The CRAFT framework:
AI prompts that work
on the first try

Most people who "use ChatGPT" type vague questions, get vague answers, and conclude AI is overrated. The tool is not the problem. The way you talk to it is. The CRAFT framework is the prompt structure we teach at Revolia, a French AI training organization: Context, Role, Action, Format, Tone. Five blocks, in that order, and the quality of the output changes on the spot.

In short: CRAFT is a five-step prompt engineering framework (Context, Role, Action, Format, Tone) that turns any instruction you give a generative AI (ChatGPT, Claude, Gemini) into a proper brief, so the first answer is one you can actually use.

Why most prompts fail

Most prompts fail because they read like a vague question instead of a precise brief: the AI cannot guess your context or the expected format, so it produces a generic answer. Take a classic prompt: "Write me an email to follow up with a client."

The AI complies. It produces a generic email: "Dear client, I hope this message finds you well...". You read it, find it flat, give up and write the thing yourself.

The model is not the weak link here. The request is. A good prompt is a brief, not a question.

You don't brief an intern by saying "make me an email". You give them the context, the role, the task, the expected format and the tone. Same with AI.

CRAFT: the 5 blocks of a prompt that works

A CRAFT prompt chains five blocks in order: Context sets the scene, Role says who the AI plays, Action states the task, Format defines the deliverable and Tone frames the style. Here is each block with a concrete example.

C: Context

Start by setting the scene. Who you are, who you work for, what the situation is. The AI knows nothing about your world: give it what a colleague would already have in mind.

Example: "I'm a sales manager at a 50-person company selling HR SaaS to mid-market HR directors. A prospect, Sophie (HR director at a 300-person company), had a demo 3 weeks ago and has ignored 2 follow-ups since."

R: Role

Tell the AI who to be. An expert? A critical reviewer? A customer? The role shapes the vocabulary, the posture, and the blind spots it will watch for.

Example: "Act as a senior salesperson with 15 years of experience in complex B2B sales."

A: Action

The task, as precisely as possible. Not "help me with" but "write", "analyze", "compare", "list", "critique". One verb, one deliverable, one goal.

Example: "Write a follow-up email aimed at getting a 15-minute call."

F: Format

This is the block people skip most, and it costs them the most. What exactly do you want? An 80-word email? A table? A bulleted list? A 3-part outline? Make the format explicit.

Example: "Format: subject line + 3 short paragraphs + 1 clear CTA. 100 words max. No opener like 'I hope you're doing well'."

T: Tone

Polite, direct, warm, factual, salesy, understated? Tone changes everything. Spell it out, and if you can, add a counter-example of the tone to avoid.

Example: "Tone: professional, warm without being familiar. Strictly avoid: canned sales language and empty courtesy phrases."

CRAFT in action: before / after

The same need, following up with a silent client, produces two opposite results depending on whether you improvise or use CRAFT.

Weak version

"Write me an email to follow up with a client who hasn't replied."

Result: a generic email, "Dear [client]", filler phrases, no specific angle.

CRAFT version

Context: "I'm a sales manager at an HR SaaS company. Sophie, HR director at a 300-person group, had a demo 3 weeks ago, then silence. Her 2 topics were: simplifying annual reviews and cutting HR time spent on recruiting."

Role: "Act as a senior B2B salesperson with 15 years in complex sales."

Action: "Write a follow-up email aimed at getting a 15-minute call."

Format: "Catchy subject (5-7 words) + 3 short paragraphs + 1 clear CTA. 100 words max."

Tone: "Professional, warm, direct. No 'just checking in', no filler openers. One concrete word about one of her 2 topics in the first line."

Result: a specific email that names her actual stakes, proposes a short call, and sounds human. Ready to send after a 10-second read.

The most common mistakes

Four mistakes come up in every training session: stuffing everything into Action, skipping Format, staying vague on Tone, and giving no context at all.

Stuffing everything into "Action"

Beginners write sandwich prompts: one 15-line block that mixes everything. The AI gets lost in it. Separate the blocks visually (line breaks, bullets, CRAFT headers) so each ingredient stands out.

Skipping the Format

If you want 80 words, say so. If you want a markdown table, say so. If you want "no introduction, go straight to the content", say so. The AI delivers literally what you ask.

Staying vague on Tone

"Professional tone" means nothing. Add a counter-example: "not corporate, no empty phrases, no marketing buzzwords". You save two iterations.

Asking without context

Context is half the quality. "Write an email" with no context guarantees a generic result. Paste a previous example, a note, a brief, a transcript, anything that anchors the request in your reality.

The CRAFT template to copy

Here is the CRAFT skeleton, ready to use: paste it into your prompt library and replace the brackets with your situation.

[CONTEXT]
I am [role] at [company / industry].
Situation: [describe precisely].
Relevant history: [what happened before].

[ROLE]
Act as [expert / persona]. Bring the posture and vocabulary of [X].

[ACTION]
[Imperative verb: Write / Analyze / Compare / List / Critique...]
Concrete goal: [what I want to end up with].

[FORMAT]
Structure: [outline / table / email / list].
Length: [X words / X lines / X bullets].
Constraints: [musts / must-nots].

[TONE]
Expected tone: [adjectives].
Strictly avoid: [counter-examples].
        

The real next level: CRAFT + examples

To go from "good result" to "calibrated result", add one or two examples of what you expect. This is few-shot prompting, and one concrete example beats paragraphs of explanation.

In practice, you paste one input → expected output pair, then give the new case to handle. The AI imitates the pattern it just saw: short subject line, opener on the prospect's topic, social proof, precise time ask.

One example is enough most of the time. Move to two or three only when the pattern is subtle (a very particular tone, strict formatting, a sorting logic) and one example doesn't get it across. Beyond that, you lengthen the prompt for little gain.

CRAFT, RACE or CO-STAR?

RACE (Role, Action, Context, Expectations) and CO-STAR (Context, Objective, Style, Tone, Audience, Response) aim at the same thing as CRAFT: structuring a prompt instead of tossing a question. CRAFT stands out with its dedicated Format block; RACE is the shortest to memorize; CO-STAR isolates the audience, which helps when the same content targets several readers.

CO-STAR has a track record: data scientist Sheila Teo used it to win Singapore's first GPT-4 prompt engineering competition. But the framework matters less than the habit. Pick the one you'll remember and apply it to every serious prompt. A structure you actually use beats a perfect one you forget.

When CRAFT is not enough

CRAFT structures a single instruction. Some jobs need more than one.

• Multi-step (agentic) tasks. When the AI has to fetch information, process it, then act on it, decompose the work into steps, each with its own prompt.
• Complex reasoning. On problems that require step-by-step thinking, explicitly ask the model to lay out its reasoning before concluding. CRAFT frames the request; it doesn't replace that instruction.
• Long, reusable system prompts. For an assistant replayed dozens of times (a custom GPT, a Project), the prompt becomes a document: rules, guardrails, examples, edge cases. CRAFT is its skeleton, not its final form.
• Guardrails. On sensitive topics (HR, legal, customer data), add what the AI must not do: never invent figures, flag uncertainty, leave human decisions to humans.

CRAFT is the core method of Revolia's AI training programs, delivered in French. If you read French, the original version of this guide is here: la méthode CRAFT pour structurer ses prompts.