What does CRAFT stand for in prompting?

Context, Role, Action, Format and Tone — the five elements that turn a vague request into a prompt that produces usable output.

What are the common prompt engineering acronyms?

CRAFT, RTF (Role, Task, Format), RACE (Role, Action, Context, Expectation) and CO-STAR (Context, Objective, Style, Tone, Audience, Response) are the most used. They describe the same core ideas, so teams should standardise on one.

How many times should I iterate on a prompt?

Two or three focused refinements — narrow, critique, extend — are normally enough. If a prompt works twice, save it as a template.

Prompting

The CRAFT Prompt Framework: Five Parts of a Prompt That Works

Q: What are the parts of a good prompt?

Background and situation, the role the model should take, one concrete task, the required output format, and the tone of voice. Anything you leave out is a gap the model fills with a guess.

Q: Why do my prompts give generic answers?

Usually because context and format are missing. Add the situation, audience and constraints, and state the exact shape of the answer you expect.

Devdiscourses Online · September 3, 2026

The CRAFT Prompt Framework: Five Parts of a Prompt That Works

Context, Role, Action, Format, Tone — a repeatable structure for prompts, with a worked example, common mistakes, and how to turn a good prompt into a reusable template.

A prompt is a structure, not a sentence.

Ask a model to "write a summary" and you get generic output. Ask with structure and you get something you can send. CRAFT is the structure.

The five elements

Context — the situation, the audience, the constraints, the background material. Most weak prompts fail here.
Role — who the model should act as. "You are an experienced HR business partner" narrows the vocabulary and the assumptions.
Action — the single, concrete task. One verb: summarise, compare, draft, critique, restructure.
Format — the shape of the answer: length, sections, bullet points, table, JSON, word limits.
Tone — plain, formal, warm, direct. Tone is what makes output sound like your organisation instead of a chatbot.

A worked example

Weak prompt: Write an email about the new AI policy.

CRAFT prompt:

Context: We are a 120-person marketing agency rolling out an approved-tools AI policy on 1 October. Staff are anxious about job security. Role: You are an internal communications lead. Action: Draft the announcement email. Format: Under 250 words, a subject line, three short sections, and one clear next step. Tone: Calm, concrete, no hype.

The second prompt produces something an editor tweaks in two minutes. The first produces something you rewrite.

Iteration beats perfection

Even a good prompt rarely lands first time. Use three moves:

Narrow it. "Cut this to 150 words and remove every adjective."
Critique it. "List three weaknesses in this draft and what a sceptical reader would object to."
Extend it. "Now produce two alternative versions: one shorter and more direct, one warmer."

Common mistakes

Asking for several unrelated tasks in one prompt.
Providing no example of what "good" looks like — one sample output usually beats three paragraphs of instructions.
Leaving format implicit, then complaining about the structure.
Never saving what worked. A prompt used twice should become a template with [placeholders].

Turn prompts into assets

Store working prompts where the team can find them, with the placeholders visible and a note on what output to expect. That single habit is the difference between individual experiments and an organisation that actually gets faster.

Try the CRAFT scoring exercise inside the AI coach, or browse ready-made role-based prompts in the prompt library.

How CRAFT compares to other prompt frameworks

People search for "prompt engineering acronyms" because there are a dozen of them and they mostly describe the same few ideas. RTF (Role, Task, Format), RACE (Role, Action, Context, Expectation), CO-STAR (Context, Objective, Style, Tone, Audience, Response) and CRAFT all exist to stop you writing a one-line request.

What they share is the useful part: say who the model is, what exactly it must do, what it needs to know, and what the answer should look like. CRAFT keeps five elements because that is the smallest set most people can recall without a cheat sheet, and it puts Context first because that is where weak prompts actually fail. Pick one framework and use it everywhere — switching between acronyms is how teams end up with no standard at all.

Frequently asked questions

What does CRAFT stand for? Context, Role, Action, Format, Tone.

What are the parts of a good prompt? The situation and background, the role the model plays, one concrete task, the required output shape, and the tone of voice. Anything missing is a gap the model fills with a guess.

Why do my prompts give generic answers? Almost always missing context and format. Add the audience, the constraints and the exact shape of the answer.

Is prompt engineering a real skill? Yes — it is structured, teachable and measurable, which is why it is assessed in our prompt engineering module. See also what prompt engineering is.

How many prompt frameworks should a team use? One. A shared structure plus a saved template library beats everyone inventing their own.

Related reading: What is prompt engineering? for the wider picture, the SAFE framework for using AI on real company data without exposure, and What is Titan OS? for an example of reading a technology decision the way this course teaches.

Also useful: How to use AI at work and Is ChatGPT safe to use at work?.

Keep reading

Prompt engineering frameworks compared: CRAFT, RTF, RACE, CO-STAR, PARTS