RTF
Role · Task · Format — The three-line starter. Say who the model is, what it must do, and what shape the answer takes.
Use it for
Everyday one-off requests where you already know what a good answer looks like: a summary, a rewrite, a short list, a first draft.
Avoid it for
Anything where the model needs background it cannot guess. RTF has no slot for context or source material, so it quietly invites the model to invent the missing pieces.
The slots
Each letter is a question the prompt has to answer.Role
Who should the model be?
Name a perspective with real expertise attached. A role narrows vocabulary and priorities — it does not add knowledge the model lacks.
Example: A technical writer who documents developer APIs for a mid-sized SaaS company.
Task
What exactly must it produce?
Start with a verb and name the deliverable. "Rewrite", "compare", "draft", "extract" beat "help me with".
Example: Rewrite the error-message reference below so each entry states the cause and the fix in plain language.
Format
What shape should the answer take?
Be mechanical: a table with named columns, Markdown, JSON with named keys, a word count. Vague format instructions produce vague structure.
Example: A Markdown table with columns: Code, What happened, What to do. Maximum 25 words per cell.
Strengths
- Memorable enough to use without looking it up.
- Format is a first-class slot, which is where most vague answers actually go wrong.
- Fast to iterate: you can rewrite the whole prompt in seconds.
Weaknesses
- No context slot, so the model fills gaps with plausible invention.
- No place to attach source documents or data.
- No success criteria, so you cannot tell the model what "done" means.
Sources and grounding
- 1Prompt engineering — best practicesOpenAI (Developer docs)Vendor grounding for the role and output-format components.
- 2Prompting best practices — Give Claude a roleAnthropic (Claude Platform docs)Anthropic documents role-setting and explicit output format as distinct techniques.