Blog Post #3: Training the AI Tutor: Crafting Smarter Prompts for Smarter Teaching

Hi everyone!

This week, I’ve been deep in the trenches of prompt engineering to guide the GPT model to produce better, more effective outputs. While the model already has a wealth of knowledge, teaching personal finance to teens requires more than just spitting out facts. It needs to explain concepts clearly, adapt to different learning styles, and engage users in a way that makes financial literacy stick. To achieve this, I focused on two powerful prompt engineering techniques: chain-of-thought prompting and few-shot learning.

1. Chain-of-Thought Prompting:

Chain-of-thought (CoT) prompting is all about getting the model to show its work. Instead of jumping straight to an answer, the model is prompted to break down its reasoning step by step. This is especially useful for teaching because it mirrors how a tutor should explain complex ideas.

For example, if a teen asks, “Why is it bad to only pay the minimum on a credit card?” the model doesn’t just say, “Because you’ll pay more in interest.” Instead it walks through the logic:

1. Line 1: “When you only pay the minimum, the remaining balance accrues interest.”
2. Line 2:“For example, if you owe $1,000 with a 20% annual interest rate, you’ll pay $200 in interest over a year.”
3. Line 3:“If you only pay the minimum, it could take years to pay off the debt, and you’ll end up paying much more than the original amount.”

By structuring prompts to encourage this step-by-step reasoning, the model can generate more engaging answers by walking the user through analytical reasoning. Note the above examples are simplified and real responses would be made to be more in depth.

2. Few-Shot Learning: Teaching the Model by Example

Few-shot learning is another useful tactic. It involves providing the model with a few examples of high-quality responses to specific questions or scenarios. These examples act as a blueprint, showing the model how to structure its answers in a way that’s clear, engaging, and educational.

For instance, I fed the model examples like:

• Question: “What’s the difference between a savings account and a checking account?”
• Answer: “A checking account is for everyday spending—like paying bills or buying groceries. A savings account is for storing money you don’t need right away, and it usually earns interest over time. Think of it like this: checking is for now, savings is for later.”

By giving the model a handful of these examples, it learns to mimic the tone, depth, and clarity of a great tutor. This is especially helpful for personal finance, where concepts can feel abstract without real-world context, helping with content depth. This ensures the model can behave in its intended manner rather than leaving it to its imagination.

3. Vector Databases: Supercharging the Tutor’s Knowledge Base

While not a prompt engineering technique, vector databases play a crucial role in making the tutor faster and more accurate. Think of them as the tutor’s “memory bank,” where it stores and retrieves information quickly and efficiently.

Here’s how it works:

• When the model encounters a question, it converts the query into a numerical representation (a vector) and searches the database for similar vectors. This is different from a text-to-text database as you would only be able to correlate a query with a data point if they have the same keyword.
• This allows the tutor to pull up relevant information instantly, whether it’s a definition, a case study, or a step-by-step explanation.

For example, if a teen asks, “How do I start investing?” the model can quickly retrieve a pre-stored explanation, complete with examples like:

• “Start by setting a goal, like saving for college or retirement.”
• “Then, research low-cost index funds or ETFs as a beginner-friendly option.”

Note that none of these data points have the key word investing or saving. This form of storage and retrieval is for semantic purposes.

At the end of the day, these are all techniques to improve on the default GPT-4o model. As the model trains, I hope it can become a better tutor.