S-Tier Prompt Writing

You wouldn't hand a stranger a stack of work and say "do something."

You'd tell them who they are first. The accountant. The editor. The coach. The job comes with a way of thinking, a set of standards, a thing they care about.

AI is the same. And by default, you've hired no one.

With no role, you get the average of everyone

When you ask a question cold, the AI answers as nobody in particular. It pulls a little from the textbook, a little from the forum post, a little from the blog written by someone who half-understood the topic. Then it blends it all into one smooth, safe, forgettable answer.

That blend is the average. And average is exactly what you don't want.

The fastest way to a better answer isn't a longer question. It's telling the AI who's answering.

The role is who, not what

This is the part people skip.

The role is the identity. Who the AI is. It's not the request you're about to make. "You are a writing teacher" is the role. "Fix my essay intro" is the work you'll hand that teacher in a second.

This lesson is only about the who. Get that part right and everything you ask afterward lands differently.

Specific beats vague, every time

"You are a teacher" works. Barely.

Watch what happens as you sharpen it:

Each line pulls a more specific version. The last one doesn't just set a subject. It sets a standard. "Getting struggling students to get it" tells the AI what good looks like before you've asked a single thing.

Details that change the answer are worth adding. Details that don't are just noise. (The teacher's eye color won't change your chemistry explanation. Their fifteen years in the classroom might.)

Yes, it's pretending. That's the point.

The AI is not really a teacher. It's a model predicting text. Telling it to "act as" a teacher is pretend, and that's completely fine.

The pretend works because the AI has read how teachers write. When you ask it to play one, it reaches for that pattern. Patient. Structured. Explaining the why. Checking you followed. You get teacher-shaped output. That's the whole trick.

But it's not a magic wand

A role changes who's answering. It doesn't change what they can actually know or do.

Telling the AI to "act as" something doesn't hand it powers it never had. It pulls from what's already inside the model. It can't conjure what isn't there.

So a few roles that quietly fail:

• Pretend to be my colleague Ana from the office. The AI has never met Ana. It'll happily invent her, then put words in her mouth. Confident, and completely made up.
• You are a superintelligence smarter than any human. Saying it doesn't make it so. It's the same model with a fancier label.
• Make no mistakes. It can't promise that, and the role doesn't unlock perfection. It just sounds more sure while being exactly as wrong.

The role decides who answers. It doesn't bend the laws of physics, invent knowledge that was never in the training, or reach information you didn't actually give it.

So when the AI is missing something, the fix isn't a fancier job title. It's handing it the missing piece. (That's a later layer.)

See it in action

Now it answers like someone who has seen the mistakes. It warns you off the "Since the beginning of time" opener. It tells you what actually loses marks. Same question. Sharper answer. Because someone specific is answering.

The takeaway

Before you ask the AI to do anything, tell it who it is.

Pick the role whose actual job is to produce what you want. Make it specific enough to set a standard. Then hand over the work.

That's the foundation. Next, we'll meet the four roles that do most of the heavy lifting: the Expert, the Teacher, the Critic, and the Creator. And when to reach for each.