There is a recurrent question when starting to use LLMs seriously : “How do I structure my prompts ?”.

Several frameworks emerged to answer that question. Let me go through the most common ones, with practical examples.

Why bother with a framework ?

A prompt is just text. Yet, a poorly written prompt yields a poor result. A well-structured one yields a precise, useful one.

The frameworks are just mnemonics. They are a checklist to avoid forgetting important context. Nothing more, nothing less.

CRAFT

Context, Role, Action, Format, Target

The most widely used one. Easy to remember, covers most cases.

Context: Building a 2D platformer in C99
Role:    Senior game engine developer
Action:  Write an AABB collision detection system
Format:  Annotated C99, Linux kernel style
Target:  Embedded / low-level developers

A good default. Start here.

RISEN

Role, Instructions, Steps, End goal, Narrowing

Better suited for multi-step tasks where the order matters.

Role:         Physics programmer
Instructions: Implement sweep-and-prune broadphase
Steps:        1) Define axis-sorted list  2) Find overlapping pairs  3) Output pair buffer
End goal:     Reusable broadphase module
Narrowing:    C99, no dynamic allocation, < 300 lines

The Narrowing part is the most valuable addition over CRAFT. It is where you exclude what you don’t want, which is often as important as what you do want.

CREATE

Character, Request, Extras, Adjustments, Type, Extras1

More creative / generative oriented. The double Extras is intentional : first pass for constraints, second pass for polish.

Character:   Senior graphics programmer
Request:     Implement a sprite batch renderer
Extras:      Must sustain 10k sprites @ 60fps
Adjustments: Vulkan only, no OpenGL
Type:        C99 source file
Extras:      Inline performance notes

The distinction between Extras and Adjustments is subtle but useful : Extras adds, Adjustments redirects.

CARE

Context, Ask, Rules, Examples

Best when you have concrete I/O examples. Closest to a proper specification.

Context:  Tile-based binary map loader
Ask:      Parse format into a struct
Rules:    No malloc, fixed buffers, C99 only
Examples: [binary sample] → [expected struct output]

The Examples part is the killer feature here. An LLM guided by examples makes far fewer wrong assumptions.

CLEAR

Concise, Logical, Explicit, Audience, Role

A meta-framework of sorts. Less about structure, more about quality attributes of each component.

Concise:   "Write a ring buffer"
Logical:   "for audio streaming, SPSC"
Explicit:  "thread-safe via atomics, no locks"
Audience:  Embedded C developers
Role:      DSP systems programmer

Useful as a review checklist once the prompt is written rather than as a writing guide.

Comparison

Framework Sweet spot
CRAFT General purpose, good default
RISEN Multi-step, ordered tasks
CREATE Creative / generative output
CARE Data transformation, precise I/O
CLEAR Quality review of existing prompt

In practice

No need to pick one and stick to it forever. CRAFT covers ~80% of daily use. Reach for RISEN when the task has sequencing constraints. Use CARE when you can supply examples — it gives the most deterministic output.

The frameworks are not mutually exclusive either. A CRAFT prompt can embed CARE’s Examples without issue.

The real takeaway is : always include what the output is for and who reads it. That single addition — Target in CRAFT, Audience in CLEAR — has the biggest impact on output quality, and is the most commonly forgotten element.2

  1. The second E in CREATE varies by source. Some use Evaluation, some use Extras. The meaning differs slightly but the intent is the same : a second pass for refinements. ↩

  2. This article was written with Claude. Asking an AI to explain how to talk to an AI is either the most efficient use of a prompt framework, or a remarkably short loop. ↩