Your Job Description is Too Generic: How to Tell AI Who You Really Are

The AI Template Trap: Why Generic Roles Fail

[0:00]Hi, this is Tanmay Kulkarni from Humanitarians AI. The hard part of using AI isn't saying what you want. It's saying who you are. Every template you'll ever pick up was written for a generic professional, and you're not one. So, this video is about a single skill, describing your own job clearly enough that the answer you get back is actually about your job. Here's what we'll cover. First, four questions you can run on any description you write. Then, what the research actually says about telling AI who you are, which is less than you'd hope. Then, a test I ran 408 times because one of those four questions turned out to matter far more than the other three. Open most AI exercises and you'll find the same shape, a paragraph with blanks in it. Your role goes here, your task goes here, your limits, your format, the people who depend on the answer. Those

The 4-Question Framework for Specificity

[0:45]Blanks are honest. Whoever wrote the exercise can't know who you are, so they leave you room to say it. The interesting question is what you put in them. Watch what most people write in that first blank, healthcare professional, software engineer, financial analyst. Those aren't jobs. They are categories, and a category gets you a category-shaped answer. So, here are the four questions. Run them in order on any description you write. First, does it rule anyone out, or does it just describe you? Most of what people write is true of them and just as true of the person at the next desk. Describing yourself is easy. Ruling other people out is the work. Second, who has the closest job to yours? Not someone in another department, the person whose work is most like yours. Pick someone far away and you'll pass without learning anything. Third, which exact words rule that person out? Point at them. If you can't point at them, your description isn't specific, it's just long. Fourth,

Case Study: The Night Pharmacist & The Beers Criteria

[1:31]Would someone else agree with you? If you're the only person who can run your own test and get your own answer, that isn't a test, it's an opinion. The order matters. The first two are quick, and they catch most weak descriptions. The third is the one people skip. And the fourth turned out to matter far more than I expected. Let's actually run it. Here's a description, a hospital pharmacist working nights who checks discharge medicines against a safety list called the Beers Criteria before the patient goes home. Question one, does it rule anyone out? It definitely describes somebody. Whether it rules anyone out is what we don't know yet, and that's the whole question. Question two, who has the closest job? Not a radiologist, not a hospital manager, the night nurse on the same ward. Same hours, same patients, same discharge. She's who we test against. Question three, which words could that nurse not write? Working nights, she works nights, too. Before the patient goes home, she's standing right there for it. But checking

"Expert Persona" vs. "Situation": What the Research Says

[2:20]Medicines against the Beers Criteria, that's the pharmacist's expertise. Those are the words doing the work. Good, except now question four, would someone else agree? So, I looked it up. In the Beers Criteria, the actual document was written by a panel of doctors, pharmacists, and nurses. A nursing institute publishes it in a series made for nurses, described as a tool to help nurses review medicines. She could write those words. She probably has. My example fails. The line I was confident about, the one I'd have put on a slide, rules nobody out. And I only found that out because question four made me check something other than my own opinion. If you want words that really do rule her out, it isn't the safety list. It's who signs off releasing the prescription in the pharmacy system before it can be handed over. She checks it, he releases it. That's the line. Notice how that went. Questions one to three were me marking my own homework, and I passed. Question four was the only one that could catch me. And it's the one most advice quietly skips. Here's what I'd been missing, and it's why this is harder than it looks. There

We Ran the Test 480 Times: The AI Consistency Problem

[3:15]Are two completely different things people mean by telling AI who you are, and nobody separates them. The first is a claim about the AI. You are a senior hospital pharmacist with 15 years of experience. You're telling it who to pretend to be. The second describes your situation, night shift, discharge medicines before the patient goes home. You're not claiming anything about the AI. You're telling it where you're standing. Same trick, same name in every guide you'll read. Two completely different things, and only one of them has actually been tested. First kind, the you are an expert kind, has been tested, and the results aren't flattering. One team gave an AI expert personalities, and then ran it through a general knowledge test. With no personality, It scored 71.6% at a short expert personality, 68. At a long detailed one, 66.3. Their words, not mine. Longer personality prompts do

