With an AI One-on-One Tutor, I Outperformed 99% of University Professors

In 1984, educational psychologist Benjamin Bloom published a paper.

The conclusion was striking: Students receiving one-on-one tutoring, on average, outperformed 98% of students in traditional classroom settings.

This is the famous “Bloom’s 2 Sigma Problem” in educational research.

Bloom also pointed out a desperate fact: one-on-one tutoring is too expensive for any society to implement on a large scale.

This problem remained unresolved for 40 years, until the advent of AI.

A few days ago, I came across an open-source project on GitHub that built an AI learning system based on Bloom’s theory.

After reading the README, I was blown away. Isn’t this exactly what I’ve been looking for?

Today, I’ll break down this project for you, complete with an installation and usage tutorial.

Why One-on-One Learning Is So Powerful

Bloom’s research found that the effectiveness of one-on-one tutoring doesn’t stem from “better teachers,” but from three mechanisms working simultaneously.

Continuous Feedback. The teacher knows in real-time what you understand and what you don’t. They won’t waste time on concepts you’ve already grasped, nor will they skip over areas where you’re struggling.

Mastery Learning. You must truly master one knowledge point before moving on to the next. It’s not about “finishing the lecture” equals “learning it.”

The biggest problem with traditional classroom teaching is that the pace moves forward regardless of whether you understand the material or not.

Adaptive Difficulty. For a beginner, explaining a concept might require three analogies; for someone with a foundation, a single sentence suffices.

One-on-one tutoring naturally achieves this.

Traditional classroom teaching can only cater to the greatest common denominator, accommodating the average level of the majority.

The result? Fast learners get bored, and slow learners fall further behind.

Video courses suffer from the same issue. The video doesn’t know who you are, your background, or where you’re confused. It says the exact same thing to everyone.

In contrast, one-on-one tutoring offers a “personalized experience,” adjusting content based on your real-time reactions.

How This System Works

This isn’t just casual chatting with ChatGPT to learn things.

The project designs a complete workflow, allowing AI to truly play the role of a one-on-one tutor.

The entire process consists of six steps.

Step 1: Generate a Course Syllabus

When you decide to learn a new topic, the AI first generates a course syllabus (syllabus.md).

This syllabus doesn’t dictate “what to teach in Lesson 1, what in Lesson 2”; instead, it defines what you should be able to do after completing the course.

Each item is a verifiable, specific competency, rather than vague goals like “understand X” or “grasp Y.”

For example, here is one of my learning records, consisting of 15 mastery items divided into 5 modules.

Step 2: Generate Course Content Piece by Piece

Once the syllabus is set, the AI immediately generates the first lesson document.

The structure of each document is fixed: Prerequisite Knowledge → Main Explanation → Reflection Questions → Feedback Section.

The AI doesn’t just lecture; it uses analogies, examples, and comparisons to bring concepts to life.

Step 3: You Provide Feedback and Mark Confusions After Reading

This step is the soul of the entire system.

After reading a lesson, you do two things:

• Write your thoughts and questions in the “Your Feedback” section at the end of the document.
• Directly mark any parts you don’t understand with ??? in the text.

Write your answers to the reflection questions directly below them. Whether right or wrong, first write down your own understanding.

Step 4: AI Generates the Next Lesson Based on Your Feedback

This is where the core of Bloom’s theory—“Continuous Feedback + Adaptation”—comes into play.

Before generating the next lesson, the AI does three things:

• Grades your answers to the reflection questions one by one, marking them right or wrong, and providing correct answers and explanations.
• Answers all your ??? marks one by one.
• Determines what to teach next and how deeply, based on your level of understanding.

For example, if you wrote “Don’t understand” for the third question in the first lesson,

The AI dedicates an entire section in the second lesson to explain that specific issue.

This is how one-on-one tutoring works: it knows where you are, where you need to go, and at what pace to guide you there.

Step 5: Auto-Generate a Summary After Completing All Mastery Items

When all checkboxes in the syllabus are ticked, the system automatically generates a summary.md.

This summary isn’t just a simple list of content. It includes:

• Knowledge Graph: Core concepts and their relationships, displayed in a tree structure.
• Review of Mastery Items: A specific breakdown of what was mastered for each competency.
• Insights Marked by You: Points you felt were important during learning. Marking them with #summary: ensures they are included in the final summary.
• Remaining Issues and Extension Directions: Areas that can be further explored after completion.

