Prompt Framework Picker: Match Job to CO-STAR, RISEN or CRAFT

TL;DR: This prompt framework picker gets you to the right structure in under a minute. Answer five yes or no questions about the job in front of you. One-line lookups and brainstorms go freeform. Ordered procedures with a finish line go to RISEN. Reader-facing writing where tone decides quality goes to CO-STAR. Fact-heavy output with a strict format and a channel voice goes to CRAFT. Short repeat tasks with a known shape go to RTF. Each branch below comes with the job signals that trigger it and a filled example you can paste into ChatGPT, Claude, or Gemini. Build the CRAFT branch at https://promptmake.net/frameworks/craft when you want labeled fields generated for you.

How the prompt framework picker works

Most people pick a prompt framework by habit. They learned CO-STAR first, so every prompt becomes CO-STAR, including a two-word lookup that needed no structure at all. The picker flips that. You start from the job, read a few signals, and let the signals choose the frame. Each framework has fields that solve a specific failure. CO-STAR has Tone and Audience because reader-facing text fails on voice. RISEN has Steps and Narrowing because procedures fail on order and scope. CRAFT has Context and Format up front because fact-heavy outputs fail on missing background and messy shape. RTF stays short because small tasks fail when you overbuild them. Match the failure you fear to the field that prevents it. The two subsections below explain the questions and how to read the answer.

The five questions, in order

1. Would one plain sentence get a usable answer? If yes, write freeform and stop.
2. Does the job describe a process with ordered steps and a clear end state? If yes, use RISEN.
3. Will an outside reader judge the output on tone and fit, such as a customer, an executive, or the public? If yes, use CO-STAR.
4. Does the output depend on background facts plus a strict format, such as a table, a bulleted brief, or a fixed template, with a set voice? If yes, use CRAFT.
5. Is it a short task you repeat often with a known output shape? If yes, use RTF (Role, Task, Format).

Reading the result

Stop at the first yes. The order matters: it filters out small jobs first, then catches procedures, then reader-facing writing, then fact-and-format work. If you reach the end with no yes, default to RTF and add a Context line.

Treat the pick as a starting frame. You can add one field from another framework when the job needs it, such as a Tone line inside RISEN for a customer-facing checklist. Two borrowed fields is the limit before you should switch frames.

Job signals cheat sheet

Scan your request for these words and situations. They map to branches faster than the full question list once you have used it a few times.

• Freeform signals: “what is,” “rename,” “give me ten ideas,” “fix this typo,” a pasted schema or code with an obvious ask.
• RISEN signals: “step by step,” “checklist,” “SOP,” “process,” “audit,” “migrate,” “review in this order,” a clear done state.
• CO-STAR signals: “email to,” “post for,” “announcement,” “reply to customer,” “pitch,” any named audience with feelings about the result.
• CRAFT signals: “summarize this data,” “brief for standup,” “turn notes into,” “in a table,” “max five bullets,” pasted source material plus a format rule.
• RTF signals: daily repeat jobs like “write alt text,” “tag these tickets,” “draft a commit message,” where the format never changes.

One worked example per branch

The five examples below use the same office setting so you can see how the signals shift the pick. Each one lists the request as it arrived, the signals that decided the branch, and a filled prompt. Copy the prompt, swap the bracketed details for your own, and run it. If your result misses, check which field was thin before you change frameworks. A weak Audience line in CO-STAR or a vague End goal in RISEN causes more bad outputs than picking the wrong acronym. The examples also show where each branch stops being worth it, which helps you catch overbuilt prompts before they waste a run.

Freeform: a quick lookup

Request: “What's the difference between median and mean for our survey report?” Signals: one question, no audience, no format stakes. Prompt: “Explain the difference between median and mean in three sentences, with one example using survey scores from 1 to 10.” Adding CO-STAR fields here would slow you down and add nothing.

RISEN: a procedure with a finish line

Request: “Help me clean up our CRM duplicates before the quarter closes.” Signals: ordered process, clear end state, scope risk. Prompt:

• ROLE: CRM operations specialist.
• INSTRUCTIONS: Produce a cleanup procedure for duplicate contact records in [HubSpot].
• STEPS: (1) define a duplicate by email and company domain, (2) export candidates, (3) pick a survivor rule, (4) merge in batches of 50, (5) spot-check 10 merges.
• END GOAL: a checklist a coordinator can run in one afternoon with zero lost deal history.
• NARROWING: no deletions, merges only; skip records owned by the enterprise team; under 300 words.

CO-STAR: reader-facing writing

Request: “Write the email telling customers about the price change.” Signals: outside readers, tone decides the reaction, named audience. Prompt:

• CONTEXT: Starter plan rises from $12 to $15 per month on November 1, 2026; existing customers keep $12 until February 1, 2027.
• OBJECTIVE: Announce the change and the grace period so customers stay subscribed.
• STYLE: short product update email.
• TONE: plain, respectful, no spin.
• AUDIENCE: small-business owners on Starter for over six months.
• RESPONSE: subject line plus body under 150 words, one link to the pricing page, no exclamation marks.

CRAFT: facts plus a strict format

Request: “Turn this feedback spreadsheet into something for tomorrow's standup.” Signals: pasted source data, fixed output shape, internal voice. Prompt:

• CONTEXT: [paste 40 rows of customer feedback with date, plan, and comment]. Only use this data.
• ROLE: product analyst who prepares standup notes.
• ACTION: summarize the top themes and flag anything marked urgent.
• FORMAT: max five bullets, each with theme, count of mentions, and one short quote; then one line listing urgent row numbers.
• TONE: neutral and direct, no recommendations.

