💥 Part 1 | ChatGPT Canvas: The AI Tool You Didn’t Know You Needed (While Top Creators Did)

Welcome to OpenAI Canvas Pro

Let me start by saying how excited I am about OpenAI Canvas. Over the years, I’ve worked with dozens of tools for writing and coding, but nothing has felt quite as interactive and efficient as this. If you’re looking for a way to collaborate with AI like you would with a teammate, you’re in the right place. Since 2022, when AI tools really started taking off, ChatGPT has been the most well-known name in the game. It's so dominant that people often mix up “generative AI” and “ChatGPT,” just like how people say "Kleenex" for tissues.

Now, OpenAI, the company behind ChatGPT, has released a new version of the tool. And I’m not just talking about an updated model. ChatGPT Canvas is a new spin on the regular version, letting you interact with ChatGPT in a more hands-on way to create exactly the type of text you need. But what does that mean for you?

That’s one of the things we’ll be exploring on this playbook. With OpenAI Canvas, you’ll learn how to write and code more effectively. Features like targeted editing, in-line feedback, adjusting reading levels, polishing your work, reviewing code, and fixing bugs make it easier to refine your text and projects.

You’ll also get hands-on experience with practical use cases, like building a game app, generating Python code from plot screenshots, and creating SQL databases from architecture images.

Plus, you’ll understand how GPT-4 was trained to power all of these features in OpenAI Canvas. In this short playbook, you’ll:

• Learn how to get in-line feedback and control the changes in your creative work by directly editing specific parts of your text or code in the model’s output.

• Discover how to use quick automation tools through a shortcut menu that lets you adjust your writing tone, modify length, improve your code, and revert to earlier versions of your work.

• Learn how to use Canvas as a research assistant by asking the model to analyze a plot screenshot and write a research report, with the ability to ask questions within the report.

• Extend its capabilities by having the model write Python code to replicate the graph from a screenshot.

• Explore how to build a video game, like Space Battleship, from scratch, edit it, and display it all in one HTML file.

• See a real-world example of creating a SQL database from an architecture image.

• Understand how the model training and design behind Canvas works.

You might be wondering, “Is this for me?”. Here’s how I see it:

• Writers: If you’re creating blog posts, course content, or reports, Canvas can refine your drafts, adjust tone, and even structure ideas better.

• Coders: For developers like you and me, it’s like having a code reviewer always ready to debug, optimize, and even add documentation.

• Creators in General: Anyone who juggles ideas, words, or code will find Canvas a game-changer.

For example, when I was working on the NewsletterAZ Course, I used Canvas to help create templates and polish my communication. The tone suggestions alone made it easy to strike the right balance between conversational and professional.

Are you ready to transform the way you write and code? Let’s dive into the details in the upcoming sections!

I. What is ChatGPT Canvas?

While the standard ChatGPT interface is effective for generating ideas, it can be limiting for tasks that require iterative editing and collaboration.

OpenAI describes Canvas mode as a new way to collaborate with ChatGPT for writing and coding tasks. While I’m not a coder myself, I do know my way around writing, and I’ve been using Canvas to help craft copy for various courses and playbook.

OpenAI Canvas is a tool that brings writing and coding together into one collaborative workspace. It is a new interface for ChatGPT that enhances the original by adding interactive features. It lets you edit the content ChatGPT generates directly while also allowing you to request edits in real time as you work. This makes the content creation process more flexible and engaging.

Canvas uses GPT-4o, a specially trained AI model, to understand what you’re working on and offer real-time suggestions. You and the AI share a side-by-side workspace. On one side, you see your project—an essay, a piece of code, or even a research report. On the other, Canvas gives feedback, edits, or alternative ideas, so you never feel stuck.

I have a good news: it’s now available to all users. It used to be exclusive to Plus and Teams users, but now anyone with access to ChatGPT can use it. Here’s what else we know: Canvas mode works on the web version of ChatGPT, so it’s not available on the mobile app or desktop app just yet. For now, you’ll need to use it through the web version.

II. Key Capabilities to Know

Canvas makes both writing and coding with AI more fun and a lot easier. For writing, it allows you to highlight specific sections of your text for targeted edits. ​You can also adjust the length and complexity of your text. Additionally, Canvas provides tools to apply grammar checks and clarify enhancements. These iterative features make writing more flexible and efficient with Canvas.

Here are some handy writing shortcuts you can use with ChatGPT:

• Suggest edits: ChatGPT provides inline suggestions and feedback to improve your text.

• Adjust the length: You can make the document shorter or longer based on your needs.

