Role, Source, Task, Format, Fence: The Pattern Behind Every Great AI Prompt

People think prompting is a bag of secret magic words. It is not. Every strong prompt I run has the same five parts under the hood. Learn them and you stop hunting for the perfect prompt online, because you can build any of them yourself.

Here they are. Role, Source, Task, Format, Fence.

Role

Tell the AI who to be. "You are an elite sales coach." "You are a skeptical CFO." "You are a documentary story editor." The role sets the lens. The same transcript read by a sales coach and read by a therapist gives you two completely different gold mines. Most people skip this and get a generic answer, then blame the tool. You did not tell it who it was.

Source

Give it the material, clearly marked. "Here is the transcript of my call, between the triple dashes." The point is to separate the words it should analyze from the words that are your instructions, so it never confuses the two. Mark your source and the AI stops guessing about what is data and what is a command.

Task

Say exactly what you want done. Vague in, vague out. "Summarize this" is weak. "Pull every objection the prospect raised, the exact words they used, and whether I handled it" is strong. The more specific the task, the more the answer feels like it was written by someone who was paying close attention, because it was.

Format

Tell it the shape of the answer. A table. Five bullet points. A two-paragraph email in my voice. A score out of ten with three fixes. If you do not specify the shape, you get an essay, and you wanted a table. This one step is the difference between output you can use immediately and output you have to reformat by hand.

Fence

Set the guardrails. "Use only what is actually in the transcript. Do not invent anything. If something is not there, say so." An AI wants to please you, so left alone it will smooth over gaps and hand you a plausible detail that never happened. The fence is how you make it tell the truth. It matters more than any other part, which is why it gets its own post.

You will not name them while you work

Role, Source, Task, Format, Fence. You will not think about these five while you write a prompt, the same way you do not think about grammar while you talk. But here is the useful part: when a prompt gives you mush, it is always because one of these five was missing. No role, no lens. No source markers, confusion. Vague task, vague answer. No format, an essay. No fence, a flattering lie.

So when an answer disappoints you, do not start over. Walk the five. Find the one you left out. Add it back, and the mush turns to gold.

And when you cannot think of a prompt at all

Let the AI write it for you. Paste your material and say: "You are an expert prompt engineer. Here is a transcript of my sales call. Write me the five best prompts I could run on this, and tell me what each one gives me." It hands you a menu. You pick one and run it. When you do not know the question, ask the machine for the question, then go get the answer.

This is the engine behind all 101 plays in my forthcoming book Transcript Alchemy, coming early July. One skill, a hundred and one variations. Learn the five parts and every play in the book is just you, turning a dial.