The Feynman technique with AI, using your own notes

To use the Feynman technique with AI, study one small idea, close the source, explain it in your own words, and ask the AI to question unclear steps. Then reopen the source and check your explanation. Keep each gap as a specific question before trying again.

The useful work is producing and checking your explanation. A fluent explanation written by the AI does not show what you can explain yourself. You can do the same exercise with a friend, a blank page, or a voice recording.

Give yourself a small, checkable source

Here is an invented library system for this exercise. These are its rules, not a description of every real library:

A title can have several physical copies. Each copy has a unique ID. A checkout records one copy ID, one borrower, and a due date. Returning that copy clears its checkout. Other copies of the same title may remain checked out.

The target is “explain why a checkout tracks a copy rather than only a title.” That is more testable than “understand databases.”

When using your own material, keep the source alongside your notes: a textbook section, documentation, or another reference you can actually check. If the note came from an AI explanation, verify the important claims before treating it as your reference.

Close the note and make the first attempt

An initial explanation might be:

The library records who has each title. When someone returns the book, that title becomes available again.

It sounds reasonable. But our example rules distinguish a title from a physical copy, and the explanation has collapsed those into one thing.

A useful listener could ask:

What if two people borrow two copies of the same title and only one returns theirs?

Pause before asking for the answer. Try to work through the case. Which record would change? What identifies the returned object? What stays the same?

This is a constructed example, not an app transcript. An AI may ask a different question or miss the gap. You can ask it yourself.

Ask for questions before corrections

If you use a general AI chat, this prompt sets a clear role:

I am practicing an explanation from memory.
Act as a curious newcomer. Ask one short question at a time
about a term, a missing step, an example, or a boundary case.
Do not explain the topic for me or supply the missing answer.
Wait for my reply before asking another question.
We will check the explanation against my source afterward.

Keep the source closed. If the AI supplies the answer anyway, attempt a fresh explanation before reading more help.

Closing the source turns the exercise into retrieval. In experiments on learning from text, retrieval practice improved later retention compared with additional study under the tested conditions. Related research found benefits on conceptual questions about science texts. These studies support practicing recall; they did not test Graphite or establish that any AI tutor guarantees learning. Roediger and Karpicke, 2006, Karpicke and Blunt, 2011

Turn the stopping point into a question

For our library example, write:

Why must a checkout refer to a copy ID rather than only a title?

This names the distinction you could not explain.

Reopen the source and locate the relevant sentences: every copy has a unique ID, and a checkout records that ID. Follow the example through: copy A returns, so checkout A clears; copy B's checkout remains.

Now make a second explanation with the source closed:

A title describes the work, but the library lends individual copies. Each copy has its own ID. A checkout connects that ID to a borrower and due date. Returning one copy clears that copy's checkout, so another copy of the same title can still be on loan.

Test a changed example: if three copies exist and two are borrowed, what changes when one comes back? Check whether your explanation handles this case.

Separate wrong, missing, and unsupported

Suppose you add: “The system automatically charges a fine when a book is late.” Our source says nothing about fines.

That claim is unsupported by this source. It is not established as false. You need another source or a narrower explanation.

Use three questions when checking your attempt:

• Does any sentence contradict the reference?
• Did I omit a step needed to explain the example?
• Did I add a claim the reference does not establish?

An AI check can miss all three. Read the relevant source yourself, especially when the distinction matters. If your notes contain an error, checking an explanation against those notes can preserve the error.

Keep a reusable record

Copy this into any notebook:

## Topic and source
What am I trying to explain? Where will I check it?

## First explanation, source closed
My own words:

## Question that stopped me
What could I not explain?

## Evidence checked
The relevant source passage, and what I changed:

## Second explanation, source closed
My revised explanation:

## New example
Can I explain the same mechanism in a different case?

## Still open
- [ ] One specific remaining question?

Keep both attempts. The difference helps you see what changed in your explanation. Return to the open question when you next study the topic; no streak or numerical grade is needed to keep that question useful.

Use Graphite to keep the session with the idea

In Graphite's browser notebook, Teach it back lets you explain a note to an AI Student that asks questions. When you finish, the Guide checks the explanation against your notes. It saves your words and the Student's questions in a teaching note, with the feedback, and adds gaps as open questions under What's missing in the note you taught.

Feedback can identify an unsupported claim, ask a question, or suggest a link. It is not an external fact check or a mastery score. The AI is the one you choose in Settings → AI, such as a key of your own (how to get one) or a connected OpenRouter account. In the browser, requests pass through Graphite's server to that AI's provider; the server stores neither your key nor notes.

Graphite does not schedule spaced-repetition reviews or turn a dropped textbook into learning material. Browser notes remain in that browser unless exported. See the learning workflow and its limits and the research behind teaching back. To keep the resulting questions and explanations connected, continue with organizing notes around ideas.