• Change reading level: Adjust the reading level, whether it's for Kindergarten or Graduate School.

• Add final polish: It checks for grammar, clarity, and consistency to make your writing more refined.

• Add emojis: Adds relevant emojis to highlight points and bring some color to the text.

Lots of people are using AI to help with coding too. Canvas has several tools that can help you create code better and faster. After creating the first version of your code, Canvas can review it and give suggestions for how to improve it. For example, if there's a syntax or logic error in your code, or if your code can be simplified or made faster, Canvas might be able to point this out.

You control the project in canvas. You can directly edit text or code. There’s a menu of shortcuts for you to ask ChatGPT to adjust writing length, debug your code, and quickly perform other useful actions. You can also restore previous versions of your work by using the back button in canvas.

For myself, I'm sometimes a bit lazy about writing comments, and I think Canvas is great at that. It also helps with debugging by letting you add logs. Another feature is the ability to translate code between different programming languages like Python, JavaScript, and Java with just a few clicks. We'll also go behind the scenes and look at what it takes to train the model to create an interface like Canvas.

📌 NOTE: This works with any of the models - whether you're using GPT-3.5, GPT-4, or any of the other versions. It doesn’t matter which one you choose. So it means you can use o1-preview or o1-mini as well.

III. Why Should You Use OpenAI Canvas?

Let me explain why I think it’s a game-changer for you, based on experience and what I’ve seen it achieve for others.

What Makes It Unique?

• Collaboration: It feels like you’re brainstorming with a team, even if you’re working alone.

• Precision: Instead of vague suggestions, it focuses on exactly what you ask for. For example, I once needed to write a research report on the AI Mastery AZ course. Instead of starting from scratch, I uploaded a screenshot of key data. Canvas analyzed it, helped me draft the report, and even suggested a catchy headline. In coding, I used it to design a self-contained HTML file for my game, combining creativity and precision in one place.

• Versatility: Whether you’re writing a marketing email, debugging code for a web app, or creating a SQL database schema, Canvas adapts. I know you’re busy. You want tools that make your life easier, not harder.

• Save Time and Get More Done: Time is your most valuable resource, and OpenAI Canvas helps you save a lot of it.

• Perfect for Beginners and Experts: You don’t need to be a pro to use Canvas. If you’re just starting out, Canvas helps you learn. It doesn’t just fix problems; it explains why they exist. If you’re experienced, you’ll love how it lets you focus on creative problem-solving instead of repetitive tasks.

Who Should Use It?

If you:

• Write reports, newsletters, or content (like the "Viral Automated AI Replica Playbook").

• Code anything from simple Python scripts to interactive apps like the game example in this playbook.

• Need a reliable assistant to improve your work without adding extra effort.

Then Canvas is perfect for you. The sooner you start using it, the faster you’ll adapt to the tools shaping tomorrow’s workflows. It’s fast, smart, and designed to help you do more with less. That’s why I believe you’ll find it as essential as I do.

IV. How Was OpenAI Canvas Trained?

In this part, we'll go behind the scenes and look at what it takes to train the model to create an interface like Canvas. You learn how Canvas can make your writing and coding a lot better, faster, and more fun. If you're interested in knowing how the model behind Canvas is trained, here we're going to have a very brief overview of the model development steps.

This part is optional, you can read it or not, but I know for some people, it would be interesting.

1. Training the Model to Become a Collaborator

So, OpenAI trained GPT-4o to collaborate as a creative partner. It knows when to open Canvas, make specific edits, and even rewrite content when needed. It also understands the broader context of your work to provide more accurate feedback and suggestions for your document.

One of the key innovations in Canvas is that all post-training was done using synthetic data. They also used techniques like distilling outputs from OpenAI’s o1-preview model to teach the model its core behaviors. This method helped the model quickly improve writing quality and adapt to new ways users interact with the tool. All without relying on human-generated or human-collected data.

OpenAI’s research team focused on developing core behaviors for this model, including triggering Canvas for writing and coding, generating various content types, making specific edits, rewriting documents, offering inline critiques, and more. To track progress, they used 20 automated internal evaluations that they created, and they continued refining as they launched the product.

Canvas was built by a team of designers, engineers, researchers, and product managers who worked together from the very start until the product's release. Throughout the process, they developed a product development framework that could be helpful for you too. Here's a general lifecycle for developing product features like Canvas:

1. Developing model behavior specifications

2. Designing reliable and trustworthy evaluations

3. Iterating quickly with prompts to establish results

   

The baseline performance of Canvas is achieved through prompting, using techniques like in-context learning and a few-shot examples. Once you’re confident in the product feature, you can fine-tune the model's behavior with more targeted training. In the case of Canvas, OpenAI used synthetic methods, including a process called distillation, to refine the model.

