Prompt Frameworks Compared: CO-STAR, RISEN, CRAFT & RTF

Prompt frameworks give your team shared labels so nobody forgets context, tone, or output shape. CO-STAR, RISEN, CRAFT, and RTF are the four most copied patterns in mid-2026 chat work. This hub compares them side by side, shows when each wins, and gives a team pick guide so you stop debating acronyms in Slack. You leave knowing which skeleton fits a client email, a research brief, a creative brief, or a fast daily task. Draft structured blocks at https://promptmake.net/frameworks/costar and sibling paths for RISEN and CRAFT. PromptMake generates prompt text only. It does not run models or store your company data. Guests get about three runs per day per path. Free accounts get about five.

What prompt frameworks are (and why teams adopt them)

A prompt framework is a fixed set of field names you fill before you send a message to ChatGPT, Claude, Gemini, or an API. The labels force you to separate audience, goal, evidence, tone, and format instead of typing one vague paragraph.

Teams adopt frameworks when three people write prompts for the same job and get three different outputs. Shared labels make review easier. Managers can scan a CO-STAR block in ten seconds. Onboarding docs can say use RISEN for research summaries without re-teaching prompt theory.

Frameworks are not magic. They do not add facts the model lacks. They reduce omission errors. You still paste source text, still edit output, still run preview tests on Custom GPTs and standing skills.

Freeform prompts still win for quick one-off questions and for engineers who already think in JSON schema. Frameworks win when output shape must stay stable across dozens of runs per week.

Shared anatomy across all four

Every framework here covers role or audience, the task verb, constraints or style, and output format. Some add explicit context blocks. Some add success criteria. The difference is emphasis and field count, not philosophy.

Pick one primary framework per team library. Mixing CO-STAR on Monday and CRAFT on Tuesday for the same ticket-reply job creates review chaos.

When a framework beats freeform

Use a framework when the same job repeats weekly, when multiple writers touch the same prompt, when output must land in fixed headers, or when you onboard non-specialists who need a checklist.

Skip frameworks for exploratory chat, one-line fixes, or API calls that already use structured output schemas. Those paths have their own contracts.

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

CO-STAR is the default teaching framework in many corporate prompt workshops. Six fields map cleanly to marketing and operations work where tone and audience matter as much as the task verb.

Context holds background the model must respect: product facts, prior message, policy snippet. Objective states the outcome in one sentence. Style names register: executive brief, support reply, lesson plan. Tone names warmth or directness. Audience names who reads the result. Response defines format: bullets, table, word cap, headers.

CO-STAR shines on customer-facing copy, internal comms, and training content where a wrong tone costs trust. It is heavier than RTF for simple extraction jobs.

Draft a CO-STAR block at https://promptmake.net/frameworks/costar when you want labeled fields without building a template from scratch. Edit every fence before production use.

CO-STAR field cheat sheet

Context: paste evidence, not vibes. Objective: one verb-led sentence. Style: genre label. Tone: emotional register. Audience: job title or reader type. Response: structure and length.

Weak CO-STAR fills Style with creative and Tone with friendly. Strong CO-STAR names Style as QBR slide outline and Tone as direct, no hype.

CO-STAR example shape

Context: SaaS billing product, duplicate charge ticket, refund policy three business days. Objective: Draft customer reply approving refund. Style: Support email. Tone: Empathetic, policy-clear. Audience: Frustrated SMB admin. Response: Subject line, body under 150 words, internal CRM note as bullets.

That block gives reviewers six scan points. Missing Audience is the most common team mistake.

RISEN: Role, Instructions, Steps, End goal, Narrowing

RISEN adds procedural depth. Role sets perspective. Instructions hold rules and fences. Steps list numbered workflow the model should follow. End goal restates success in one line. Narrowing adds constraints: word limits, banned phrases, citation rules.

RISEN fits multi-step jobs: research synthesis, code review comments, lesson sequences, audit checklists. It beats CO-STAR when order of operations matters and when you need explicit never lines inside Instructions.

RISEN can feel long in fast chat UIs. Trim Steps after preview tests prove which steps the model skips anyway.

