Prompt Frameworks: RTF, RACE, CRISPE & RISEN

A prompt framework is a fixed schema that gives your instruction structure — instead of free-form wording, you fill in named fields like role, task, and format. The five most important for beginners are RTF, RACE, CRISPE, chain-of-thought, and RISEN. Each solves a different problem, and all of them can be learned in minutes.

Why use a framework at all? Because the most common beginner mistake is not too little technique but missing structure. A vague prompt forces the model to guess your intent — and the more open the input, the more average the output. A framework prevents this by walking you step by step through the building blocks a good brief needs. If you want to solidify the basics first, start with our guide to [prompt engineering fundamentals](/magazin/prompt-engineering-fundamentals).

A note up front: frameworks are not magic formulas. They work because they give the model more context, more clarity, and a clear expectation for the result. Once you understand the principle, you can mix your own variants instead of slavishly working through an acronym. That is the focus here: not just the letters, but when each schema is the best choice.

What is a prompt framework and why do you need one?

A prompt framework is a repeatable template that breaks a good prompt into named components. Instead of wondering each time what you forgot, you check off a fixed list: role, context, task, format. This makes your results more predictable and saves time.

The real value lies in reproducibility. A prompt that worked once can be saved as a schema and applied to hundreds of similar tasks. Frameworks are therefore the bridge between a one-off good input and a professional, team-ready [prompt library](/magazin/15-prompt-engineering-techniques). Working with templates means you stop rewriting from scratch and instead improve one central asset.

The table below maps the five frameworks in this article to their use case:

How does the RTF framework (Role, Task, Format) work?

RTF is the leanest framework of all: you set only three things — who the model is (Role), what it should do (Task), and in what form the answer comes (Format). For most everyday tasks, nothing more is needed.

Precisely because it is so compact, RTF is perfect for getting started and for quick requests where an elaborate schema only costs time. Its strength lies in the third component: explicitly enforcing the format is one of the most effective levers there is, because it prevents free-form prose that you would have to clean up afterward.

Prompt

You are an experienced LinkedIn copywriter (Role). Write a post explaining why small teams benefit from AI tools (Task). Format: a hook in the first line, three short paragraphs, and a question to readers at the end — 150 words maximum (Format).

Make sure to keep the format genuinely concrete. "Write a list" is weaker than "Respond as a numbered list with at most five points, each point one sentence." The more precise the format requirement, the less you have to fix afterward.

How does the RACE framework (Role, Action, Context, Expectation) work?

RACE extends RTF with two crucial building blocks: the Context (background information) and the Expectation (your expectation for quality and scope). The four elements are Role, Action, Context, Expectation — a complete brief without feeling overloaded.

The gain over RTF is the context. A model without context averages over everything it has learned about the topic; with concrete context, it targets your case. The expectation, in turn, controls measurable properties like length, tone, or audience. RACE is therefore the most versatile all-purpose framework and a good default when you are unsure which to pick.

Prompt

You are an experienced nutritionist (Role). Create a weekly lunch plan (Action) for a working family with two children who has little time to cook and eats vegetarian (Context). The plan should include five dishes, each preparable in under 30 minutes, with a short shopping list (Expectation).

Compare that with a bare "give me lunch ideas" — the difference in usefulness is enormous. RACE forces you to supply exactly the information that makes the result usable.

How does the CRISPE framework work?

CRISPE is the most detailed of the five frameworks and is built for complex, creative tasks. The six building blocks are Capacity (capability/expertise), Role (role), Insight (background knowledge), Statement (the actual task), Personality (tone and style), and Experiment (allow multiple variants).

The special value lies in the last two blocks. Personality gives you fine control over the voice — from sober and factual to playful and provocative. Experiment explicitly allows the model to offer several solution paths instead of committing to the first one. For brainstorming, brand communication, and concept work, that is worth its weight in gold.

Prompt

Act as an AI with deep expertise in brand strategy and advertising psychology (Capacity). You are the creative director at an agency (Role). Context: a sustainable Swiss start-up is launching reusable coffee cups and wants to stand out from cheap products (Insight). Develop three claim suggestions for the launch campaign (Statement). Tone: confident, warm, free of eco clichés (Personality). Deliberately deliver different styles and briefly justify each one (Experiment).

CRISPE is powerful but also demanding. Use it where creativity and differentiation count — for a quick email it would be overkill.

How does chain-of-thought prompting work?

Chain-of-thought (CoT) is not an acronym with fields but a reasoning framework: you ask the model to write out its intermediate steps before giving an answer. Instead of jumping straight to the result, the model thinks "aloud" — and this markedly improves accuracy on logic, math, and multi-step tasks.

