How to Learn With AI: A 6-Step Method From a Real Session

To learn with AI, don't ask it for answers. Ask it to explain one idea at a time, then quiz you so you answer in your own words.

I tried this on November 9, 2026. I used Claude on the Max plan, inside Claude Code. I learned six project management concepts from the PMBOK Guide. The definitions here are as taught in my course, which used the 7th edition. Below are the method, a prompt you can copy and my real answers.

Key Takeaways

• Learn one concept per round, not a whole topic at once.
• Give the AI your guess first, or ask it to correct a common misconception.
• Ask for an analogy from your own job.
• End every round with one question you must answer in your own words.
• In my session, every one of my six answers had a gap. The gaps were the lesson.

Why doesn't asking AI for answers help you learn?

Asking AI for answers does not work because a clear answer feels understood while you read it, and reading never proves you can explain the idea yourself.

I had this problem with project management. The course was done and my notes were saved. But I had no record of explaining any of it in my own words. I had collected the material without checking my understanding.

So I set the AI up as a tutor, not a search engine. A tutor explains, then makes you talk.

What is the 6-step method for learning with AI?

The method is six fixed steps per concept: a one-line definition, a corrected guess, an analogy from your own job, three core ideas, an application to your work, and one check question you answer in your own words.

1. One-line definition. What the thing is, in one plain sentence.
2. Correct a guess. I state what I think it means. If I have no guess, the AI corrects a common misunderstanding instead.
3. An analogy from my job. I'm an architect, so I asked for building and design examples.
4. Three core ideas. Only three. Anything more waits for the next round.
5. Apply it to my work. Where this idea would change something I actually do.
6. One check question. A scenario I must answer in my own words. No yes-or-no questions.

In my session, I had a real guess only once. For the other five concepts, the AI corrected a common misconception instead. One example was "stakeholders are just the client who pays."

In my experience, step 6 did the most work. Every gap I found showed up there.

What does a real AI learning session look like?

A real session runs one concept per round and ends each round with a scenario question you answer without notes; I worked through six concepts in one evening, and two of those rounds are below, translated from Korean.

Round: tailoring

The AI explained tailoring as adapting a standard method to one project. Its analogy was a standard building specification versus the specification written for one specific site.

Then it asked me a scenario question. A design competition penalizes floor areas that are off by more than 3%. The team has three people and four weeks. What would you add, and what would you cut?

My answer: "Build an area-check system. Cut my personal free time."

The feedback was direct. Adding an area check was right. But I also needed to say when to run it: on every version and two days before the deadline.

Cutting free time was wrong. Tailoring adjusts the process to the project. You can add checks, and you can cut paperwork, meetings and procedures. You don't cut people's rest.

Round: value delivery

The question gave me two options right before a deadline.

• Option A: area error 0.5%, building rules met, but no space open to residents.
• Option B: area error 2.5%, inside the 3% limit, with half the ground floor open to residents.

In the scenario, the client's main request was a building open to residents.

My answer: "Option B. It's within the limit, it passes the rules, it gives residents open space, and it follows every guideline."

I picked the right option for a weak reason. Option A also passed the limit and the rules. Passing the rules only gets you into the comparison. Value decides between the options that pass.

Only one of my reasons actually counted: the open space, because that was what the client asked for.

What did all six rounds look like in one table?

All six rounds ended with a gap, and the missing part was always the action rather than the label: when to run the check, how to respond, or what to share. On a phone, swipe the table sideways to see every column.

Concept	Check question	My answer (translated)	What was missing
PMBOK structure	Which management area would you cut most, and which can never be cut?	"Building codes can never be cut. What can be cut depends on reading the brief carefully."	Codes are a hard limit, not a management area. Named nothing to cut
Tailoring	Add one thing and cut one thing	"Build an area-check system. Cut my personal free time."	When to run the check. Cut paperwork, not people's time
Value delivery	Option A or Option B, and why?	"Option B. It's within the limit… it follows every guideline."	Used rule-passing as the reason
Life cycle	Where do two decision gates go, and what gets expensive after each?	"One at the site layout and one at the floor plan. Changing the layout turns the whole plan upside down."	Why the second gate matters: boards, area tables and renderings must be redone
Uncertainty	Classify three situations and respond	"1. Ambiguity 2. Issue, threat 3. Opportunity."	Labels right. No response for any of them
Stakeholders	How do you engage judges you can't meet, and your firm's principal?	"I just don't meet the judges directly. I hold a meeting the principal attends and brief them."	Following the no-contact rule isn't engagement. Share with the principal before the meeting

What pattern showed up in my own answers?