Generate a RISEN skeleton at https://promptmake.net/frameworks/risen. Same honesty: text output you paste into your tool of choice.

RISEN versus CO-STAR in one job

For a support reply, CO-STAR front-loads tone and audience. RISEN front-loads workflow steps: extract ask, check policy, draft reply, add internal note. Pick CO-STAR when voice consistency is the risk. Pick RISEN when skipped steps caused past failures.

RISEN Steps discipline

Keep Steps to five or fewer for v1. Each step starts with a verb. If a step repeats Instructions, merge them. Long step lists invite the model to skip middle steps.

End goal should be measurable: returns three sections with headers beats write a good reply.

CRAFT: Context, Role, Action, Format, Target audience

CRAFT compresses CO-STAR-like ideas into five fields with Action as the sharp task verb. Context holds evidence. Role sets voice. Format locks structure. Target audience names the reader.

CRAFT works well for content teams, product marketing, and educators who want labels without six boxes. It sits between RTF speed and CO-STAR nuance.

Use https://promptmake.net/frameworks/craft when your team prefers five fields and clear Action language over separate Style and Tone lines.

CRAFT Action verbs

Action should be one primary verb: summarize, compare, extract, rewrite, critique, outline. Stack secondary verbs in Context instead of Action.

Format examples: markdown table with three columns, JSON array of strings, five bullet executive summary.

CRAFT example for product marketing

Context: Launch blog for API rate limits feature, technical buyers, no competitor bashing. Role: B2B product marketer. Action: Draft hero section and three benefit bullets. Format: H2 plus bullets under 40 words each. Target audience: Platform engineers evaluating observability tools.

RTF: Role, Task, Format (plus Context when needed)

RTF is the lightest framework here. Role, Task, Format fit on one screen. Add Context as a fourth paste block when the job needs source material. Add Constraints when you need never lines.

RTF wins on Gemini 3.5 Flash, GPT-5.5 Instant, and daily chat where speed beats ceremony. It is the best default before you escalate to CO-STAR or RISEN for high-stakes external copy.

This blog ships a dedicated RTF examples article with paste-ready blocks. This comparison page treats RTF as the baseline other frameworks extend.

RTFC upgrade path

When RTF previews drift, add Context with pasted source. When fences fail, add Constraints with three never lines. When tone still slips, switch the job to CO-STAR with explicit Tone and Audience fields.

RTF one-liner that still works

As a security reviewer for a fintech backend team, list high-confidence issues in this PR diff. Format: Must fix and Should fix sections, bullets only, two sentences max per bullet.

That sentence is valid RTF without labels. Labels help teams; prose helps solo speed.

Side-by-side comparison table (mental model)

Field count: RTF three core, CRAFT five, CO-STAR six, RISEN five with heavier Steps and Narrowing.

Best for tone-heavy external copy: CO-STAR. Best for procedural internal work: RISEN. Best for marketing and education balance: CRAFT. Best for fast daily tasks: RTF.

Review time: RTF fastest to scan. RISEN slowest unless Steps stay short.

Onboarding difficulty: RTF lowest. CO-STAR moderate. RISEN highest when teams overfill Steps.

Map each recurring job in your library to one primary framework. Document the choice in your wiki so new hires do not invent a fifth pattern.

Team pick guide: which framework for which job

Start from job risk, not acronym popularity. High external tone risk points to CO-STAR. Multi-step internal workflows point to RISEN. Content and campaign drafts point to CRAFT. Daily extraction and quick rewrites point to RTF.

Run a two-week pilot with one framework per job type. Log preview failures. If the model skips steps, move support jobs from CO-STAR to RISEN. If prompts feel bloated, move research summaries from RISEN to CRAFT.

Assign a framework owner who updates shared templates when policy changes. Framework rot looks like empty Style fields and copied lorem in Context.

Pick by team role

Support and success: CO-STAR or RISEN depending on whether tone or steps failed last quarter.

Marketing and content: CRAFT or CO-STAR. CRAFT when Action clarity matters. CO-STAR when Tone and Audience need separate boxes.

Engineering and data: RTF or RISEN. RTF for quick reviews. RISEN when numbered verification steps matter.

