How I Use ChatGPT to Learn Technical Topics Without Copying Answers

I use ChatGPT a lot for learning technical subjects, especially topics where a textbook gives me the formula but I still do not feel that I understand the idea.

For me, the useful part is not getting an answer faster.

It is being able to keep asking:

Why?

What does this symbol mean?

Why is this statement true?

Can you keep using the same example?

Where exactly did my reasoning go wrong?

That changes ChatGPT from an answer generator into something closer to an interactive tutor.

My rule is simple:

I want ChatGPT to reduce the time between being confused and understanding the concept. I do not want it to replace the thinking that creates the understanding.

This article describes the workflow I actually find useful when studying technical material.

I start with the exact thing I do not understand

I usually get better results when I do not begin with a broad request such as:

Teach me regression.

Instead, I give ChatGPT the exact point where I became confused.

For example:

The textbook says residuals are realizations of random errors
assuming the regression model is correct.

I understand what a residual is mathematically,
but I do not understand what "realization" means here.
Explain only that part first.

That is much more useful than asking for an entire chapter summary.

When I am reading a textbook, lecture note, or practice problem, I will often upload a screenshot and point to one sentence, formula, bullet, or answer choice.

The goal is to remove one specific misunderstanding before moving on.

I make ChatGPT keep the same notation

One surprisingly common source of confusion is unnecessary notation changes.

Suppose my textbook uses:

X1, X2, X3

and defines a principal component as:

Z1 = 0.8X1 + 0.1X2 - 0.5X3

If the explanation suddenly switches to:

x, y, z

or introduces a completely different example, I may spend more effort translating the explanation than learning from it.

So I often say:

Continue using exactly the same X1, X2, X3 example.
Do not introduce new notation unless necessary.

For technical subjects, continuity matters.

I ask “why” more than “what”

ChatGPT is very good at returning definitions.

But definitions are often not what I am missing.

For example, I might already know:

LOOCV uses one observation as the validation set at a time.

My real question may be:

Why can LOOCV have higher variance than k-fold cross-validation?

The useful discussion starts when I ask for the mechanism behind the statement.

Other prompts I use are:

Why does this follow from the previous equation?

What would have to be true for this statement to be false?

Give me a counterexample to my interpretation.

I know the formula. Explain the intuition behind it.

Those questions force the conversation away from memorization and toward structure.

I tell ChatGPT my current interpretation

One of my most useful habits is to explain the concept back to ChatGPT in my own words.

For example:

Here is how I understand it:

PCR builds principal components without using Y,
then regresses Y on those components.
PLS uses Y while constructing the directions.

Is that understanding correct?
If not, tell me exactly which part is wrong.

This is much better than repeatedly asking:

Explain PCR and PLS again.

Why?

Because now the model has something specific to diagnose.

It can separate:

what I already understand

from:

the exact misconception that remains

That makes the next explanation much more efficient.

I ask it to explain every answer choice

For multiple-choice practice, I do not find it very useful to receive only:

The answer is C.

I want:

A — why it is wrong
B — why it is wrong
C — why it is correct
D — why it is wrong

This is particularly useful when two choices sound almost identical.

A prompt I use is:

Do not just tell me the correct answer.
Explain what each choice is claiming,
then explain why each one is true or false.

That turns one question into four small concept checks.

I ask for a new problem instead of the solution

When I am stuck on a calculation, there is a temptation to ask ChatGPT to finish it.

Sometimes I do need a worked example.

But a stronger learning method is often:

Do not solve my exact problem yet.
Create a similar problem with different numbers,
and walk through that one first.

Then I return to the original problem and try again.

This is especially useful for:

statistics
regression
ANOVA
probability
linear algebra
machine learning calculations

The important part is that I still have to transfer the method back to the original question.

I use one-question-at-a-time tutoring when I am really stuck

If a topic is difficult, a full explanation can become another wall of text.

In that situation I use a much more restrictive prompt:

Do not explain everything at once.
Ask me one question at a time.
Use my answer to decide the next question.
Do not reveal the final answer unless I ask for it.

This works because I cannot passively scroll through the explanation.

I have to respond.

OpenAI’s Study Mode is built around a similar idea: guided questions, step-by-step explanations, knowledge checks, and active participation rather than only returning a final answer.

I sometimes use explicit prompts like the one above because I can control the style very precisely, but Study Mode is also a useful option when I want the tutoring behavior built into the conversation.

I separate conceptual understanding from arithmetic

Technical study often mixes two different tasks:

understanding the method

and:

performing the calculation

I try not to confuse them.

For example, if I am calculating a standard error or an ANOVA quantity, I may first ask:

Explain why this is the correct formula.

Then I reproduce the arithmetic myself with a calculator or spreadsheet.

If my number differs from ChatGPT’s number, I can ask:

Here are my intermediate values.
Find the first step where our calculations diverge.

That is much more educational than replacing the whole calculation with a generated answer.

I use follow-up questions aggressively

A major advantage of conversational AI is that the first answer does not have to be perfect.

I frequently follow with questions such as:

I still do not get the second part.

Use a smaller numerical example.

Show me where this value came from.

As compared to what?

