Prompt Frameworks: CO-STAR, CRISPE, RTF, RACE, RISEN

What a prompt framework is

A prompt framework is an acronym that reminds you what to put in a prompt. Each letter is a component — a role, the context, the task, the format of the answer. Fill in each one and you have a structured prompt.

They became popular in 2023 as people moved from one-line requests to longer prompts. Their real value is as a teaching device: a team that habitually writes “summarise this” starts writing prompts with an audience, a purpose and a format.

None of them appears in the prompting guides published by OpenAI, Anthropic or Google, which cover the same ground directly — see what is prompt engineering.

CO-STAR

CO-STAR stands for Context, Objective, Style, Tone, Audience and Response. It is the best-documented of the frameworks: it comes from the prompt engineering playbook published by the data science and AI team at GovTech, the Singapore government’s technology agency, in 2023. It was popularised by the winner of Singapore’s GPT-4 prompt engineering competition, who credited the GovTech team.

Because it separates style, tone and audience, it suits writing tasks where voice matters — marketing copy, announcements, customer emails.

Context: We are a B2B HR software company launching a mobile app for employees.
Objective: Write a LinkedIn post announcing the launch.
Style: Clear and practical, like a product update.
Tone: Confident, not hyped.
Audience: HR managers at mid-sized companies in India.
Response: Under 150 words, one line per benefit, no hashtags.

CRISPE

CRISPE stands for Capacity and role, Insight, Statement, Personality and Experiment. “Experiment” means asking for several variations to choose from.

It is often described as having been developed by OpenAI. It was not. The earliest version we could find was added to a community-maintained list of ChatGPT prompts on GitHub in February 2023 — where the C briefly stood for “Clarifying role” before becoming “Capacity and role”. It is not part of OpenAI’s guidance.

It suits exploratory work where you want options: naming, positioning, alternative drafts.

RTF

RTF stands for Role, Task, Format — the shortest of the five. We could not trace it to any original author or first publication; it appears across blogs and social posts without a source.

Its brevity is its strength. For everyday requests, naming a role, stating the task and specifying the format covers most of what matters.

Role: You are an editor for a technical blog.
Task: Shorten this paragraph without losing any technical detail.
Format: One paragraph, under 80 words.

RACE

RACE stands for Role, Action, Context, Execute, in the version published by Trust Insights, a marketing analytics consultancy, which has since moved on to a longer framework of its own. A widespread variant ends in “Expectation” instead of “Execute”; we found no traceable source for it.

It works as a compact checklist for routine requests, with context given its own slot.

RISEN

RISEN stands for Role, Instructions, Steps, End goal and Narrowing — where narrowing means the constraints. It traces to a 2024 post by a social-media educator rather than a publication.

The explicit Steps and End goal make it suit multi-step tasks with standard models. With reasoning models, prescribing steps can work against you — state the end goal and constraints, and let the model plan, as chain-of-thought prompting explains.

Which framework should you use?

It matters much less than the choice suggests. Lay them side by side and they share a core — role, context, task, format — that is exactly what the providers’ own guidance recommends.

What the frameworks leave out matters more. None includes examples, the single most reliable way to steer output (few-shot prompting). None mentions testing the prompt on real inputs. And none distinguishes reasoning models, which need less prescription, not more.

Our recommendation: use one as a teaching checklist if your team finds it helpful — CO-STAR for writing tasks, RTF for everyday ones — and then add examples, a defined output format and a small test set. That combination is what actually makes prompts reliable.