Executives and chiefs of staff: RTF with tight Format. They will not maintain six fields.

Pick by model class

Fast chat models on GPT-5.5 Instant or Gemini 3.5 Flash: RTF or CRAFT first. Reasoning-class models on GPT-5.6 Sol or Claude Opus 5: goal plus constraints plus format often enough; add CO-STAR when tone audits fail.

Agent Skills and Custom GPTs: embed one framework in standing instructions so every chat inherits the same labels.

Draft frameworks on PromptMake before you scale

Blank fields stall teams. https://promptmake.net/frameworks/costar turns a rough job brief into labeled CO-STAR text. Sibling generators cover RISEN and CRAFT with the same pattern.

Workflow: write a three-sentence seed offline, generate once, paste into chat or into Custom GPT Instructions, run four preview prompts, save winning text in your internal library.

Honest limits: PromptMake outputs prompt text. It does not execute jobs, store company PDFs, or push configs into OpenAI, Anthropic, or Google builders. You copy and edit locally.

Guests get about three runs per day per tool path. Registered free users get about five. Use a run to beat blank-page delay, then edit every fence.

For the same standing job on Claude Skills or Gemini Gems, open https://promptmake.net/skills and map framework fields into SKILL.md or Gem Persona blocks without copying ChatGPT instructions verbatim across hosts.

Common mistakes when teams adopt prompt frameworks

Mistake 1: Adopting four frameworks at once so nobody knows which template to open.

Mistake 2: Filling Style and Tone with empty adjectives instead of concrete labels.

Mistake 3: Pasting frameworks without Context when the job needs source text.

Mistake 4: Treating generator output as final without preview tests.

Mistake 5: Using CO-STAR for one-line extraction tasks that RTF handles in ten seconds.

Mistake 6: Skipping Format so every output arrives as a wall of prose.

Mistake 7: Changing framework per writer instead of per job type.

FAQ

What are the main prompt frameworks in 2026?

CO-STAR, RISEN, CRAFT, and RTF are the four most common labeled patterns for chat and agent work. CO-STAR emphasizes context, objective, style, tone, audience, and response format. RISEN adds numbered steps and narrowing constraints. CRAFT compresses to context, role, action, format, and target audience. RTF keeps role, task, and format as the minimum viable spec.

When should I use CO-STAR versus RTF?

Use CO-STAR when tone, audience, and response shape all affect quality and external trust. Use RTF when you need a fast repeatable spec for internal drafts, extraction, or quick reviews. Move from RTF to CO-STAR when preview tests show tone drift or missing audience awareness.

How is RISEN different from CRAFT?

RISEN foregrounds numbered Steps and an End goal plus Narrowing constraints for procedural jobs. CRAFT foregrounds a single Action verb with Format and Target audience for content and marketing work. Pick RISEN when skipped steps caused failures. Pick CRAFT when clarity of action and format matters more than step order.

Can PromptMake generate CO-STAR and other framework prompts?

Yes. https://promptmake.net/frameworks/costar drafts CO-STAR blocks. https://promptmake.net/frameworks/risen and https://promptmake.net/frameworks/craft cover the other named generators. PromptMake returns text you copy into ChatGPT, Claude, Gemini, or standing GPT and Skill configs. It does not run the model for you.

How do I choose one framework for my team?

List your top five recurring prompt jobs. Assign one framework per job based on risk: tone-heavy jobs get CO-STAR, step-heavy jobs get RISEN, content jobs get CRAFT, quick jobs get RTF. Pilot two weeks, log failures, adjust. Document the choice in your internal wiki so onboarding stays consistent.

Do prompt frameworks work with reasoning models?

Reasoning-class models such as GPT-5.6 Sol and Claude Opus 5 often need clear goal, constraints, and format more than long role prose. RTF or tight CRAFT blocks frequently suffice. Add CO-STAR Tone and Audience fields when external copy still drifts after reasoning upgrades.

Where can I learn RTF in depth on this blog?

Read rtf-framework-prompts-examples on promptmake.net for copy-paste RTF and RTFC blocks with marketing, engineering, and support samples. This comparison page situates RTF against CO-STAR, RISEN, and CRAFT so you pick the right depth for each job.