This was an experiment to get ChatGPT (GPT-4 version) to write an emoji memory game from scratch. Play the game!

See the full v2.0.0 conversation with ChatGPT here.

Tell ChatGPT what issues I want to fix, let it propose fixes, test those fixes, provide feedback, and merge everything when it's ready.

It took about 40 minutes to refactor the code and improve the design and animation. Some things that I observed:

• ChatGPT can refactor code, but it was initially quite conservative, proposing only a few changes. I had to encourage it to make more changes to improve readability and maintainability.
• It can improve design and add animation upon prompting. It did a much better job with the user interface during this round than it did in the initial round. It does seem to have a design sensibility, though it did initially suggest a quite garish title background that it removed after my feedback.
• Context length became an issue here. For the last task in this round of improvements, I asked ChatGPT if it knew what that task was, which I had told it in the first message. It didn't seem to be aware of it, so we had to start over. It would be good if the ChatGPT interface let you know when you had exceeded the model's context window so you could know when to give it missing context.

See the full v1.0.0 conversation with ChatGPT here.

I tried as best as possible to let ChatGPT write the code, while I provided feedback.

I did, however, instruct it to use React with Tailwind. This is because I'm not very familiar with React or Tailwind (as I primarily program in Python, though I'm also familiar with HTML, CSS, and JavaScript). So it was easier for me to play the role of ChatGPT's robot arms, and I also wanted to learn about React and Tailwind from this experience.

It took about 50 minutes to complete this, though I did have one false start before the successful conversation due to a Git conflict. (I had a GitHub repository remotely, ChatGPT initialized a separate repository locally, and it struggled to connect the two.)

Overall, this was a good experience, but I think it would be challenging for someone with less familiarity with programming, HTML, CSS, and JavaScript. Here are some takeaways:

• ChatGPT can create complete full applications (although this one is simple). It's able to understand how to map requirements to a functioning application. And it's able to act on feedback.
• It sometimes gives unclear instructions. For example, it created a component called GameBoard.js and a function called GameBoard within it. It would tell me to write some code in "GameBoard," but it wasn't clear if it meant the GameBoard function or the GameBoard component more generally.
• It may not create optimal architectures. For this app, for example, it shoved pretty much everything into GameBoard.js. It probably would have been better to separate functionality into different components, like ScoreBoard.js, Card.js, RestartButton.js, etc. Similarly, the GameBoard function is pretty packed and should be refactored.
• This said, it won't always take the same approach. As mentioned above, I had a false start with a Git issue. During that attempt, ChatGPT did separate functionality into different components.
• It may not be great at design. The board design in this game is pretty basic. Even when I asked for something nicer, ChatGPT didn't output CSS that made the board much more attractive. Perhaps this is because the current GPT-4 version of ChatGPT still doesn't have multimodality, so it can't see images, and I can't show it what the boring board looks like. Maybe when ChatGPT gets multimodality I can show it an example of a cool card, and have it copy that. It's also possible (perhaps even likely) that I didn't prompt it in the right way to get a better design.

Overall, I think there's a lot of potential here, but it requires learning about ChatGPT's quirks and being sure to prompt it in a way that gets the best outcomes. For example, if I had told it initially that I wanted to separate distinct functionality into distinct components, it probably would have built the application that way.

Here are two ways I would improve this app further (with ChatGPT's help, of course):

• Refactor the GameBoard.js component into multiple components, and the GameBoard function into multiple functions
• Improve the design of the board and cards, and add some light animation