The Three ChatGPT Prompt Frameworks Worth Memorizing
CRAFT, RTF and CARE — what they actually mean, when each one fits, and the situations where ignoring all three is the right call.
On this page 4
Prompt "frameworks" are mostly mnemonics for stuff you'd do anyway if you slowed down. But three of them actually pay rent. Memorize these and you can stop bookmarking Twitter threads.
CRAFT — Context, Role, Action, Format, Tone
The most complete framework. Use it when the output quality matters and you can spare 30 seconds to write a real prompt.
- Context — the situation, audience, background.
- Role — who the model is playing.
- Action — the specific task.
- Format — bullets, table, JSON, length.
- Tone — voice, register, energy.
"Context: I run a B2B newsletter for ops leaders. Role: you are an editor with 10 years of experience. Action: pick the strongest of these 3 subject lines. Format: a short verdict + one sentence why. Tone: blunt, no hedging."
RTF — Role, Task, Format
The 80/20 version. Use it for quick, repeated requests where CRAFT is overkill. Most chat sessions should live here.
"You are a code reviewer. Find the bug in this function. Reply with: line number, problem, one-line fix."
CARE — Context, Action, Result, Example
Best when you need a specific shape of output. The example does most of the work; the rest just primes it.
"Context: I'm writing an FAQ page. Action: turn this paragraph into a Q&A. Result: 3 Q&A pairs. Example: Q: How long does shipping take? A: Usually 3-5 business days."
When to skip frameworks entirely
For exploratory questions ("brainstorm 20 angles for…"), frameworks hurt. They narrow the model when you want it wide. Save the structure for tasks with a right answer; leave the open-ended ones loose.
More to read
Keep reading
How to Write Midjourney Prompts That Actually Look Cinematic
Six concrete levers — subject, lens, lighting, color, composition and rendering — that turn a generic Midjourney output into something worth printing.
Five Prompt Patterns That Quietly Outperform Everything Else
Role priming, constraints-first, worked examples, chain-of-thought and "rate then revise" — when each one wins and a copy-paste template for each.