Of course, things aren’t as neat and straightforward in reality. After creating a baseline, it’s important to revisit your evaluations. As new edge cases arise, you’ll need to update your model behavior specifications. This may mean adjusting the way you train the model.

In the real world, developing the model is an ongoing, iterative process, constantly refining and improving based on new insights and variations.

2. Training Model Behavior

Defining model behavior is all about being precise and thorough, thinking through edge cases, and understanding how different parts of the system interact. For Canvas, I think they had to consider how it works with other tools, like search and DALL-E image generation, for example. As the models get more complex, the complexity of the system grows, and so do the specifications for model behavior.

From my experience, fixing language model behavior is similar to fixing software bugs. Just like in software engineering, you need to identify issues and reproduce them to understand what went wrong. By doing this, you can create a more reliable and robust system that performs well across different scenarios.

In software engineering, dissecting a problem involves upgrading different components of your model to identify the root cause. Once you pinpoint the issue, you need to address it without disrupting other parts of the system. This is where model behavior gets tricky. It's during training that you might start noticing bugs or issues with how the model behaves when it's in use. These issues only come to light once the model is interacting with real-world data. As you iterate, you can refine the model through targeted improvements, fixing specific problems while making sure the rest of the system continues to function smoothly.

3. Challenges When Training ChatGPT Canvas Model

a. The Decision Boundary

OpenAI said that one of the biggest challenges was figuring out when to trigger Canvas, this is called the decision boundary. They trained the model to recognize prompts like "write a blog post about the history of coffee beans," which would trigger Canvas, while avoiding over-triggering on general questions like "help me cook a new recipe for dinner."

For writing tasks, the focus was on getting the triggers right, even if it meant sacrificing accuracy in some other areas. The model reached an 83% accuracy rate compared to the baseline zero-shot GPT-4 with prompted instructions. However, the quality of these baselines can change depending on the specific prompts used. Different prompts can lead to varying degrees of accuracy, and even cause the baseline to perform poorly in different ways.

For example, the errors might be evenly spread across coding and writing tasks, leading to distinct patterns of mistakes and different types of suboptimal performance.

For coding, OpenAI intentionally trained the model to avoid triggering Canvas too much, especially to prevent interfering with power users. This approach was guided by user feedback, as they said they wanted to make sure they weren’t disrupting their workflow.

b. Editing Behavior

Another challenge we faced was tuning the model's editing behavior once Canvas was triggered. This meant deciding when to make very specific, targeted edits versus rewriting the entire content. For example, if the prompt says, "change the second paragraph to be shorter," it's better for the model to focus on just that paragraph instead of rewriting everything.

So, they trained the model to make these targeted edits when users select specific text or when the instruction is clear about what needs to be changed. Otherwise, the model would favor rewriting the entire content. This behavior is continuously refined as we gather more feedback and improve the model.

For writing and coding tasks, here’s how they prioritized improving canvas targeted edits. GPT-4o with canvas performs better than a baseline prompted GPT-4o by 18%.

c. Comments/ Suggestions Quality

Lastly, they continue to focus on training the model to produce high-quality outcomes, ensuring that it can consistently deliver precise and useful results. Comments and suggestions require careful iteration.

Unlike the first two model behaviors, which are easily adaptable to automated evaluation, this required thorough manual reviews.

It's hard to measure quality automatically, so they relied on human evaluations to assess the common quality and accuracy of triggering suggestions and comments.

The results were promising: the Canvas model performed 30% better in accuracy and 16% better in quality compared to the zero-shot GPT 4-0 with responsive instructions. This demonstrated that synthetic training significantly improved response quality and behavior, especially when compared to zero-shot prompting with detailed instructions.

That wraps up the overview of how OpenAI trained the Canvas model. While this playbook focused on sharing these insights, the key takeaway is that thorough, manual review and iterative improvements have helped us build a more accurate and higher-quality AI tool.

Next Part

What does this mean for you? OpenAI Canvas is practical, not theoretical. It’s trained to tackle the exact problems you face, whether you’re refining a viral AI video script or debugging a database project. You’re using a tool that’s been built, tested, and improved through millions of interactions, so you can focus on results, not the process.

So, this part is just the beginning, next, you will see how to make use of these features I mentioned and boost your productivity like 10x better, faster.

Next Part: Canvas for writing: Hidden features in Canvas people don’t often use but we found them really exciting.