The effect is well documented. In the original study by Wei et al. (Google, 2022), CoT lifted a large model's success rate on the GSM8K math benchmark from roughly 18 to 58 percent — simply by asking it to reason step by step. The reason: by formulating intermediate results, the model gains more computation room and grounds each step on the previous one.

Prompt

An online shop had 1,240 orders in March with an average value of 38 euros. In April, the order count rose by 15 percent, but the average value fell by 4 euros. What was the revenue in both months, and how large is the difference? Think step by step and show your calculation.

CoT combines with any other framework — you simply append the instruction. But note the boundary: on pure taste or style questions it adds nothing and only lengthens the answer. Modern reasoning models also often think step by step internally; with them, the explicit instruction is frequently redundant. For an in-depth treatment, see our guide to [chain-of-thought prompting](/magazin/chain-of-thought-prompting).

How does the RISEN framework work?

RISEN is built for multi-step, clearly delimited workflows. The five building blocks are Role (role), Instructions (clear instruction), Steps (steps to follow), End goal (end goal), and Narrowing (constraints and limits).

The difference from RACE lies in the explicit Steps and the Narrowing. While RACE describes a brief as a whole, RISEN breaks it into an ordered sequence and at the same time defines what should be left out. That makes it ideal for tasks where order matters and you want to deliberately limit the solution space — such as analyses, audits, or structured reports.

Prompt

You are an experienced SEO consultant (Role). Analyze the following homepage of a tradesperson's business and deliver concrete improvements (Instructions). Proceed in this order: 1. Check the title and meta description, 2. evaluate the heading structure, 3. identify missing local signals (Steps). The goal is a prioritized list of the three most impactful measures (End goal). Limit yourself to on-page factors and ignore backlinks and load time (Narrowing).

The Narrowing is the often-underrated trump card: by explicitly stating what is *not* part of the task, you prevent rambling answers and keep the model on course. If you want to dive deeper into individual building blocks like role prompting or structured output, our overview of [15 prompt engineering techniques](/magazin/15-prompt-engineering-techniques) is worth a read.

Which framework fits which task?

The choice depends on the task type and the effort you want to invest. A simple rule of thumb: start with the leanest framework that could work, and reach for a more detailed one only when the output demands it.

• Quick everyday tasks (email, short text, summary): RTF almost always suffices.
• Complete briefs with context (product copy, a plan, an explanation for an audience): RACE is the reliable standard.
• Creative and differentiating work (campaigns, concepts, exploring variants): CRISPE plays to its strengths.
• Logic, calculation, multi-step reasoning: append chain-of-thought to any other framework.
• Structured, ordered workflows (analyses, audits, reports): RISEN with its steps and limits.

The most important advice at the end: understand the building blocks, not just the acronyms. Nearly all frameworks consist of the same core elements — role, context, task, format, limits — just weighted and named differently. Master these elements and eventually you won't need an acronym at all; you'll intuitively place the right blocks. Until then, frameworks are the best scaffold for writing good prompts reliably and repeatably.

Save your proven schemas as templates in one place, rather than rewriting them from memory every time. That turns a one-off hit into a repeatable standard your whole team can use — which is exactly what Prompt2Love is built for. If you are right at the start, our introduction to [how to write effective AI prompts](/magazin/write-effective-ai-prompts) will help.

Frequently Asked Questions

Do I even need a framework, or is a clear prompt enough?

A clear prompt is always the goal — a framework is just the scaffold that helps you get there. At the start it prevents you from forgetting important building blocks. With experience, the schemas merge into an intuition, and you no longer need them consciously.

Which framework is best for beginners?

RTF (Role, Task, Format) is the ideal entry point because it has only three fields and delivers noticeably better results right away. As soon as context matters, switch to RACE. Together, the two cover the bulk of all everyday tasks.

Can I combine multiple frameworks?

Yes, and it often makes sense. Chain-of-thought can be appended to virtually any other framework, and RACE or RISEN can be combined with structured output. Just be careful not to stack too many competing rules — otherwise the compliance rate drops.

Do these frameworks work with all AI models?

The core principles hold across all major models. Individual providers, however, have quirks: Claude responds especially well to detailed roles and XML tags for structure, while reasoning models often handle chain-of-thought internally. Test your important prompts on every model you run in production.

What about frameworks like CO-STAR or STAR?

There are dozens of acronyms, and many overlap heavily. CO-STAR, for instance, is a marketing-oriented variant focused on style, tone, and audience. Once you master the five building-block logics presented here, you can place any further framework in minutes, because it simply recombines the same core elements.