The pattern in my own answers was that I described what I would not break instead of what I would do, and the AI only surfaced it in the last round of the session.

That pattern matters for my job. I'm good at following the rules. I need to get better at saying how we win once the rules are met.

I would not have seen this from notes alone. It only showed up because I had to answer in my own words.

What prompt should you copy to start?

This prompt sets the AI up as a six-step tutor that stops after each check question, and you paste it at the start of a new chat with the brackets replaced.

This prompt condenses the six-step instructions I saved in Claude Code for my session. I haven't yet run this exact text as a pasted prompt.

I want to learn [topic]. I am a [your job] and I know [nothing / the basic terms / some practice].

Teach me one concept at a time in this order:
1. A one-line definition.
2. Correct my guess: say which part of my guess is right and which part is wrong. If I give no guess, correct a common misunderstanding.
3. An analogy from my job as a [your job].
4. Only three core ideas.
5. How it applies to my real work.
6. One scenario question I must answer in my own words. No yes-or-no questions.

After I answer, tell me what was right, what was missing and what was wrong. Then fill in a short concept card with my answer copied word for word. Then wait. Do not move to the next concept until I say so.

Start by showing me a short map of the topic and suggest 3 to 6 concepts in a sensible order.

How do you keep what you learn with AI?

A concept card keeps your own answer to the check question word for word, so what you review later is what you understood rather than the AI's summary.

Field	What goes in it
Concept	The name
One-line definition	From step 1
First guess and the correction	From step 2
My own words	My answer to the check question, unedited
Status	Heard → Explained → Used at work
Next review	1, 3, 7, then 21 days later

Here is my real tailoring card, shortened and translated:

• Concept: Tailoring
• Definition: Adapting the methods, steps and tools of a standard to one project. It can add as well as cut.
• Correction: I said to read the brief. I also needed to look at the team's size, time and experience.
• My own words: "Build an area-check system. Cut my personal free time."
• Status: Explained
• Next review: November 10, 2026

The status column keeps me honest. "Heard" means I listened. "Explained" means I answered in my own words. "Used at work" needs a real file or task where I applied it.

The review schedule is my own rule, not a proven formula. I ask the AI to quiz me only when I say so, and if I miss a card, it goes back to one day. As of November 9, 2026, I haven't done my first review.

Where does this method break down?

This method has been tested on one learner, one session and one paid tool only, and the list below records its known limits.

• This is one learner, one session, one tool. I ran the method only in Claude Code, with saved instructions. I haven't tested the pasted prompt in a plain chat or in other chatbots.
• I only ran it on a paid plan. The method is plain text and needs no special feature. But I used the Max plan and haven't tried a free plan.
• The AI can be wrong. My course notes were saved on my computer, and Claude Code could read them. Its explanations pointed back to those notes. Without notes, ask where each claim comes from and check it.
• Analogies can mislead. An analogy helps you start. It is not the definition.
• It is slower than reading answers. Six concepts took one evening. Whether I remember more, I'll only know after my reviews.
• The questions feel like a test. That is the point, but it can be tiring. One concept per round keeps it light.

Who is this method best for?

This method works best for a reader who already has material but no confidence in it. Course notes, a book, or a topic you keep hearing about at work all fit.

It suits less well if you need one quick fact. For that, a normal question is fine.

What is your first step today?

Your first step is to pick one topic you already have notes on. Paste the prompt, type your honest guess about the first concept, and answer the first check question without looking at your notes.

About the author: JH-A is an architect who tests AI tools on real office tasks.

This post was drafted with the help of AI writing tools, based on my own tests and notes.

I also use AI to keep work notes between tools: How I Make Claude Code and Codex Share One Worklog. More about how I test tools: About Charrette AI Room.

Related guides

Two more posts turn one round of this method into a short, checkable practice session:

• How to Gamify Boring Tasks: A 15-Minute Mission Template
• REA for AI Agents: What It Can Prove About an App

Frequently Asked Questions

Can I use the six steps with a free AI chatbot?

Yes, the method is plain text and needs no special feature. I only ran it in Claude Code on the Max plan, so I have not tested the pasted prompt in a plain chat or on a free plan.

What if I have no guess about a concept?

Say so and ask the AI to correct a common misconception instead. In my session I had a real guess only once, so the AI corrected a misconception for the other five concepts.

How long does one round take, and how many concepts should I cover?

One concept per round, and the round ends when you answer the check question in your own words. Six concepts took me one evening, which is why I kept every round short.

Should I trust the AI's explanations without my own notes?

Not without checking. My explanations pointed back to notes saved on my computer; without notes, ask where each claim comes from and confirm it against a source you trust.