The CRAFT Prompting Framework: How to Write Better AI Prompts

You've probably discovered that the quality of the response you get from AI often depends heavily on the quality of the instructions you give it.

Ask a vague question and you'll often get a vague, generic answer. Give the AI clear context, explain what you need and describe how you want the answer presented, and the results can be dramatically better.

That's where prompting frameworks such as the CRAFT prompting framework can help.

CRAFT provides a simple structure for creating more effective AI prompts. Rather than trying to remember the perfect words to use, it encourages you to give an AI tool the information it needs to understand your situation, the role you want it to take, what you want it to do and what a good response should look like.

What is the CRAFT prompting framework?

CRAFT is a framework for structuring prompts when working with generative AI tools such as ChatGPT, Microsoft Copilot, Claude and Gemini.

The acronym commonly represents:

• C: Context
• R: Role
• A: Action
• F: Format
• T: Tone

Instead of simply asking AI a question, CRAFT encourages you to think about the information the AI needs before producing its response.

You don't necessarily need to use every element in every prompt. For a simple request, a short instruction may be perfectly adequate. CRAFT becomes particularly useful when the task is more complex or when the quality, style and structure of the output matter.

C: Context

Context gives the AI the background information it needs to understand your request.

One of the most common mistakes when using AI is assuming that it understands your situation without being told. Although an AI model may have extensive general knowledge, it doesn't automatically know the circumstances surrounding your particular task.

For example, rather than saying:

"Give me some ideas for improving customer service."

You might provide context such as:

"We are a UK-based software company providing accounting software to small businesses. Our customer service team handles most enquiries through email and live chat. Recent customer feedback suggests that responses are accurate but sometimes feel impersonal."

The second prompt gives the AI considerably more to work with.

Useful context might include your organisation, audience, objectives, current situation, challenges, constraints or information that the AI should take into account.

R: Role

Role describes the perspective, expertise or position you want the AI to adopt when approaching the task.

For example:

"Act as an experienced learning and development consultant."

Or:

"Approach this as a marketing specialist with experience of selling professional services to UK businesses."

Giving the AI a role can help establish the perspective from which it should approach the task.

However, roles should be useful rather than decorative. Telling an AI to be a "world-leading, internationally renowned, genius marketing expert" rarely adds anything meaningful. A specific role connected to the task is usually much more effective.

A: Action

Action is the actual task you want the AI to complete.

This should be as clear and specific as possible.

Useful action words might include:

• Analyse
• Compare
• Summarise
• Rewrite
• Create
• Identify
• Explain
• Evaluate
• Generate
• Recommend

For example, instead of:

"Help me with this report."

You could say:

"Review the report and identify the five most important findings for the senior management team. Explain why each finding matters and identify any areas where the evidence doesn't support the conclusions."

The clearer you are about the action required, the less the AI has to guess.

F: Format

Format tells the AI how you want the response structured or presented.

This is particularly useful because the best answer isn't always a long block of text.

You might ask for the response as:

• A table
• A bullet-point list
• A step-by-step guide
• An email
• A report
• A meeting agenda
• A checklist
• A social media post
• A presentation outline
• Structured data such as JSON

You can also specify length and structure.

For example:

"Present the answer as a table with columns for the issue, potential impact, recommended action and priority. Include no more than eight recommendations."

This saves time because you're telling the AI what a usable answer looks like rather than asking it to generate something and then repeatedly reformatting the response.

T: Tone

Tone describes how you want the response to sound.

This can be particularly important when creating content that will be read by other people.

You might ask for a tone that is:

• Professional
• Friendly
• Conversational
• Confident
• Empathetic
• Persuasive
• Informal
• Technical
• Plain English

Again, specificity helps.

Rather than simply asking for a "professional" tone, you could say:

"Use a professional but conversational tone. Avoid corporate jargon and overly formal language. Write as though you're explaining the subject to an intelligent colleague who isn't an expert."

That gives the AI a much clearer indication of what you're looking for.

A CRAFT prompt example

Let's put the complete CRAFT prompting framework together.

Imagine you've been asked to create an email encouraging managers to attend a leadership development programme.

A basic prompt might be:

"Write an email promoting our leadership course."

It will probably produce something usable, but the AI has to make a lot of assumptions.

Using CRAFT, you might write:

Context: We are launching an internal leadership development programme for employees who have recently moved into their first management role. Many are technically experienced but have received little formal management training.

Role: Act as an experienced learning and development communications specialist.

Action: Write an email introducing the programme and encouraging new managers to register. Focus on the practical benefits, including handling difficult conversations, delegating effectively, motivating team members and managing performance.

Format: Write an email of approximately 250 words with a short opening, three key benefits and a clear call to action.

Tone: Friendly, encouraging and down to earth. Avoid corporate jargon and don't make the programme sound like remedial training.

The second prompt gives the AI far more useful information about what you're actually trying to achieve.

Why does the CRAFT prompting framework work?

CRAFT works because it reduces ambiguity.

Generative AI is remarkably good at interpreting instructions, but it still has to work with the information you provide. When information is missing, the model may need to make assumptions about what you intended.

Sometimes those assumptions are correct. Sometimes they're not.

Providing context, role, action, format and tone gives the AI clearer boundaries within which to work.

It also encourages the person writing the prompt to think more carefully about what they actually need.

That's an often overlooked part of effective prompting. If you can't clearly describe the outcome you want, it becomes much harder for either AI or another person to produce it.

Do you need a prompting framework every time you use AI?

No.

If you want to know how to calculate a percentage or need a quick explanation of a concept, constructing a five-part CRAFT prompt would probably be unnecessary.

Modern AI models are also much better at interpreting natural language and understanding conversational context than earlier systems were. You don't need to treat every interaction with AI like you're programming a computer.

Prompting frameworks are most useful when the task is important, complex or likely to produce a generic response without additional direction.

Think of CRAFT as a checklist rather than a rigid formula.

If the AI isn't giving you the result you need, ask yourself:

• Have I provided enough context?
• Would a specific role help establish the right perspective?
• Have I clearly explained the action I want completed?
• Have I described the format I need?
• Have I explained the appropriate tone?

Often, one of those elements is what's missing.

CRAFT is a starting point, not the end of the conversation

Perhaps the biggest mistake people make when learning about prompting is believing that they need to create one enormous "perfect prompt" that produces the perfect answer first time.

That's rarely necessary.

One of the biggest advantages of generative AI is that you can have a conversation with it.

Use CRAFT to create a strong initial prompt, review the response and then refine it.

You can ask the AI to expand an idea, challenge its own recommendations, simplify an explanation, provide alternatives or ask you questions before continuing.

For more complex tasks, you can even ask the AI what additional information it needs from you to produce a better result.

Effective prompting isn't really about discovering secret words or complicated formulas. It's about communicating clearly with AI.

The CRAFT prompting framework simply gives you a practical structure for doing exactly that.