Transforming education with Artificial Intelligence: we created 8,000 activities for public schools

By: Otacilio Saraiva Maia Neto

At AprendiZAP, our mission is to be the best and most practical digital learning tool for students at public schools in Brazil. Looking at today’s educational challenges, we recognized the urgent need to encourage autonomous and participatory learning, which can be achieved through the adoption of active teaching methodologies. With this approach, students are encouraged to interact and create solutions to real-world problems, fostering more meaningful learning.

In this context, our engineering team was challenged to create at least six active learning methodology proposals for each of the 1,428 lessons available for lower secondary education on AprendiZAP. In addition, it was necessary to organize a content review process with a team of teacher reviewers to ensure the excellence of the material produced.

The Power of Artificial Intelligence

Considering the large volume of content—more than 8,000 activities by the end of the project—manual creation would have been unfeasible. Therefore, we chose to use generative artificial intelligence. We began our preliminary tests with ChatGPT, which quickly showed potential, although it required refinement.

Given the volume of work, using ChatGPT manually would have been impractical. We therefore automated the process using the OpenAI API, which provides access to the gpt-3.5-turbo model at an affordable cost. Using the API also allowed us to adjust certain parameters that are not available in ChatGPT, such as "temperature," which influences the degree of creativity in the model’s responses.

The Art of Prompt Engineering

A vital part of our work involved prompt engineering. A prompt is the instruction given to an artificial intelligence model, and the quality of the responses generated by GPT depends directly on the quality of the prompt used.

During the first weeks of the project, we devoted significant effort to defining and refining the prompts that would be used to generate the activities. We were assisted by iFood engineers who were also studying the subject and provided us with valuable insights.

In addition, we found a guide to best practices in prompt engineering developed by OpenAI. Based on this guide, we adopted recommended practices for structuring and improving our prompts. Among these, the importance of defining the desired output format through examples and focusing on asking "what to do" rather than "what not to do" stands out.

One particularly interesting suggestion was to use artificial intelligence itself to improve our prompt. We therefore made the following request to the model:

Thus, we combined the techniques we learned, qualitative analyses of generated test packages, and feedback from the AI model itself to arrive at a satisfactory prompt. We also adjusted the "temperature" parameter for each subject, as we noticed that, in some subjects, a higher temperature led the AI to produce results that were less connected to reality.

The final prompt used was:

With this prompt, we were able to populate the metadata according to each lesson for which active methodology proposals would be generated. We also identified 12 active methodologies to be used in the project, allowing for a greater variety of educational activities.

Teacher Review Process

Given the amount of content generated, we needed a solid review strategy. The challenge was to present the material on a platform and interface with which the teacher reviewers were familiar. Several CMS platforms were considered, but we chose Google Docs because it was already part of their routines.

Using the Google Drive and Google Docs APIs, we were able to automate part of this process. The materials generated by GPT were converted into Google Docs files and stored in a shared Google Drive folder. This allowed the teacher reviewers to conveniently access, review, and edit the material.

Google Drive also provided additional benefits, such as automatic backups, protection against accidental file deletion, and version control, which allowed us to track all changes made to the documents.

The teacher reviewers played a crucial role in assessing the quality of the generated materials. To speed up the process and make it more efficient, we established an emoji code. Teachers used the ✅ emoji to mark files that had been reviewed and were ready for use. They used the ❌ emoji to indicate proposals that had been rejected, meaning that we would need to make adjustments and generate the material again, for example, by trying one of the other available active methodologies.

After the review process, we used the Google Drive API again to extract the files ready for use, filtering only the files that began with the ✅ emoji. These files were then incorporated into our content database.

Our data team also played a vital role in this process, being responsible for segmenting the full text into its respective sections using regular expressions. This teamwork and the successful integration of different skills were fundamental elements in completing this project.

Conclusions

The use of artificial intelligence proved essential to making this ambitious project possible. Without AI, producing content on such a large scale would have been impractical. Thanks to GPT and refined prompt engineering, we were able to generate an enormous amount of high-quality educational material.

We learned a great deal from this project. Prompt engineering, material review, and the use of APIs from tools such as Google Docs and Google Drive—all of these things contributed to the project’s ultimate success. The key was understanding how technology could assist the process without ever forgetting the ultimate goal: creating the best possible learning experience for our teachers and students.

Although artificial intelligence is a powerful tool, it is important to remember that it is only that—a tool. We still rely heavily on human judgment and experience to ensure that the generated content is of high quality and genuinely useful for teaching.

This project is only the beginning. With continuing advances in AI and educational technology, we look forward to seeing what the future holds for education in Brazilian public schools. At AprendiZAP, we will continue exploring every possibility to improve learning and provide the best opportunities possible for our students.