Here is the failure this fixes, and you have almost certainly lived it.
You ask for something. What comes back is not wrong. It is accurate, fluent, reasonable. And then you spend the next four messages on it: shorter, make that a table, drop the intro, the audience is engineers, not executives. By the time it is usable you have typed more in the corrections than you did in the original request.
Every one of those corrections was information you had at the start and did not supply. Not because you were lazy, but because there was no obvious place to put it. A blank chat box does not prompt you for anything.
That is what RTF is: a place to put it.
Role, Task, Format. Who is answering, what exactly they are producing, and what it has to look like when it lands.
Three lines, always in that order, and the ambiguity that generated your four follow-ups is gone before the model writes a word.
What each slot is actually doing
- Role fixes the vantage point. Not the model's credentials, the model's seat in the room. A security engineer and a support lead describe the same outage differently because they notice different things.
- Task fixes the deliverable. A verb, an object, and the constraints that change the work. "Summarise this" is a wish; "extract the five decisions made and who owns each one" is a task.
- Format fixes the shape. Table, numbered steps, email, six bullets, 150 words. This is the slot that turns "nearly right" into "paste it in".
It is model-agnostic, which matters more than it sounds. The same three lines behave the same way on Claude, ChatGPT, Gemini and Copilot, because none of this is a trick that depends on how a particular model was tuned. You are not exploiting anything. You are just no longer making it guess.