The internet is full of prompt frameworks. CRAFT. RACE. CLEAR. CO-STAR. RTF. RISEN. Every month someone publishes a new one, usually with a clean graphic and a claim that it produces "10x better results."
I got tired of taking these claims on faith, so I tested four of the most popular frameworks against each other with the same task. Same model (Claude 3.5 Sonnet), same temperature, same task — just different frameworks structuring the prompt. Here's what I found.
The four frameworks
Before the test, a quick rundown of each:
CRAFT — Context, Role, Action, Format, Tone You establish the background context first, assign a role, specify what action to take, define the output format, and set the tone. Heavy on setup, good for content creation tasks.
RACE — Role, Action, Context, Execute Leaner version. Role first, then the action, then context to support it, then a directive to go. Less structured around format and tone, more focused on getting the task done fast.
CLEAR — Concise, Logical, Explicit, Adaptive, Reflective This one's different — it's not a template structure, it's a set of quality criteria for the prompt itself. Concise: no filler. Logical: ordered reasoning. Explicit: leave nothing implicit. Adaptive: match the model's capabilities. Reflective: ask the model to evaluate its own output.
CO-STAR — Context, Objective, Style, Tone, Audience, Response The most comprehensive framework here. Adds explicit Audience and Style dimensions that the others skip. Good for audience-aware writing tasks.
The test task
I picked a real-world task that most people would actually use AI for:
"Write a LinkedIn post about AI tools for marketing managers."
I ran each framework as a structured prompt with that core task, keeping everything else equal. Here are the prompts and what came out.
CRAFT prompt
Context: Marketing managers at mid-size B2B companies are overwhelmed with content demands and looking for ways to use AI tools to move faster without sacrificing quality.
Role: You are a B2B content strategist who regularly advises marketing teams on AI adoption.
Action: Write a LinkedIn post about AI tools for marketing managers.
Format: LinkedIn post. 150–200 words. No bullet lists — flowing paragraphs. Hook line first, no label or header.
Tone: Practical and direct. No hype. Written by a peer, not a vendor.
Output quality: Strong. The role + context combination gave the model enough background to make specific claims rather than generic ones. The tone specification ("no hype, written by a peer") produced something that didn't read like an ad. Format compliance was perfect.
Weakness: The prompt itself was 100 words of setup for a 180-word output. High overhead for simple tasks.
RACE prompt
Role: B2B content strategist
Action: Write a LinkedIn post about AI tools for marketing managers
Context: Audience is busy marketing managers at mid-size companies who are skeptical of AI hype but open to practical tools
Execute: Write the post now
Output quality: Decent but noticeably flatter. The model produced a post that was technically correct but generic. It defaulted to a bullet list format despite the context implying LinkedIn-style prose. No tone specification meant it hedged toward neutral corporate language.
Weakness: RACE gives you role and context but skips format and tone, which are often the difference between usable and unusable output. The "Execute" step doesn't add anything — it's just a directive.
CLEAR prompt
The CLEAR framework applies as criteria rather than a template:
Write a LinkedIn post about AI tools for marketing managers.
Apply these criteria to your prompt interpretation:
- Concise: interpret my ask without padding
- Logical: structure the post with a clear narrative arc
- Explicit: don't assume — if something is ambiguous, make the most reasonable choice and tell me
- Adaptive: write for a LinkedIn audience, not a blog audience
- Reflective: after writing, note one thing you'd change if given more information about the specific audience
Output quality: Interesting. The Reflective criterion produced something genuinely useful — the model wrote the post and then added a note saying "I assumed marketing managers at growth-stage companies; if this is for enterprise marketers the risk framing should shift." That kind of transparency is valuable when you're iterating.
Weakness: CLEAR is less a prompt structure and more a quality checklist. It doesn't tell you what information to include (no role, no format spec), so you still have to supply those separately. It's a complement to other frameworks, not a standalone replacement.
CO-STAR prompt
Context: Marketing managers at mid-size B2B SaaS companies are under pressure to produce more content with smaller teams. AI tools are proliferating but trust is low after hype cycles.
Objective: Write a LinkedIn post that makes marketing managers curious about specific AI tools without overselling.
Style: Conversational but informed — like advice from a respected colleague, not a vendor.
Tone: Honest and grounded. Acknowledge the skepticism. Don't promise ROI.
Audience: Marketing managers, 5–15 years experience, familiar with tools like HubSpot and Figma, skeptical of AI hype.
Response: Single LinkedIn post, 150–200 words, no bullet lists, hook in first line.
Output quality: The best of the four for this task. The explicit Audience dimension did real work — the model referenced "teams that already use Figma for design" and "HubSpot workflows," which made the post feel targeted rather than generic. The Style vs. Tone distinction (both present in CO-STAR, absent in the others) let the model navigate a specific voice: informal register but substantive content.
Weakness: CO-STAR prompts are long. If you're writing many variations or running quick iterations, the overhead adds up. And the Style/Tone distinction, while useful, can be confusing — the line between them isn't obvious.
What the test actually showed
Running the same task through four frameworks revealed something more useful than "which framework wins": it showed which components do the most work.
The components that consistently improved output:
- Audience specificity (CO-STAR's explicit audience field)
- Tone with exclusions ("no hype," "don't promise ROI")
- Format precision (word count, prose vs. bullets)
The components that mattered less than expected:
- The "Execute" step in RACE — pure overhead
- Naming the framework at all — models don't care what you call the structure
The CLEAR framework's Reflective criterion is worth stealing for any framework: asking the model to note its own assumptions produces better transparency and makes iteration faster.
Which framework for which situation
Here's my honest take after running these repeatedly:
Use CRAFT for content creation where you have time to set up the context properly. The structure guides you to specify everything that matters: context, role, format, tone. Works well for blog posts, email copy, social content.
Use RACE for quick, low-stakes tasks where you need something fast and don't care about format perfection. Good for first drafts, brainstorming, quick summarization.
Use CLEAR's reflective criterion in any prompt where you're iterating. Ask the model to surface its assumptions — you'll catch errors faster.
Use CO-STAR for audience-aware writing — anything where knowing who reads it changes what you say. Product launch copy, sales emails, public-facing content. The explicit audience and style fields do real work.
For the fundamentals of what makes any prompt work, the beginner track breaks down prompt anatomy before you add framework structure to it.
Frameworks are training wheels
Here's the honest truth about all of these: after a few months of intentional prompting practice, you stop thinking in framework terms. You internalize the components — role, context, format, tone, audience, constraints — and apply them fluidly without a checklist.
The frameworks are useful when you're starting out because they force you to think through dimensions you'd otherwise skip. Most beginners forget to specify output format. They forget to describe their audience. They forget to set a tone. Frameworks make those omissions obvious.
But "I used CO-STAR for this" is not a meaningful statement. "I specified a 200-word LinkedIn post for skeptical marketing managers with a grounded peer tone" is. The framework is just a mnemonic for getting there. Once you've internalized it, let it go.
Start with CO-STAR if you're new to structured prompting. It covers the most ground. When it starts feeling like a checklist you can complete in your head without the acronym, you've leveled up.
