The Feynman Method: Why You Forget 90% of What You Read (And the 4 Prompts That Fix It)

Why a clear paragraph fools you

You read something yesterday. Can you explain the actual insight now, not just name the topic? Most people can't, and the article's first move is to say that's normal brain behaviour, not stupidity.

It leans on Hermann Ebbinghaus's 1885 memory experiments: without reinforcement you lose roughly 42% after 20 minutes and up to 80% after a month. That is the forgetting curve (memory decays on a predictable schedule unless you refresh it). The deeper trap is the fluency illusion (mistaking the ease of reading something for actually understanding it).

A clear paragraph feels like knowledge, but it's only recognition: you've seen the idea, you haven't used it. The line that earns its quote is the whole article in one sentence — "Reading is proof of attention. Explaining is proof of understanding."

Feynman's notebook of things he didn't know

Feynman's trick wasn't studying what he knew; it was hunting what he couldn't yet explain. Preparing for his Princeton exams he opened a notebook titled "Notebook Of Things I Don't Know About" and worked through each gap one at a time.

His students later wrote it up as four steps: pick one concept and put it on a blank page; explain it as if teaching a 12-year-old, plain words only; find every place you go vague or stuck and mark it; go back to the source, close the gap, and lock it with an analogy.

Each step forces the one thing reading alone never does, which is to pull the idea back out of your head. The method is content-free, so it works the same on physics, finance, or a term you nodded at in a meeting.

The science that backs it

Feynman never tested his method; 60 years of cognitive science did it for him, and the piece names three findings.

The Protege Effect (you learn more by teaching than by studying): Koh, Lee and Lim (2018) found students who taught material scored 10 to 20% higher, because teaching forces retrieval. Retrieval practice (pulling information out beats cramming more in): Karpicke and Blunt (2011), in Science, found self-testing beat rereading on every measure. Desirable difficulty (Robert Bjork's term): learning that feels easy is shallow, and the discomfort of typing your own clumsy explanation is the moment the memory hardens, not a sign you're failing.

Three settled results, one direction: teaching beats studying, retrieval beats rereading, difficulty beats comfort.

The four prompts, and Claude's real job

The workflow turns those three effects into four Claude prompts you run in about 20 minutes. Prompt 1 (Concept Map) extracts the five load-bearing ideas from a topic, on the logic that "most topics have 50 facts; only 5 matter." Prompt 2 (the 12-Year-Old Test) writes a plain model answer, then stops and makes you type your own version, which is the retrieval rep that does the actual work.

Prompt 3 (the Gap Finder) is the part a book can't do: Claude grades each of your explanations STRONG, WEAK, or WRONG against the source and names the single highest-leverage gap to restudy. Prompt 4 (the Analogy Lock) ties each idea to two everyday analogies, shows where each one breaks, then hands you one summary sentence to re-read tomorrow.

The load-bearing idea is Claude as gap-finder, not author: its real value is scoring your retrieval against ground truth, not writing the answer for you.

The uncomfortable thesis, and the case against it

That maker-then-checker split is worth naming across three worlds. In software it's LLM-as-a-judge (a second model graded only on catching the first one's mistakes); in finance it's the four-eyes principle (a second reviewer signs off before anything is final); here it's the same move pointed at a human learner, with the model acting as four-eyes on your own understanding.

The piece's sharpest claim is that "people who use AI for active recall will out-think the people who use AI to summarize" — a direct challenge to any reading setup whose output is a summary you skim.

The honest case against it is that the method can quietly automate the very illusion it exists to break: the model writes the answer, you read it, feel fluent, type a near-paraphrase, and a polite grader marks it STRONG. It only works if the Gap Finder is genuinely strict and actually has the source to judge against, because a model judge can be confident and wrong. Used honestly, it converts a passive summary into one real retrieval rep, and that is the keeper.

Vocabulary

• forgetting curve — memory fades on a predictable schedule unless you actively refresh it
• fluency illusion — mistaking how easily you read something for actually understanding it
• active recall / retrieval practice — pulling an idea out of your head; it sticks far better than rereading
• Protege Effect — you learn a thing more deeply by teaching it than by studying it
• desirable difficulty — if learning feels easy it is shallow; the effort is where memory forms
• LLM-as-a-judge — a second model whose only job is grading the first one's output

If you're building — what to watch for

• If you are building a reading or research pipeline, notice the grain this cuts against: summarise-and-surface optimises for a fast skim, and the piece argues the skim is exactly where the learning leaks out.
• The reusable pattern is maker-then-checker: one pass produces, a separate pass grades against ground truth. It works on agents and on humans, and the checker only earns its place if it actually holds the source.
• A grader that cannot see the source, or that is tuned to be agreeable, will mark a paraphrase as correct, and you have re-automated the illusion you were trying to break. Strictness and grounding are the whole feature.
• The cheapest upgrade to any summary step is a retrieval step: make the reader type the idea back before showing the model answer. Ninety seconds of typing beats another paragraph of prose.

Reading it critically

• "Forget 90%" is oversold. Ebbinghaus's ~80%-at-31-days figure is for nonsense syllables with zero reinforcement; meaningful, connected material decays much slower. The number is real, the framing is not.
• One study (Koh 2018, a 10-20% gain) gets stretched into "remember it for years" — a real effect with an overclaimed magnitude. Check the study's scope before quoting it.
• The whole method rests on the Gap Finder being accurate. A model grading your explanation "against the source" assumes it has the source and judges strictly; a polite one marks a paraphrase STRONG.
• The "18-month window to out-think everyone" line is motivational packaging on a salesy thread, not evidence. Keep the method, drop the urgency pitch.