Uplevel your prompt craft in ChatGPT with the CREATE framework

This post explores the various components of crafting high-quality prompts for different Artificial Intelligence (AI) tools like DALLE-2 and ChatGPT. I share the CREATE framework to communicate best practices and critical guidelines. The framework aims to help people write better prompts and improve their prompt craft skills.

What is prompt craft?

Prompt crafting refers to the process of creating a specific prompt or set of prompts for a language model like GPT to generate text from. It can involve carefully selecting words, phrases, and even entire sentences to guide the model in a certain direction and produce a desired output. This can be used for a variety of applications, such as writing creative fiction, generating code, or answering questions.

First unedited ChatGPT output from prompt: What is prompt craft?

Although many experts believe that prompt craft is an awkward stage we will overcome with better-trained models and technology, at this moment, the quality of our input directly impacts the quality of the output.

When I say quality output, I also mean alignment with our idea of success. You get better results with better inputs; that is why we should all pay attention to improving our prompt craft.

I have used generative AI tools since 2021 and expanded my prompt craft in various media tools, like image and art generators: DALLE-2, Midjourney, text generators: Jasper and OpenAI’s tools: ChatGPT and GPT-3 Playground.

I am also leveraging generative AI capability within existing platforms. I use Notion and Mem. This changes the experience to an inline prompt, subtly different from engaging with a chat interface or the structured exchanges of the other tools mentioned.

With the ability for developers to plug into the ChatGPT tool, we, as educators and creators, will experience more inline AI as the capability gets layered into our toolset.

I have inputted thousands of prompts across these tools, and the CREATE framework for prompt craft captures all of the critical lessons I have learned.

CREATE Framework for High Quality Prompt Craft in AI Tools

C	Clarity	Clearly define the task or intent of the prompt, including specific information about the output.
R	Relevant info	Provide relevant details, including specific keywords and facts, the tone, audience, format and structure.
E	Examples	Use examples in the prompt to provide context and direction for the output.
A	Avoid ambiguity	Focus on the key information and delete unnecessary details in the prompt.
T	Tinker	Test and refine the prompt through multiple iterations. Explore different input versions to discover the best results.
E	Evaluate	Continuously evaluate the output and adjust the prompt as needed to improve the quality.

The acronym “CREATE” is a helpful guide for crafting high-quality prompts for AI tools. Each letter represents an important step in the process.

The first four CREA are all part of prompt writing, where TE, the final two are a cycle of reviewing and editing your prompts.

Let’s look at each in more detail, with some examples from ChatGPT to help.

Clarity

Clearly define the task or intent of the prompt, including specific information that the AI tool should use to generate the output. This directs the tool to generate an accurate and relevant output.

Instead of LESSON PLAN ON Y5 FRACTIONS, use:

Plenary and formative assessment tasks for lesson plan, equivalent fractions, Year 5

Plenary assessment tasks for a lesson on equivalent fractions in Year 5 could include:

• A written quiz or test on the concepts covered in the lesson, such as identifying equivalent fractions and simplifying fractions
• A problem-solving activity in which students must apply their understanding of equivalent fractions to real-world situations
• A group discussion or presentation in which students explain and demonstrate their understanding of equivalent fractions to their classmates

Formative assessment tasks for the same lesson could include:

• A “exit card” where students write down one thing they learned in the lesson and one question they still have
• A “turn-and-talk” where students discuss what they learned with a partner and then share out to the class
• A formative assessment task that can be done during the lesson, such as a quick quiz, or a mini-activity that checks for understanding
• A self-assessment task where students reflect on their own understanding of the concept and identify areas they need to work on.

First unedited ChatGPT output

Instead of LETTER TO PARENTS JOINING IN 2023, use: Short friendly letter introduction to parents joining in 2023

Instead of PERSUASIVE TEXT ABOUT AI, use: Two persuasive paragraphs about teachers learning AI prompt writing skills

Blocking or Chunking for Clarity

Unlike image and artwork tools, Midjourney, Stable Diffusion and DALLE-2, writing tools allow us to block the task into chunks. Along with a precise overall aim, you can use phrases like:

• Start with a short opening paragraph
• List 3 versions of a closing sentence
• Continue with the next paragraph in the same tone

This is a much more straightforward way to approach longer-form work, and in my experience always leads to a better outcome.

You also get the ability to work in more flexible ways with chunks rather than having to rework longer sections.

Relevant info

Provide relevant and concise information, including keywords and facts, to guide the AI tool in generating accurate and appropriate output.

Consider expanding on some of the following components: Audience, Format, Tone, Length and Style.

Elaborate in specific detail about the task or intent of your prompt. Include anything that is relevant and helps with precision and clarity.

Let’s look at the parent letter from earlier:

Instead of LETTER TO PARENTS JOINING IN 2023, use:

Short letter introduction to parents joining Bark Street Primary School in 2023, children in upper school, first contact about teaching and learning details, first day for students Wednesday 1st February 2023.
Length: 100 words.
Tone: engaging, upbeat and friendly
Keywords: excitement, friendships, learning, inquiry, fun, summer, connection.
Style: Concise, professional

“Welcome to Bark Street Primary School! We are thrilled to have your children join our Upper School community starting on Wednesday, February 1st, 2023. We believe that learning should be an exciting and inquisitive journey, filled with friendships, fun and connection. Our dedicated teachers are committed to fostering a love of learning and inquiry in every student. We can’t wait to see the amazing things your children will accomplish this year. Summer may be winding down but the excitement and opportunities for growth at Bark Street are just beginning. We look forward to getting to know you and your children. Let’s make this a great school year together!”