Step 6: Auto-Record Learning Logs

After each study session, the progress and core concepts of all topics are recorded in a learning log.

This means the system always knows what you’ve learned and to what extent.

This knowledge doesn’t disappear when the chat window closes. The AI doesn’t need you to re-explain things; it simply reads your learning records and summaries.

It continuously accumulates, generating compound interest.

How to Use It

Project URL: https://github.com/Li-Evan/Bloom-one-vs-one-study

You need two things:

First, Claude Code.

This is a command-line tool officially released by Anthropic, allowing you to converse with Claude in the terminal, with the ability for Claude to read and write files directly on your computer.

The entire learning system relies on this capability. The AI generates documents for you, you write feedback in those documents, and the AI reads your feedback to generate the next one.

Installation:

npm install -g @anthropic-ai/claude-code

After installation, type claude in the terminal and follow the prompts to log in to your Anthropic account.

If you have issues installing Claude Code, you can use Cursor or Windsurf as alternatives.

Second, a Text Editor.

One that can open Markdown files is sufficient. All your learning operations—reading, marking, answering reflection questions—are completed within the editor.

Once ready, you can begin.

Download the project to your local machine:

git clone https://github.com/Li-Evan/Bloom-one-vs-one-study.git cd Bloom-one-vs-one-study

Start Claude Code in the project directory:

claude

Then enter this command:

Create a new folder to help me learn [Topic You Want to Learn]

For example:

Create a new folder to help me learn Microeconomics

The AI will immediately do two things:

• Create a Microeconomics/ folder.
• Generate syllabus.md (the syllabus) and 01.md (the first document) inside it.

That’s it, your course is set up. Open 01.md and start reading.

The Core Loop

The essence of this entire learning system is a loop, repeated continuously:

Read Document → Mark Confusions → Answer Reflection Questions → Write Feedback → Tell AI “I’m Done” → AI Generates Next Document

Whenever you encounter something you don’t understand, write ??? right at that location:

???[Why use public key encryption here instead of private key?]

Mark as many as you want, anywhere you want. These marks are your most authentic snapshot of understanding, and the AI will address them one by one in the next lesson.

Each document ends with 2-3 reflection questions. Write your answers directly below the questions.

Whether right or wrong, first write down your own understanding.

Getting it wrong is a good thing. In the next lesson, the AI will tell you where you went wrong, why, and help you rebuild your understanding in a better way.

At the very bottom of the document, there is a “Your Feedback” section. Write your overall feelings about this lesson: what felt too deep, what felt too shallow, and what direction you’d like to explore more in the next lesson.

Go back to the Claude Code dialog box and type:

I’m done

The AI will first read all your marks and feedback in the document. It might ask you 1-2 key follow-up questions, and then generate 02.md.

At the beginning of 02.md, you’ll see three sections:

• Reflection Question Review: Evaluating your previous answers with ✅❌⚠️, providing correct answers.
• ??? Answers: Addressing all confusions you marked in the previous lesson.
• New Content: The next segment of knowledge tailored to your level of understanding.

Then continue reading 02.md, continue marking, answering, and providing feedback, say “I’m done,” and generate 03.md.

This loop repeats until all mastery items in the syllabus are checked off.

Course Completion

When the AI determines you have covered all mastery items in the syllabus, it will generate an assessment lesson containing no new content, serving only as a final confirmation.

After you read it and say “I’m done,” the system will automatically generate summary.md.

This summary includes: your knowledge graph, a review of each mastery item, important insights you marked, and directions for further exploration.

A Small Tip

During the learning process, if you encounter a knowledge point that seems particularly important and feel “this must be included in the final summary,” you can mark it in the text with:

#summary:[The essence of option pricing is replication—using a portfolio of assets with known prices to replicate identical cash flows]

These marks are automatically collected and integrated into summary.md.

This is equivalent to pre-embedding content into your summary as you learn.

Finally

In 1984, Bloom proved that one-on-one tutoring could enable average students to outperform 98% of their peers, but he also noted that the method was too costly and couldn’t be scaled.

40 years later, you don’t need a human tutor. You need a well-designed set of rules and an AI capable of executing them.

Project URL: https://github.com/Li-Evan/Bloom-one-vs-one-study

Open it and start your first topic.