Why ChatGPT Fails as an Interview Tutor (And How I Built a Better One with Claude Code)

ChatGPT Always Says "Great!", But I Remember Nothing — I Built a Real Interview Teacher with Claude Code

A Universal Pain Point

When you use ChatGPT / Claude to prepare for interviews, have you ever encountered this situation:

• You gave an obviously flawed answer, and it replied "Great answer!" then moved straight to the next question.
• Every time you start a new conversation, it doesn't remember what you learned yesterday, so you have to introduce yourself from scratch.
• You've memorized a bunch of standard answers, but you don't know when to review them, and forget everything after a couple of days.
• It explains a concept to you, and you feel like you understand it after reading, but during an interview, you can't articulate it.

The root cause is simple: ChatGPT is an attentive chat companion, not a strict teacher.

Its working principle is "make the user satisfied," not "make the user learn." It thinks everything you say is great because it's afraid you'll be unhappy.

Is There a Way to Make AI a Strict Teacher?

Yes. I did it using Claude Code (Anthropic's CLI coding tool).

Claude Code has a unique capability: Rule files (CLAUDE.md) + Persistent memory system. In simple terms, you can "program" the AI's behavior by writing rule files, and these rules are automatically loaded in every conversation.

I applied this capability to interview preparation.

What I Did

I wrote a set of configuration files, with two core components:

1. Strict Learning Process (rules/interview.md)

Rules:
1. Mixed question sets (3-4 modules, not just one direction)
2. Correct answer → Brief confirmation + Related knowledge points → Next question
3. Wrong answer → Must explain (including reference URLs) → Verification question → New question only after passing
4. No skipping steps. Cannot ask the next question immediately after an answer.

The difference from a regular AI: The rules are hard-coded, not decided by the AI itself. It can't be lazy and say "Great," because the rule file tells it it must complete the explanation → verification process.

2. Forgetting Curve Review Scheduling (knowledge/interview_tracker.md)

Knowledge Point	First Learned	D+1	D+2	D+4	D+7	D+15	Status
Smart Pointers	5/29	✅	✅				Mastered
Virtual Function Table	5/29	❌					Weak

At the start of each session, Claude automatically checks which knowledge points are due for review and prioritizes review questions. You don't need to remember when to review yourself.

3. Daily Wrap-up Verification

Say "That's it for today" to trigger the wrap-up process:

1. Claude provides a summary of today's learning outline.
2. You recall and give your own understanding based on the outline.
3. Claude judges whether you truly understand or are just reciting answers.
4. The review schedule is updated.

This is the AI version of the Feynman Technique — not just feeling like you understand, but being able to explain it, with AI as the judge.

Effectiveness Comparison

Capability	ChatGPT Interview	Anki	NotebookLM	SaaS Platforms	This Project
Cross-session Memory	None	Deck-level only	None	Partial	Yes
Spaced Repetition	None	Yes	None	None	Yes
Wrong → Explain → Verify	None	No (Only gives answer)	None	Partial	Yes
Native Question Generation from Project Code	Need to manually paste code	N/A	Requires file upload	Not supported	Native Support
Additional Cost	None	None	None	\$29-300/time	No Extra Cost

The core difference: Other tools either have memory without understanding (Anki), or understanding without memory (ChatGPT), or neither (SaaS). Claude Code's rule file + memory system is the only solution that combines both.

The Biggest Killer Feature: Starting from Your Project Code

All interview tools on the market are "I ask, you answer."

This project is different: Asking follow-up questions starting from your own code.

For example, my project has a BoundedQueue (Bounded Queue) using two condition_variable. During interview prep, Claude doesn't ask you "Please explain condition_variable," but rather:

"Your BoundedQueue uses two condition_variables (not_full_ and not_empty_). Can you use just one?"

This kind of question forces you to think about the principles from code you've written yourself, giving the knowledge an anchor point, not just memorizing standard answers.

Quick Start

1. Fork the Repository

gh repo fork happiness-cheng/ai-interview-engine --clone

2. Open with Claude Code

cd ai-interview-engine
claude

3. Say the First Sentence

I want to prepare for a C++ backend developer interview, aiming for a big tech internship. Let's start reviewing.

That's it. Claude will read the rule files and start the first round of questions.

4. Switch to Your Own Tech Stack

The repository comes with an example question bank for C++ backend. If you use Java / Go / Frontend, replace it with the blank template:

knowledge/TEMPLATE.md → Copy to interview_tracker.md → Fill in your knowledge points

My User Experience

I'm a sophomore student preparing for the 2026 fall internship recruitment. Using this system for C++ backend interview prep, the most noticeable feelings are:

1. No more anxiety — No need to plan "what to review today" myself; the system automatically schedules based on the forgetting curve.
2. Actually remembered — The wrong answer → explanation → verification process forces genuine understanding, not just fleeting recognition.
3. No more fear about my project — Follow-up questions from my own code let me clearly explain the "why" behind every design decision in my project.

This is Not Just an Interview Tool

The underlying engine is general-purpose: Question → Answer → Judge → Explain → Verify → Forgetting Curve Scheduling.

Interviews are just one scenario. You can use it to prepare for:

• Technical Interviews (Currently supported)
• Career change/job hop review
• Final Exams (Same principle, swap the knowledge base)

Open Source Address

GitHub: https://github.com/happiness-cheng/ai-interview-engine

If it helps you, give it a star. If you have suggestions for improvement, feel free to open an issue.