A prompt framework is a named, reusable structure that specifies which components a prompt should contain and in what order, such as the role of the model, the task, the relevant context, the constraints and the expected output format. In prompting, frameworks serve as checklists for prompt authors: they reduce the chance that an instruction omits information the model needs to produce a useful answer.
Most prompt frameworks are expressed as mnemonic acronyms, for example RTF, RISEN or CO-STAR, with each letter standing for one component. They are not algorithms or model features; they are writing conventions that make prompts more complete, more consistent across a team and easier to review, test and maintain.
How a prompt framework structures an instruction
A prompt framework works by decomposing an instruction into labeled slots that the author fills in before sending the prompt. Language models respond to everything in their input, so a prompt that silently omits the audience, the length or the format forces the model to guess, and guessing produces variable output. A framework makes those decisions explicit.
The slots map onto a fairly stable set of core prompt components that recur across almost every framework. The table below lists them with the failure each one prevents.
| Component | What it specifies | Failure it prevents |
|---|---|---|
| Role or persona | Expertise and perspective the model should adopt | Generic, shallow answers |
| Task or objective | The concrete action to perform | Answering a different question |
| Context | Background facts, data and situation | Plausible but irrelevant output |
| Audience | Who will read the result | Wrong level of detail or jargon |
| Constraints | Length, scope, rules and exclusions | Overlong or off-limits content |
| Output format | Structure such as a list, table or JSON | Output that cannot be parsed or reused |
| Examples | Sample inputs and desired outputs | Misread style or labeling scheme |
Order matters less than completeness, but many practitioners place stable instructions first and variable input last, and separate sections with prompt delimiters such as headings or XML-style tags. Clear separation helps the model distinguish instructions from data, which also reduces the risk of text inside the data being treated as a command.
Common prompt frameworks and their components
Dozens of prompt frameworks circulate in training materials, and they differ mainly in how many components they require and which ones they emphasize. None of them is a formal standard, and their letters are sometimes expanded differently by different authors. The table summarizes widely cited examples.
| Framework | Components | Typical use |
|---|---|---|
| RTF | Role, Task, Format | Quick everyday requests |
| TAG | Task, Action, Goal | Short goal-oriented instructions |
| RACE | Role, Action, Context, Expectation | Business writing and analysis |
| RISEN | Role, Instructions, Steps, End goal, Narrowing | Multi-step procedures |
| CO-STAR | Context, Objective, Style, Tone, Audience, Response | Content where voice and audience matter |
| CRISPE | Capacity, Insight, Statement, Personality, Experiment | Exploratory and creative work |
The differences are smaller than the acronyms suggest. Almost every framework combines the role, the task and the shape of the result; the longer ones add audience, tone, steps or explicit exclusions. Choosing a framework is therefore mostly a matter of matching its depth to the task.
Prompt frameworks compared with related prompting concepts
A prompt framework is often confused with neighboring concepts in the same field. The key distinction is that a framework describes what a prompt should contain, while a template, a pattern or a reasoning technique describes something more specific.
| Concept | Level of abstraction | What it fixes |
|---|---|---|
| Prompt framework | Checklist of components | Which information must be present |
| Prompt template | Concrete text with placeholders | Exact wording, variables filled at runtime |
| Prompt pattern | Reusable solution to a recurring problem | An interaction behavior, such as flipped interaction |
| Reasoning technique | Method for eliciting reasoning | How the model thinks, as in chain-of-thought |
| Context engineering | System-level design of model input | What enters the context window and when |
In practice these layers stack. A team may use a framework to decide the components, encode them in a versioned prompt template, add a reasoning technique inside the task section and manage the whole input through context engineering in an application.
Choosing and applying a prompt framework
The right framework is the simplest one that removes ambiguity for the task at hand. Short factual questions need little structure, while prompts that run thousands of times in production benefit from every component being explicit and tested. The decision table below gives typical choices.
| Situation | Suitable structure | Reason |
|---|---|---|
| One-off simple question | No framework or RTF | Extra structure adds little value |
| Marketing or customer content | CO-STAR or RACE | Tone and audience drive quality |
| Multi-step analytical task | RISEN | Explicit steps reduce skipped stages |
| Data extraction for software | Role, task, format plus examples | Strict output format enables parsing |
| Shared team prompt library | One agreed framework for all prompts | Consistency simplifies review |
Applying a framework well means filling each slot with specific content, not with generic phrases. A role such as “expert” adds little; a role such as “a tax adviser explaining rules to small business owners” constrains vocabulary, depth and assumptions. Every filled slot should then be validated through prompt evaluation on representative inputs.
Limitations and misconceptions about prompt frameworks
The most common misconception is that a named framework guarantees better results. Frameworks improve completeness, but the quality of output still depends on the accuracy of the context, the clarity of the task and the capability of the model. There is little rigorous evidence that one acronym consistently outperforms another.
- Overstructuring simple requests makes prompts longer, costlier and sometimes less natural without improving answers.
- Rigid templates can hide missing facts: a filled slot looks complete even when its content is vague or wrong.
- Model differences matter: a structure tuned for one model may need adjustment after a model update, a problem known as prompt drift.
- Role statements are not expertise: assigning a persona does not give the model knowledge it lacks.
Experienced practitioners treat a framework as a starting checklist, not a rule, and refine prompts through iterative testing against real inputs.
The short version
A prompt framework is a checklist for asking an AI model for something. Instead of typing a quick request and hoping for the best, the writer goes through a short list: who the model should act as, what exactly it should do, what background it needs, who the answer is for and what the answer should look like.
It works like a good briefing for a new colleague. A clear brief gets a useful result the first time, while a vague one leads to rework. The acronyms are simply memory aids for remembering what a good brief contains.
Where it makes a difference
A customer support team at a software company uses a language model to draft replies to about 1,500 tickets per day. Early prompts were written ad hoc by each agent, and drafts varied widely in tone, length and accuracy, so agents rewrote most of them.
The team adopts a single framework based on CO-STAR. Each prompt now states the context (product version and ticket history), the objective (resolve or escalate), the style and tone (calm, concise, no blame), the audience (non-technical customers) and the response format (greeting, three-step fix, closing line under 150 words). The shared structure becomes a versioned template in the prompt library.
After two weeks of testing against a sample of past tickets, far fewer drafts need substantial rewriting, and reviewers can spot problems faster because every prompt follows the same layout.