필사 모드: How to Study with AI — 8 Techniques That Turn an LLM into Your Best Tutor

• Introduction — The 2 Sigma Problem and the Promise of AI Tutors
• The Traps First — Three Ways AI Ruins Studying
• Technique 1 — The Socratic Tutor
• Technique 2 — Feynman Role-Play
• Technique 3 — Quiz Generator plus Spaced Repetition
• Technique 4 — The Error Analyzer
• Technique 5 — The Difficulty Ladder (i+1)
• Technique 6 — Contrast Learning
• Technique 7 — AI as Reviewer (You Generate, It Critiques)
• Technique 8 — The Mock Interviewer
• The Verification Loop — Not Getting Eaten by Hallucination
• Tying It into One Workflow — a 4-Week Example for a New Technology
• Conclusion
• References

Introduction — The 2 Sigma Problem and the Promise of AI Tutors

In 1984 the educational psychologist Benjamin Bloom published a famous observation: students who received 1:1 tutoring performed on average two standard deviations (2 sigma) better than students in conventional classrooms. That does not mean the top 2% got better — it means an average classroom student rose to roughly the top 2% level. The problem was cost: you cannot give everyone a personal tutor. Bloom called this the "2 sigma problem," and for forty years it remained the holy grail of educational technology.

LLMs are the closest tool yet to that grail. You now have a partner with infinite patience, tuned to your level, available at any hour. But there is a trap: used wrongly, AI does not help you study — it studies instead of you. This article names the traps first, then lays out eight usage patterns grounded in learning science, each with a prompt. It is the final part of the trilogy, following the role knowledge map and the interview playbook.

The Traps First — Three Ways AI Ruins Studying

The fluency illusion. As the secrets of memory showed, reading a well-written explanation produces the feeling of knowing, but that feeling is not memory. Because AI explanations are unusually smooth, the illusion gets stronger. The gap between understanding-while-reading and being able to reproduce it yourself only reveals itself in the exam room or the interview.

Generation dependence. The core of the testing effect from Roediger and Karpicke's research is that memory is built not by input but by retrieval. When AI generates the answer, your retrieval practice disappears wholesale. It is the same reason your coding muscles do not grow when AI writes all the code.

Hallucination and uncritical absorption. LLMs are wrong in plausible ways — especially with numbers, recent information, API details, and paper citations. Absorb without a verification habit and you accumulate misconceptions bundled with confidence.

The common antidote to all three traps is one sentence: don't make the AI answer — make it question you.

Technique 1 — The Socratic Tutor

The foundational setup: explicitly cast the AI as a tutor who never gives the answer.

You are a Socratic tutor. Rules:
1. Never state the answer directly.
2. When my answer is wrong, ask one guiding question at a time so I
   discover the error myself.
3. Only if I get stuck twice in a row, give one small hint.
4. When I reach the answer, point out what my explanation still missed.
Topic: [e.g., pod scheduling in Kubernetes]
Start with your first question.

The power of this setup is that it hard-wires retrieval practice into the very structure of the conversation. Studying-by-reading becomes studying-by-answering.

Technique 2 — Feynman Role-Play

The biggest weakness of the Feynman technique (explain it simply) is that alone, you cannot see your own gaps. Casting the AI as a relentless beginner removes that weakness.

I will now explain [topic]. You play two roles at once:
1. A curious beginner: ask "why?" at every point my explanation
   doesn't hold up.
2. A strict examiner: afterwards, list every place I hand-waved
   or was imprecise.
Reply "start explaining" when ready.

The places you stumble while explaining are exactly the holes in your knowledge. You then restudy only those holes, which shrinks your review scope dramatically.

Technique 3 — Quiz Generator plus Spaced Repetition

Having the AI convert what you just learned into test items gives you unlimited retrieval practice.

What I just studied: [paste notes]
Create 5 questions from this. Conditions:
- 3 multiple-choice (with plausible wrong options), 2 short-answer
- One question at a time; grade and explain after I answer
- The explanation must include what was missing from MY answer

Combine this with spacing: ask it to "re-test me on X from three days ago," or have wrong answers exported as Anki-style cards. It is exactly the principle that review just before forgetting builds the strongest memory.

Technique 4 — The Error Analyzer

More important than the fact that you were wrong is the type of misconception behind it. AI is excellent at this diagnosis.

Question: [question]
Correct answer: [answer]  /  My answer: [my answer]
Do not grade me. Infer what misconception most likely led me to
this answer, then create 2 check questions that would verify
whether I hold that misconception.