Your Turn: The 4-Step Specificity Challenge

[4:01]More damage. So, the thing everyone does, piling on job titles, years of experience, specialties, cost accuracy on that test. And the more they piled on, the more it cost. This is a pre-print, which means other researchers haven't checked it yet. I'm telling you that because it matters. The second group asked a narrower question. When does it help? They found a trade, not an improvement. Telling the AI who to be made answers deeper. It also made them less clear. Where it helped, questions asking for advice in medicine and psychology. Where plain asking one, questions asking it to explain something in finance, law, science, and technology. Also a pre-print. A made-up test set, only a few AI models, and the marking was done by another AI. The authors say so themselves. So, I went looking for study on the second kind, the situation kind, the one the exercise actually asks you to write.

I couldn't find one. And that second paper measures every expert personality as one thing. It doesn't ask whether the answer changed because the AI was told it was an expert or because the words described the job differently. Every test I could find is of the first kind, which means the advice isn't wrong. It just hasn't been tested yet. Those are very different things, and the difference is good news because untested is something any of us can go and fix. Which leaves you where it left me. The exercise gives you a sensible instruction. Write a description of your job, hand it to an AI, and ask whether it's specific enough. And that's a genuinely useful habit. But specific enough is exactly the kind of judgment where an AI can sound completely certain and be reading something else entirely. So, how much weight should you put on its answer? Both easy answers are wrong.

Trust it completely. And you're trusting something nobody has tested. Refuse to use it, and you've thrown away a second pair of eyes that costs you nothing. I didn't want to guess, so I tested it. So, here's what I did and why. If the advice is to hand your description to an AI and let it name the closest job, then the whole thing rests on the AI giving you a steady answer. So, that's what I measured. I wrote 24 job descriptions across healthcare, finance, and engineering. Some deliberately sharp, some deliberately vague. Then I asked two AI models the same question about each one five separate times. 480 answers in total. Here's the first thing that came back. Ask an AI who has the closest job to yours five times over and you get roughly two and a half different answers. 79% of descriptions got more than one answer from the first model,

71% from the second. Not different wording, different jobs. And because the pass or fail depends on which job it picked, that moves, too. On the first model, 42% of descriptions got both a pass and a fail. Same AI, same question, nothing changed but the attempt. Then the useful part. Stop asking the AI to pick the closest job, tell it name that person yourself and the flip-flopping drops by about two and a half times. 42% down to 17, 17 down to eight. The two models agreeing with each other goes from 67% up to 88. One sentence of change, a steadier test. Now the limits, because they matter. 24 descriptions, two models, one task, and I wrote the descriptions myself. This doesn't show AI is bad at this. It shows the answer moves and that where it moves is something you control. So, the advice was right. Tighten the description until it fits one

Person. Keep doing that, but it's worth being clear about why it works. Not because being specific is obviously better and careful here, not because of those accuracy numbers, either. Those were measured on the other kind, the you are an expert kind. Borrowing them would be me making the exact mistake I've just spent four minutes describing. Tighten it for the reason question three gives you. Words you can't point at aren't doing any work. And run the closest job test because it steadies things as long as you pick that person instead of asking an AI to pick for you. That part I did test, so the exercise gives you good practice. Where it could be stronger is spelling out the reason and that's something to add, not something to replace. It's exactly what happened to my own pharmacist example four minutes ago. Your turn and this is four steps, not something to paste.

First, write your job description. Second, name the person whose job is closest to yours, closest not furthest. If you're reaching for someone in another department, you've picked wrong. Third, underline the one phrase that person could not write. Fourth, and this is the one, delete that phrase and read it again. If it still sounds fine, it was never specific. It was just long. You've done it right if you find a description you were happy with and can't point to the words doing the work. You've done it wrong if everything passes first try. That means you marked yourself generously or you picked someone who was never really close. And if you do hand it to an AI, tell it who that person is. Don't ask. Four questions, the fourth is the one that catches you because it's the only one that isn't you marking your own homework. This is Tanmay Kulkarni from Humanitarian's AI.