Prompt frameworks that work: RTF, CO-STAR, CRISPE, and when to use each

What a framework actually does

A prompt framework is a memory aid. It is a fixed list of things to include so you do not forget the audience, the format, or the constraint that would have saved you three rounds of rewrites.

That is all it is. The model does not recognize "CO-STAR" as a special mode. It reads your words. A framework helps because it forces you to write the words a good brief needs, in a consistent order, every time.

So the honest question is not "which framework is the most powerful" but "which one will I actually use at 4:45 on a Thursday." The answer depends on the task, and this guide gives you a table for that.

If you have not read the prompt engineering guide, the short version: strong prompts contain a role, a task, context, a format, constraints, and examples. Every framework below is a different way of slicing those same six ingredients.

The frameworks, one by one

RTF: Role, Task, Format

The smallest useful framework. Who the model is, what it should do, what the output looks like.

Role: You are a senior operations manager at a regional logistics company.
Task: Write a one-page memo proposing that we move dock scheduling to 15-minute appointment windows.
Format: Four sections with headings: Problem, Proposal, Costs, Next steps. Under 400 words.

RTF fits quick, well-defined tasks where you already have the context in your head and can drop it into the Task line. It fails when the audience matters or the source material is large, because there is no slot for either. You end up jamming context into the task and forgetting the constraints.

Use RTF for: memos, rewrites, quick drafts, code snippets, lists, explanations.

Persona-Task-Format and its cousins

Persona-Task-Format is RTF with a stronger emphasis on the persona: a job title plus a point of view, an experience level, and even a temperament. "You are a cautious CFO who has been burned by optimistic forecasts" gets a more useful critique than "You are a CFO."

You will see close relatives with different labels: TAG (Task, Action, Goal), APE (Action, Purpose, Expectation), RISEN (Role, Instructions, Steps, End goal, Narrowing). They are all the same idea with one or two extra slots. RISEN's Steps and Narrowing (constraints) are the useful additions; pick it when the task has an order of operations.

Use Persona-Task-Format for: critiques, reviews, practice conversations (a tough customer, a skeptical interviewer), and any task where the perspective of the reviewer is the product.

CO-STAR: Context, Objective, Style, Tone, Audience, Response

CO-STAR is the framework for writing that other people will read. Its six slots map neatly onto what a good communications brief contains.

• Context: the background the model needs.
• Objective: what you want the writing to achieve, as opposed to what it is.
• Style: the writing style, often by reference ("like a plain-English bank letter," "like our past product updates").
• Tone: the emotional register (warm, neutral, firm, apologetic).
• Audience: who reads it and what they already know or feel.
• Response: the format and length of the output.

Separating Objective from the task is CO-STAR's best idea. "Announce the price increase" is a task. "Have customers finish the email feeling the increase is fair and knowing exactly what happens if they do nothing" is an objective, and the model writes differently when it knows that.

Use CO-STAR for: customer emails, announcements, marketing copy, difficult messages, anything audience-facing. It is heavier than RTF, and for a two-line Slack message it is overkill.

CRISPE: Capacity and Role, Insight, Statement, Personality, Experiment

CRISPE comes from the early wave of ChatGPT prompt guides and has a slot the others lack: Experiment, which asks for multiple variations. The rest are Capacity and Role (who the model is), Insight (background), Statement (the task), and Personality (style and tone).

That Experiment slot makes CRISPE the exploration framework. When you do not yet know what you want, asking for three deliberately different versions and reacting to them is faster than trying to describe the perfect one up front.

Capacity and Role: You are a product marketer who has named dozens of B2B software features.
Insight: We are launching automated text reminders for physical therapy clinics. Our brand voice is plain and unflashy. Competitors use names like "SmartRemind" and we want to avoid that register.
Statement: Propose names for this feature.
Personality: Practical, a little dry, no puns.
Experiment: Give me three sets of five names. Set A: literal and descriptive. Set B: one-word names. Set C: names built around the outcome for the clinic (fewer no-shows, a calmer front desk). For each name, add a six-word rationale.

Use CRISPE for: naming, brainstorming, options for a headline or a structure, early-stage strategy where you want a spread of ideas rather than one answer.

Chain of Density: a technique for summaries

Chain of Density is different in kind. It is not a checklist for building prompts; it is a procedure for producing a better summary. It was described in a 2023 research paper by Griffin Adams and co-authors, and the idea travels well outside research.