What is the distance from where to where?

Can you draw the logic as a sequence of steps?

This is important because many technical explanations fail at one hidden assumption.

A follow-up question can expose that assumption.

I ask for visual or geometric explanations when algebra is not enough

Some concepts become much easier once I can picture them.

For example:

principal components
projection
residuals
orthogonality
bias-variance trade-offs

may be mathematically defined in a few equations, but the geometry often makes the idea click.

So I ask:

Explain this geometrically.

or:

Show me what the three distances in this diagram represent.

The best explanation is not always the most formal one.

Sometimes I need the formula. Sometimes I need the picture. Sometimes I need both.

I make it distinguish two ideas that sound similar

A lot of technical confusion comes from terms that are related but not identical.

Examples include:

validation set vs test set

bias vs variance

residual vs random error

principal component regression vs ordinary multiple regression

model quantization vs KV-cache quantization

Instead of asking for two independent definitions, I ask for the boundary:

What is the exact difference between these two?

Then:

Give me an example where A applies but B does not.

That usually reveals the distinction much faster.

I use ChatGPT to generate practice, not just explanations

Once I think I understand something, I want to test whether that understanding survives a new question.

I ask:

Give me one similar problem without the answer.
Wait for my response.

After I answer:

Grade my reasoning, not just the final number.

If I get it right, I may increase difficulty:

Now give me one where the obvious shortcut does not work.

This exposes whether I learned the method or merely copied the pattern from the previous example.

I deliberately challenge the answer

ChatGPT can sound confident even when it is wrong.

OpenAI explicitly warns that ChatGPT can produce incorrect facts, fabricated references, and overconfident answers, and recommends verifying important information from reliable sources.

So when something feels wrong, I do not treat confidence as evidence.

I may ask:

Are you sure?
Verify the claim.

or:

What assumption are you making here?

or:

Show me a primary source that supports this technical claim.

For a textbook problem, the textbook and course materials remain the reference point.

For software behavior, I prefer current official documentation.

For research claims, I prefer the original paper when possible.

I do not use the model’s agreement as proof

This is an important trap.

Suppose I write:

So high correlation always means high variance, right?

A weak conversation can drift into agreement.

A better prompt is:

Do not assume my interpretation is correct.
Try to falsify it first.

That small change makes the interaction more useful.

I want correction, not reassurance.

I sometimes ask for the shortest possible explanation

Long answers are not always better.

When I am stuck on one link in a chain of reasoning, I may say:

Explain this in three sentences only.

or:

Do not give me the whole topic again.
Explain only why step 2 leads to step 3.

This prevents the answer from burying the missing idea under material I already understand.

I sometimes ask for the full derivation

The opposite is also useful.

If I understand the intuition but do not understand where a formula comes from, I ask for every algebraic step:

Derive this formula line by line.
Do not skip algebra.
Tell me which rule is used at each step.

The important point is that I choose the level of detail based on the specific gap in my understanding.

My default prompt pattern

A reusable pattern I like is:

I am studying [topic].

Here is the exact statement/problem I am looking at:
[paste it]

Here is my current understanding:
[explain it in my own words]

Tell me exactly where my understanding is correct or incorrect.
Keep the same notation as the source.
Do not introduce a new example unless it is necessary.
After explaining it, ask me one question to check whether I understood.

This prompt does several useful things at once.

It gives ChatGPT:

context
my current mental model
constraints on the explanation
a requirement to check understanding

That is much stronger than:

Explain this.

My workflow for a difficult topic

When a concept is genuinely difficult, my process is usually:

1. Read the original material.

2. Identify the exact sentence, formula, or step I do not understand.

3. Ask ChatGPT about only that point.

4. Explain the concept back in my own words.

5. Let ChatGPT challenge my interpretation.

6. Work through a small example.

7. Solve a new problem myself.

8. Verify important claims against the original source.

The key step is number 4.

If I cannot explain the idea back without copying the generated wording, I probably do not understand it yet.

What I do not want ChatGPT to do

For learning, I try to avoid using it as:

an answer copier
an essay ghostwriter
a substitute for reading the source
a substitute for calculations I need to know how to perform
an authority that never needs verification

That does not mean I never ask for a complete solution.

A worked solution can be useful.

The difference is whether I use the solution to understand the reasoning or simply to avoid doing the reasoning.

The most useful change is to ask for resistance

If I had to reduce my whole workflow to one idea, it would be this:

Do not configure ChatGPT to agree with you. Configure it to make you think.

Useful instructions include:

Ask me before telling me.

Challenge my interpretation.

Find the first incorrect step.

Give me a similar problem.

Do not change the notation.

Make me explain it back.

Verify the important claim.

Those prompts turn the model into a much better learning tool.

Bottom line

I get the most value from ChatGPT when I use it interactively.

The pattern is not:

Question
↓
Answer
↓
Done

It is:

Confusion
↓
Specific question
↓
Explanation
↓
My interpretation
↓
Challenge
↓
Practice
↓
Verification
↓
Understanding

That takes more effort than copying an answer.

But that effort is the point.

For technical learning, I want AI to shorten the path to understanding without removing the part of the process where I actually have to understand something.