A powerful combination: practice with tools that ship detailed explanations — like this site's JLPT mock quiz or TOEFL reading practice — then bring only your repeated error patterns to the AI for a root-cause diagnosis.

Technique 5 — The Difficulty Ladder (i+1)

Krashen's "comprehensible input" principle from what fluency really is applies beyond language: learning happens slightly above your current level, at i+1. AI does this calibration precisely.

I am studying [topic]. My current level: [summary of what I know].
Give me one practice task exactly one step harder than my level.
If I solve it, raise the difficulty one notch; if I get stuck
twice, lower it half a notch.

A textbook gives everyone the same staircase; AI carves a staircase just for you, in real time. This is precisely the point where tutoring beats lectures.

Technique 6 — Contrast Learning

Confusable concept pairs (process vs. thread, authentication vs. authorization, RAG vs. fine-tuning) stay confusing if studied separately. The boundary only appears when they are placed side by side.

A: [concept A], B: [concept B]
1. Create 3 decisive questions that tell these two apart.
2. Give one scenario where A applies and B fails, and one where
   B applies and A fails.
3. Explain the most common reason people confuse them.

As the interview playbook showed, a large share of interview questions target exactly this discrimination ability.

Technique 7 — AI as Reviewer (You Generate, It Critiques)

For code, writing, and design, order is everything: you create first, then hand it to AI for review. Reverse the order and you lose the generation effect — the memory boost from producing something yourself.

Here is my [code/design/answer]: [paste]
1. Review in this order: critical problems → improvements → style.
2. For each point, explain WHY it is a problem at the level of
   principles.
3. Do not give me a fixed version — only point the direction so
   I can fix it myself.

Condition 3 is the crux. The moment you receive the fixed version, learning ends and copying begins.

Technique 8 — The Mock Interviewer

This is the prompt promised in the interview playbook's 8-week plan.

You are a senior interviewer at a [company type]. Run a technical
interview for a [role] position. Rules:
- One question at a time, like a real interview. Up to 2 follow-ups
  per answer.
- If I answer vaguely, demand a concrete example.
- After a 45-minute portion (4-5 questions), stop and give me a
  scorecard: problem solving / communication / depth / collaboration
  signals, each out of 5, with evidence.
- End by naming the single thing I should fix first.

You get the real-interview pressure and instant feedback that used to be impossible to produce alone — on infinite repeat. Your behavioral story bank can be rehearsed the same way.

The Verification Loop — Not Getting Eaten by Hallucination

A safety layer that spans all eight techniques.

• Demand sources: "Give me the official docs or paper behind this claim. If you are not sure, say you are not sure."
• Separate facts from judgment: version numbers, API signatures, figures, and citations must be cross-checked against official docs. Concept explanations and practice-question generation are comparatively safe territory.
• Verify in reverse: for anything important, ask "which part of your explanation just now is most likely to be wrong?" It is surprisingly good at finding it.
• Never paste in personal data or company code: even as a study tool, the input boundary stands.

Tying It into One Workflow — a 4-Week Example for a New Technology

Weaving the eight techniques into an actual sequence — say, learning Kubernetes:

1. Week 1 — map and input: have AI generate a roadmap (diagnose your level with technique 5) → read official docs and tutorials → end each day with generated quizzes (technique 3).
2. Week 2 — retrieval and practice: get hands-on in the Kubernetes playground → when stuck, the Socratic tutor (technique 1) → error analysis (technique 4).
3. Week 3 — generation and review: build a small project yourself → AI review (technique 7) → contrast learning on the concept pairs that still confuse you (technique 6).
4. Week 4 — explanation and exam: full explanation via Feynman role-play (technique 2) → write up what you learned on a blog (visibility!) → finish with a mock interview (technique 8).

The cycle is read → retrieve → generate → explain, with AI playing a different role at each stage. That is the point: AI is most powerful not as an answer sheet but as a question machine, examiner, reviewer, and interviewer.

Conclusion

Forty years after Bloom posed the 2 sigma problem, a personal tutor finally sits in everyone's pocket. But having a tutor and using a tutor well are different things. What happens when a good student meets a good tutor — being questioned constantly, explaining out loud, having errors precisely diagnosed — is exactly what these prompts recreate. Be the one who receives questions, not answers. In the AI era, that is the side where skill accumulates.

References

• Benjamin Bloom (1984), "The 2 Sigma Problem"
• Roediger & Karpicke (2006), "Test-Enhanced Learning" — the testing effect
• Stephen Krashen — the comprehensible input hypothesis
• Khan Academy — Khanmigo (an AI tutor case)
• Anthropic — Claude for Education