The procedure: ask for a first summary that is deliberately vague. Then, over several rounds, ask the model to find one to three specific entities from the source that the summary is missing (names, numbers, dates, decisions) and rewrite the summary to include them without making it longer. Each round trades filler for substance.

You are an editor. I will give you a document. Produce increasingly dense summaries in four rounds.

Round 1: write a summary of about 80 words that is deliberately general, with few specific names, numbers, or terms.

For each following round: identify one to three specific, informative entities from the document that the previous summary is missing (people, numbers, dates, decisions, defined terms). Rewrite the summary to include them at the same word count. Compress, merge sentences, and cut filler such as "the document discusses" to make room.

Rules: never drop an entity from an earlier round. Never add anything that is not in the document. Keep each summary self-contained.

Output for each round:
Missing entities: [LIST]
Summary: [TEXT]

Document:
[PASTE DOCUMENT]

The fourth round is usually the one you want. The earlier rounds exist to force the model to notice what it left out, which is the whole point.

Use Chain of Density for: executive summaries, abstracts, briefing notes on long reports. Do not use it for meeting notes where you need actions and owners; a structured extraction prompt does that better (see how to summarize documents with AI).

Worked example: one request, three frameworks

The request: explain a 6 percent price increase to the existing clients of a small bookkeeping firm.

RTF version. Role: bookkeeping firm owner. Task: write the price-increase email. Format: under 200 words, plain paragraphs. Result: correct, quick, a little cold. It states the number, gives a date, and thanks the client. Fine for a first draft you will edit heavily.

CO-STAR version. Adds the objective (clients should feel the increase is fair and know they can call), the audience (long-time small-business owners, some of them price-sensitive), the style (like a note from a trusted advisor, not a vendor notice), and the tone (warm, direct, no apology spiral). Result: the draft explains what has changed in the service over the past two years before the number appears, offers a call, and gives the effective date twice. It reads like something the owner would send.

CRISPE version. Same inputs, plus Experiment: three versions, one leading with the value delivered, one leading with the number, one as a short note with a link to a longer explanation. Result: a spread you can react to. Most owners pick one and borrow a sentence from another.

The frameworks did not produce different intelligence. They produced different amounts of information about the reader. The CO-STAR draft was better because you told the model who the reader is and what you want them to feel.

Decision table

Where frameworks break

They do not fix missing material. A perfect CO-STAR brief for "summarize the contract" with no contract attached produces a confident summary of nothing. Paste the source.

They do not prevent hallucination. Every framework above will happily produce a fake statistic if you ask for "compelling data." Add the constraint "do not invent figures; write [NEEDS DATA] where a number is missing" to any framework, and verify what comes back. The fact-checking guide has the routine.

Labels are not magic. Writing "Tone:" before a word does not make the model understand tone better. What helps is the example that follows the label. "Tone: like this past email of mine: [PASTE]" beats "Tone: professional yet approachable."

Long frameworks crowd out iteration. If filling the template takes ten minutes, you will stop iterating, and the second and third messages are where quality comes from. For most daily work, a short prompt plus two follow-ups beats a long prompt plus zero.

They age. Reasoning modes in the current assistants make some older advice ("think step by step," elaborate persona descriptions) less necessary. What has not aged is context, format, and constraints. Those will still matter in 2030.

Build your own house framework

The framework you will use is the one you wrote. Take the six ingredients, keep the ones that matter for your recurring tasks, and save the result as a note or as the instructions in a custom assistant (a Custom GPT, a Claude Project, or a Gemini Gem; see the comparison guide).

A house framework for a customer-support lead might be: Ticket, Customer history, Policy that applies, Desired outcome, Tone example, Format. A house framework for an analyst might be: Question, Data (with headers), Definitions, Known caveats, Output table columns. Neither has a catchy acronym. Both beat CO-STAR for those jobs, because they hold exactly the information those jobs need.

Write yours, use it for two weeks, then cut whatever you never filled in.

Next steps

• Start with the fundamentals in the prompt engineering guide if you skipped it.
• Use the prompt builder to assemble a prompt from the six ingredients without memorizing any acronym.
• Try the summarization techniques in how to summarize documents with AI.
• Pull ready-made prompts from brainstorming and strategy and writing and email.
• Read the glossary entries on chain-of-thought and few-shot prompting.
• The course turns these frameworks into exercises built on real work samples.