Stop telling AI who to be. Tell it what you're trying to do.

Prompt framing is the part most people skip. They dress the model up instead of saying the job out loud.

"You are an expert in marketing."
"Act as a senior copywriter."
"You are a world-class strategist."

That feels like progress. It isn't. That's a costume. The model already knows how to sound like those people. What it still doesn't know is what you are trying to get done.

I call the fix framing. One plain line near the top:

Here's what I'm actually trying to accomplish.

Not the tool. Not the persona. The finish line.

"Write an email" is a task.
"Here's what I'm actually trying to accomplish: get a nervous first-time patient to show up Thursday at 2 without sounding salesy" is a job.

Same request on the surface. Totally different answer.

People reach for roles because courses teach roles. "You are an expert in…" is about as far as a lot of training goes. No shame. I used that too when I started. It just doesn't remove the guesswork. It only changes the accent on the guess.

Framing does something else. It tells the model which outcome wins if two good answers collide. Length, tone, and format matter. The why matters more. Without it, the model optimizes for "sounds smart." With it, the model optimizes for "helps you ship."

Try this on your next real request. Before the bullets. Before the constraints. Before you paste the doc.

Write one sentence:

"Here's what I'm actually trying to accomplish: ____."

Make it concrete enough that a stranger could judge the answer. If you can't fill that blank, the model can't either. It will invent a destination. Sometimes a nice one. Rarely yours.

Dinner with a friend works the same way. "Make something nice" only works because they already know you. The model doesn't know you. It has the words.

So give it the destination first. Then add context, intention, constraints, and the missing details. Framing is the north star those four hang under.

Once you stop asking AI to figure out who to pretend to be, and start telling it what you're trying to do, the output finally starts looking like the thing you already had in mind.