people write when they write their own
Most people never write a role line at all. Of 1,678 briefs submitted to Briefing Fox, 102 arrived with one already in the goal: “act as…”, “you are a…”, or the same in another language. Those came from 63 people, and they are the practised end of the user base. Their goals run to a median of 262 words, more than ten times the typical request.
Even there, the same pattern shows up. About one in six of those role lines leans on a status word: world-class, elite, award-winning, renowned. And fewer than one in ten says, inside the role line, who the role is working for.
So the people who care enough to write a role are mostly writing a job title with a compliment attached.
The delete test
Here is a test you can run on any role line in ten seconds. Delete the adjective and ask what the AI now knows less about.
“You are a world-class copywriter.” Delete “world-class.” The AI knows exactly as much as before: it is writing copy. The word described how good the copywriter is. It said nothing about what they know, who they write for, or what they would refuse to do. The model was always going to try its best. Telling it to be excellent does not tell it what excellent means here.
Now try it on a line that works. “You are a copywriter for a small accounting firm whose clients are nervous first-time business owners.” Delete “small accounting firm” and the AI loses its register. Delete “nervous first-time business owners” and it loses its reader. Every phrase you remove takes information with it. That is the difference between a role line and a compliment.
One role line, rebuilt three times
Take a request that comes up constantly: review a grant application before it goes out.
The compliment. “You are an expert grant writer.” True, flattering, and nearly empty. The AI will produce the review any competent reviewer would produce for any grant.
Add the level and the situation. “You are a grant writer who has sat on review panels for small community foundations, reviewing a first-time applicant’s proposal for a local literacy programme.” Now the AI knows which side of the table it is on, how big the funder is, and that the applicant is new, so basic structural problems are likely and worth pointing out plainly.
Add the standard. “You judge proposals the way a tired panel member does on the fortieth application of the day: if the need and the outcome are not clear in the first paragraph, say so before anything else. Do not soften the verdict.” This is the part almost nobody writes, and it is the only part that tells the AI what counts as a problem.
The final version never says “expert”. It does not need to. Every sentence is doing a job the adjective only pretended to do.
What a role line has to carry
A role line earns its place when it carries three things you could not delete without losing something:
- Where the role sits. Not the field in general but the seat: the reviewer, not the applicant; the in-house lawyer, not the litigator; the teacher marking the essay, not the student writing it.
- Who it is working for. The reader, client or audience, specific enough that the AI can guess what they already know and what will lose them.
- The standard it holds. One rule a real person in that seat would apply without being told. What they flag first, what they refuse to say, when they stop.
Level of experience belongs in there too, when it changes the judgement. “Senior” is information if it means the AI should skip the basics. “Elite” never is.
What we changed
The finding at the top of this piece was ours, so we fixed it before writing about it. As of 23 September 2026, Briefing Fox no longer tells the model it is elite, and every role line it writes has to pass the delete test: no status words, and it has to say where the role sits, who it works for and one standard it holds. A tool that asks you who you are, what you need and who it is for should not then hand the AI a line that says only how impressive it is.
Before your next AI task, write the role line, then run the delete test on every word. Keep the ones that take information with them when they go.
Try Briefing Fox free at briefingfox.com
It is the line that tells the AI who it is for this task, such as “You are a grant reviewer for a small community foundation.” It sets the perspective, the assumed audience and the standards the rest of the answer is held to.
Not on its own. Words like “expert”, “elite” or “world-class” describe how good the role is, not what it knows or who it serves, so they give the model nothing new to work with. Delete the adjective: if the AI knows no less without it, it was not doing anything.
Three things: where the role sits (the specific seat, not just the field), who it is working for, and one standard a real person in that seat would apply without being told. Add the level of experience only when it changes what the AI should assume.
Rarely. In 1,678 briefs submitted to Briefing Fox, 6.1% arrived with a role line the person had written themselves, and fewer than one in ten of those said who the role was working for.