RTF, CO-STAR & RISEN: Prompt Frameworks Explained
A prompt framework is a named skeleton of slots you fill so nothing important gets dropped: RTF (Role, Task, Format) for simple well-defined tasks, CO-STAR for writing aimed at an audience, and RISEN for complex, constraint-heavy work that needs ordered steps and explicit exclusions. VantagePrompt picks one for you from the use case and what it extracted from your input.
Three frameworks cover most real work, and choosing between them is a smaller question than it looks: it is decided by the shape of the task, not by how thorough the framework appears. This is the comparison page — what each one is for, how to choose, and the exact rules VantagePrompt applies when it chooses for you. Each framework has its own guide for the slot-by-slot detail.
By Andrei Bădulescu, Founder at VantagePrompt·Updated
A prompt framework is a fixed skeleton for the instructions you hand a model. Instead of one long run-on paragraph, you fill named slots — role, task, format, audience, steps — so nothing important gets dropped and nothing ambiguous slips through. Three frameworks cover most real work: RTF, CO-STAR, and RISEN.
You can apply them by hand, but VantagePrompt also applies a fitting framework for you automatically — so the payoff here is knowing what is happening under the hood and why your optimized prompt looks the way it does.
What are the three frameworks, in one line each?
- RTF — Role, Task, Format. Three slots, no ceremony. The default for well-defined transformations where you already know the output shape.
- CO-STAR — Context, Objective, Style, Tone, Audience, Response. Six slots, built for writing, because persuasion depends on who is reading and how it sounds.
- RISEN — Role, Instructions, Steps, End-goal, Narrowing. Five slots, two of which exist nowhere else: an ordered verifiable plan, and an explicit statement of what to leave out.
How do I choose a framework by hand?
Pick by the shape of the task, not by how thorough the framework looks: a known output shape wants RTF, an audience wants CO-STAR, and stacked constraints want RISEN. Nothing in your input to structure yet means a plain scaffold is fine.
| Framework | Slots | Best for | Weakness |
|---|---|---|---|
| RTF | Role, Task, Format (3) | Well-defined transformations with a known output shape — most everyday coding requests. | No room for audience, tone, or ordered steps. |
| CO-STAR | Context, Objective, Style, Tone, Audience, Response (6) | Marketing copy and creative writing, where persuasion depends on who is reading. | Invites invented audiences and tones when the brief is thin. |
| RISEN | Role, Instructions, Steps, End-goal, Narrowing (5) | Complex reasoning, structured output, and constraint-heavy coding — Steps and Narrowing stop drift. | Verbose; overkill for a one-line transformation. |
| Plain scaffold | None | Very sparse prompts where you do not yet know the shape of the task. | No structural guarantees at all. |
More structure is not always better. RISEN on a one-line rename request buries the actual ask under empty ceremony. Match the framework to the depth of the task, not to how impressive it looks.
What rules does VantagePrompt use to select one?
The use case decides it outright wherever the use case already implies the shape of the work. Where it does not — Coding, and Auto — the choice falls to what the classifier extracted from your input.
| Use case | Framework | Condition |
|---|---|---|
| Marketing, Creative | CO-STAR | Always — these are writing. |
| Reasoning, Structured | RISEN | Always — these are decomposition work. |
| Coding | RISEN, else RTF | RISEN when the request is complex: explicit success criteria, or at least two hard constraints. Otherwise RTF. |
| Auto | CO-STAR, RISEN, RTF or none | CO-STAR when an audience and a tone were both extracted; then RISEN if the request is complex; then RTF if a task, a target and a format were extracted. Too sparse for any of those and no framework is applied. |
| Image, Audio, Video | None | Never — these tools have their own grammars. |
Read that table in reverse when the result is not the framework you wanted. Getting RTF where you expected RISEN means the input carried no constraints to hold; getting a plain scaffold on Auto means the classifier found neither an audience-and-tone pair nor a task-target-format triple to work with.
Why do image, audio and video skip frameworks entirely?
Because those tools already have a grammar of their own. Midjourney syntax, audio parameters, camera moves and shot durations are the structure a generator reads, and layering a prompt-engineering framework on top would replace a language the target model understands with one it does not. VantagePrompt therefore assigns no framework for those three use cases and writes in the target grammar instead.
The practical upshot for everything else: pick the right use case and write a clear input, and the framework choice follows. Knowing RTF, CO-STAR, and RISEN just lets you read the result critically — and tighten your input when you want a different shape.
Frequently asked questions
- What is the difference between RTF, CO-STAR, and RISEN?
- RTF has three slots (Role, Task, Format) and suits well-defined transformations. CO-STAR has six (Context, Objective, Style, Tone, Audience, Response) and exists for writing, which is why it separates Tone and Audience. RISEN has five (Role, Instructions, Steps, End-goal, Narrowing) and is the high-structure option: Steps forces an ordered plan and Narrowing forces at least one explicit exclusion.
- Which prompt framework should I use?
- Choose by the shape of the task. A known output shape and a single action wants RTF. A deliverable with a reader wants CO-STAR. Stacked constraints and ordered steps want RISEN. If your input is too sparse to fill any of them, no framework is the honest answer — write a clearer request first.
- Do I have to pick a framework myself in VantagePrompt?
- No. When you optimize a prompt, VantagePrompt selects a framework from your use case and what it extracted from your input, then structures the optimized prompt accordingly. Knowing the three frameworks just lets you read the result critically.
- Which framework should I use for marketing copy?
- CO-STAR. Its separate Tone and Audience slots are where persuasion lives — copy that ignores who is reading it falls flat. Fill each slot only from what you actually know: empty slots are better than fabricated ones.
- Why do image, audio, and video prompts skip frameworks entirely?
- Those tools have their own grammars — Midjourney syntax, audio parameters, camera moves — that replace prompt-engineering frameworks. VantagePrompt therefore assigns no framework for the image, audio, and video use cases.
- Is more structure always better?
- No. RISEN on a one-line rename request buries the actual ask under empty ceremony. Match the framework to the depth of the task, not to how thorough it looks.
Sources
Put it into practice.
Run this technique in the optimizer.