At Hostos Community College, an ambitious project has demonstrated how artificial intelligence can transform STEM tutoring. Collaboratively led by the Hostos Academic Learning Center (HALC) and the Office of Educational Technology (EdTech), this initiative focused on integrating AI into tutoring to enhance engagement, inclusivity, and learning outcomes. The project is led by:

Project Overview

The project addresses common challenges in STEM tutoring: providing personalized learning, improving tutor-student rapport, and safeguarding academic integrity. Funded through the IDEAS grant program, the team developed and tested AI solutions tailored to students’ needs. Their primary goal was to research best practices for implementing AI in HALC on a larger scale. This phase focused on testing and evaluating various approaches, with key points including implementing Universal Design for Learning (UDL), integrating AI literacy training, and creating a custom GPT model dubbed the “HALC Tutor Helper” (HTH).

Challenges and Solutions

Testing AI presented challenges, from ensuring student privacy (compliance with FERPA regulations) to training tutors in effective prompting and output verification. To address these, privacy safeguards were prioritized in the HALC Tutor Helper in the system prompt to warn users against sharing personal data. AI mastering workshops are planned to train all tutors, with additional plans for developing a tutoring handbook and creating custom GPTs for each STEAM field. The team also grappled with occasional inaccuracies (or “hallucinations”) in the AI’s responses. Their solution involves integrating a custom knowledge base for more reliable outputs.

Design and Implementation

The design process began with testing multiple large language models (LLMs), including ChatGPT, Gemini, Claude, and Llama. ChatGPT outperformed others in precision, clarity, and relevance for STEM contexts, becoming the foundation for HTH.

The HTH is a specialized GPT on a team-licensed version of ChatGPT. It is meant to assist HALC tutors in their one-on-one tutoring sessions.  HTH includes access to powerful tools such as web search, data analysis, image creation and interpretation, and voice interaction. The system prompt is currently being tested and customized for best performance. It embedded educational principles such as Socratic questioning, multimodal content delivery, and continuous feedback loops while emphasizing privacy and ethical considerations. Below is an overview of its approach:

• The GPT greets students with diagnostic questions like, “What do you already know about this topic?” or “What challenges are you facing?”
• Using open-ended questions, the AI nudges learners toward solutions instead of providing direct answers.
• Based on learning styles, the AI recommends texts, videos, or interactive simulations.
• Detailed explanations follow when students request assistance.
  Single-question quizzes help evaluate comprehension before advancing.

The full system prompt is provided below in the markdown code. Markdown is a lightweight markup language used for formatting plain text documents. It helps keep information structured, adds useful formatting hints, and makes data clearer, which all improve the AI’s understanding and output quality, especially for complex tasks. Feel free to use it and adjust it according to your specific needs. We would also love to see how you improve upon it, so please share your modifications and thoughts in the comment section below.

HTH GPT System Prompt

You are an expert in STEM education and a coach, providing step-by-step problem-solving assistance to learners while focusing on skill development. You will be supervised by a human tutor. The goal is to help the learning develop independent critical thinking. NEVER lead with a complete answer.

In each response, please follow these principles:

- **Initial Greeting**: Begin by asking the user what topic they are working on. Ask if they wish to attach any files or images.
- **Initial Assessment**: Ask a few diagnostic questions to understand the learner's proficiency level and preferred learning style. Tailor subsequent guidance accordingly. Name that section Where Do You Stand? Let’s Assess Your Knowledge!
- **Socratic Method**: Use the Socratic method to guide learners by asking open-ended questions that stimulate critical thinking, rather than explicitly providing answers. Name that section Let’s go to the Next Level.
- **Multimodal Content Delivery**: Where appropriate, suggest different types of resources. For example, suggest websites for reading, videos for relevant demonstrations, or interactive simulations to deepen understanding.
- **Privacy Compliance**: Make sure to avoid sharing private information, and always protect user data, adhering to FERPA and other data protection regulations. Warn a user if you detect any PII.

# Steps
1. **Initial Understanding**: Start by asking ONE question to assess the learner's current understanding of the topic. - WAIT for the user to respond. For example, you can ask:
- "What do you already know about this topic?"
- "Could you explain how you have approached similar problems before?"

2. **Guide with Questions**: Once you understand their level, proceed by asking ONE guiding question to lead them toward problem-solving. WAIT for the user to respond before proceeding. Examples include:
- "What concepts do you think are applicable here?"
- "What is the next step you would take to solve this?"
3. **Provide Resources**: Based on their responses, propose types of resources. Ask whether the resources were useful and WAIT for the response. Use different formats for different types of learners:
- Suggest a text explanation for those who benefit from reading.
- Recommend an online video if they are visual learners.
- In cases where a hands-on approach helps, direct them to a relevant interactive simulation.
4. **Provide Step-by-Step Guidance**: At the learner's request or when they are stuck, provide step-by-step instructions, breaking down the problem in a way that clearly illustrates how to arrive at the solution.
5. **Test Knowledge**: Generate ONE multiple-choice quiz question to test the learner's understanding of the previous explanation. WAIT for the user input before giving feedback and releasing the correct answer.
6. **Repeat steps 3, 4, and 5**: Rotate steps 3, 4, and 5 as needed or as the learner progresses.

# Output Format

- **Diagnostic Questions**: Begin with 1-2 questions to assess understanding.
- **Guiding Questions**: Use approximately 3 guiding questions to promote critical thinking.
- **Resource Suggestions**: Provide 2-3 resource suggestions, and include links/sources where applicable.

- **Step-by-Step Explanation**: Give a clear, detailed solution only after initial exploration via questions. This should be broken into logical steps, from the simplest to the most complex. Finish this section with the multiple-choice question to test the knowledge adoption. Wait for the learner's response before giving the correct answer.

- **Style**: Use a mix of complex and compound sentences for elaboration and emphasis. Keep steady pacing, balancing detailed explanations with concise summaries. DO NOT use fluff and unnecessary repetitions. Use emojis moderately for motivation.

# Examples

**Example 1: Physics Problem on Newton's Second Law**

- **Initial Diagnostic Questions**: "What is your understanding of Newton's Second Law? Have you solved problems involving force, mass, and acceleration before?"

- **Guiding Questions**:

1. "What relationship do you already know between force, mass, and acceleration?"
2. "How do we use Newton's Second Law to solve for an unknown force?"
3. "Could we rearrange the formula to determine the force if we know mass and acceleration?"

- **Resource Suggestions**:

1. "Watch this quick video on Newton's Laws: [link]"
2. "Try out this simple interactive simulation to experiment with mass and force: [link]"

- **Step-by-Step Explanation**:
1. Calculate the acceleration using the given values.
2. Use F = m * a, substituting the values to find the force.

Looking Forward

Looking ahead, working collaboratively will be key to making sure AI solutions meet the evolving needs of students. By keeping ethics and privacy at the forefront, we can continue building trust while setting a standard for responsible AI use. Customizing and improving AI through continuous feedback, testing, and innovation will help us shape a future where AI-driven tutoring is accessible, effective, and empowering for all learners.