First unedited ChatGPT output

Examples

We can add examples to guide the output along with increased clarity and including relevant information.

Example content is helpful when you are trying to generate content in a specific tone or similar to your previous work.

Sometimes it is easier to show an example of what you are looking for rather than describe every aspect and component of what you are trying to create. You offer clear guide rails when you do both.

This provides a context in the input that changes the quality and alignment of the output.

Instead of WRITE A BLOG POST INTRODUCTION EXPLAINING THE MENTAL MODEL CHESTERTON’S FENCE, use:

Write a blog post introductory paragraph, explaining the mental model chesterton’s fence, 100 words.

Match the tone and style to this text: “Oh, you have finished quickly! There is nothing like the feeling you get when you realise you have underplanned a lesson, and there is still the best part of 40 minutes still to go with 35 six-year-olds. My map was the carefully prepared, beautifully presented lesson plan (external) and a thousand priorities I was holding in my head for teaching in the early years (internal).”

Have you ever found yourself questioning the reasoning behind a decision or rule, only to realize that the purpose behind it is not immediately obvious? This is where Chesterton’s Fence comes into play. Named after the English writer, G.K. Chesterton, this mental model highlights the importance of understanding the reasons behind established practices before attempting to change or remove them. It reminds us to consider the history and context behind a decision before making changes, rather than blindly following or disregarding it without understanding its purpose. In the classroom, this can mean taking the time to understand the reasoning behind a lesson plan or teaching method before making any adjustments.

This one is a subtle output variation you need to explore yourself, but adding a template or example is helpful. Try it with writing exemplars or something you have previously published.

Avoid ambiguity

A key improvement we all can make with our prompt craft is to be more concise. Avoid ambiguity by deleting anything extraneous to the intention.

Focus on the key information and delete unnecessary details in the prompt.

I see many resources, social media posts and prompt craft downloads with so much unnecessary ambiguity. Deleting extraneous text is also a great way to tinker (see the next section) with your prompt.

Much of my concision, I think, comes from prompting in Midjourney, where there is no human-like interface or chat protocol. We all get politely sucked into anthropomorphising ChatGPT as we say please and thank you.

As Ethan Mollick points out tools like Midjourney lack this interface confusion:

“This confusion isn’t a problem in other types of generative AI. Since AI image generators are not as verbose as ChatGPT, people using them have learned they need to spend time doing “prompt engineering:” editing and playing with prompts to come up with the results they want.“

Let’s improve a prompt he uses to illustrate some of the complexity we have expanded on in this post. Here is his comparison of an essay prompt.

And the second prompt he uses:

write an essay with the following points. use an academic tone. use at least one clear example. make it concise. write for a well-informed audience. use a style like the New Yorker. make it at least 7 paragraphs. vary the language in each one. end with an ominous note. -Humans are prone to error -Most errors are not that important -In complex systems, some errors are catastrophic -Catastrophes cannot be avoided

Let’s review this for concision and to avoid unnecessary ambiguity or input.

write an essay with the following points. use an academic tone. use at least one clear example. make it concise. write for a well-informed audience. use a style like the New Yorker. make it at least 7 paragraphs. vary the language in each one. end with an ominous note. -Humans are prone to error -Most errors are not that important -In complex systems, some errors are catastrophic -Catastrophes cannot be avoided

Much of what I have deleted is a natural part of our expression and includes many references to actions.

All the AI tool is doing is predicting the next word and generating text – perhaps we don’t need to continue asking it to write, use, or make.

Here is the output from the edited prompt.

My experiment in concision shows there is little difference when we remove those verbs, increase brevity, and avoid ambiguity.

Tinker

One of the lessons you learn from working with image and art generators is to tinker and reroll. Rerolling is just to regenerate the output.

In ChatGPT, there is an option to do this under each output: Regenerate response.

Get into the habit and mindset to test and refine the prompt through multiple iterations, adjusting the input and analysing the output to ensure high-quality results. This step, which we refer to as prompt engineering, helps to fine-tune the prompt to get closer to the desired outcome.

Your edits can come in different forms: concision, reordering, expansion, examples etc.

Here are two quick examples, use the slider to compare.

*All the text features!

Evaluate

The final aspect of high-quality, prompt craft is a continuous evaluation of the output. With images, it can often feel much clearer, which result works for you.

For text generation, we have more work to review and adjust the prompt as needed to improve the quality of the output. That is why chunking or blocking text into sections is so helpful and easier to evaluate.

Moving slowly, despite a desire for speed, ensures that the output meets your desired quality and the task or intent of the prompt.

Here is another side by side:

This step concerns your original intention and adapting the prompt and output to align with your goals. This is what we are evaluating.

I hope the CREATE Framework helps to bring some structure and a little checklist to the process of prompt craft.

Let me know in the comments what resonates, any changes you would make, and what you think makes the most significant difference to the quality of your prompts.

---

The AI tools I used to develop this post were OpenAI’s ChatGPT which is trained on the LLM GPT-3.5 and Grammarly which is trained on user input and fixed grammatical / spelling rules.