RTF: a short repeat task

Request: “Write alt text for these product photos.” Signals: repeats daily, format never changes, little background needed. Prompt: “ROLE: e-commerce accessibility writer. TASK: write alt text for the product described below. FORMAT: one sentence under 125 characters, start with the product type, include color and material, no ‘image of.’ Product: [description].” Save it once and reuse it every day.

When two branches fit

Some jobs trip two questions. A customer-facing onboarding checklist is both a procedure (RISEN) and reader-facing (CO-STAR). The picker order sends it to RISEN, because a wrong step does more damage than a stiff tone. Borrow Audience and Tone lines from CO-STAR and add them under Narrowing.

CO-STAR and CRAFT overlap most. Pick CO-STAR when the reader sits outside your team and their reaction is the goal. Pick CRAFT when the reader is internal and the format is the goal: a standup brief, a changelog, a research summary with fixed sections.

Long jobs can switch frames partway. Use RISEN to plan a launch sequence, then CO-STAR for each announcement the plan calls for. One frame per prompt keeps each run easy to review.

Common mistakes with a prompt framework picker

Forcing structure on a lookup is the most common miss. If one sentence works, send one sentence.

Stopping at the acronym and filling fields with filler undoes the benefit. “Tone: professional” and “Audience: users” score as blank.

Picking by the framework's fame instead of the failure you fear leads to Tone fields on data summaries and missing Steps on procedures.

Stacking every field from every framework produces a twelve-line form nobody fills twice. Borrow at most two fields.

Using this page as a comparison chart misses its purpose. For a side-by-side of all four frameworks field by field, read https://promptmake.net/blog/prompt-frameworks-compared. For the broader call on structure versus freeform across your week, read https://promptmake.net/blog/when-to-use-prompt-frameworks. This page is the fast decision path for one job at a time.

Model notes for September 2026

Frameworks travel across models. GPT-5.6 Sol, Claude Opus 5, Claude Sonnet 5, Gemini 3.1 Pro, and Gemini 3.5 Flash all respond well to labeled fields. Reasoning-class models such as GPT-5.6 Sol need goal, constraints, and format more than step-by-step coaching, so RISEN Steps should list the procedure you want in the output and leave the reasoning to the model.

Fast chat tiers such as GPT-5.5 Instant and Gemini 3.5 Flash benefit most from a strict Format or Response line, because they drift on length. If a fast model keeps overrunning, tighten that one field before you switch frameworks.

Build the picked framework with PromptMake

Once the picker lands on a branch, you can fill the fields by hand or generate a first draft. PromptMake has live generators for all three named frameworks. For fact-heavy briefs, open https://promptmake.net/frameworks/craft, paste your rough request, and get Context, Role, Action, Format, and Tone sections back. For reader-facing writing, use the CO-STAR path, and for procedures, the RISEN path, both linked from the CRAFT page.

For freeform and RTF jobs, https://promptmake.net/text turns a rough idea into a model-ready prompt without forcing extra fields. PromptMake writes prompt text only and does not run models. As of September 2026, guests get about three runs per day per tool path and free registered accounts about five.

FAQ

What is a prompt framework?

A prompt framework is a set of labeled fields you fill before sending a prompt, such as Context, Role, Task, Tone, or Format. Common ones include CO-STAR, RISEN, CRAFT, and RTF. Each one targets a different kind of job and a different failure. Using the right one helps the model and helps you review the result.

How do I choose the right prompt framework?

Start from the job and ask five questions in order: would one sentence work, is it a procedure, does an outside reader judge tone, does it need facts plus a strict format, or is it a short repeat task. Stop at the first yes. Those answers map to freeform, RISEN, CO-STAR, CRAFT, and RTF. Borrow up to two fields from another framework when the job needs them.

When should I use CO-STAR instead of CRAFT?

Use CO-STAR when the reader sits outside your team and their reaction is the goal, such as customer emails and public posts. Its separate Style, Tone, and Audience fields catch voice problems. Use CRAFT when the reader is internal and the output format matters most, such as standup notes or research briefs. Both start with Context, so switching costs little.

When is RISEN the best prompt framework?

RISEN fits jobs that describe a process with ordered steps and a clear end state: SOPs, audits, migrations, and review checklists. Its Steps field locks the order, and Narrowing fences scope. Pick it when a wrong step would cause more harm than a stiff tone. Add Tone from CO-STAR if a customer will read the checklist.

Do I need a prompt framework for every prompt?

No. Lookups, quick rewrites, brainstorms, and asks that come with a pasted schema or code often work best as one clear sentence. Frameworks earn their keep on repeat jobs, shared prompts, and outputs where tone or format must hold steady. The first question in the picker exists to catch jobs that need no structure.

What is the RTF prompt framework?

RTF stands for Role, Task, Format. It is the lightest named framework and fits short tasks you repeat often, like alt text, ticket tags, or commit messages. Add a Context line when the model needs background it cannot guess. Save a filled RTF prompt and reuse it daily.

Is there a free prompt framework generator?

Yes. PromptMake offers free CO-STAR, RISEN, and CRAFT generators, including https://promptmake.net/frameworks/craft. Guests get about three runs per day per path and registered free accounts about five, as of September 2026. The tool writes prompt text; you paste it into ChatGPT, Claude, or Gemini.