The Feynman Technique for Deep Understanding
Explaining a concept in plain language exposes exactly what you do not understand — how the technique works, where the four steps came from, and how to run it well.
By Velloreth
There is a test for whether you understand something, and it is not whether you can recognize the right answer when you see it. It is whether you can build the explanation from scratch, in plain language, for someone who does not already share your vocabulary — and keep going when they ask why.
The version of this idea circulating as the Feynman technique is four steps long and takes about a minute to state. The steps are almost beside the point; what makes the method work is the specific discomfort it produces, because that discomfort is a measurement.
The four steps, as popularized
- Pick a concept you want to learn or check.
- Explain it in plain language, as if to someone with no background — writing, speaking, or at a whiteboard.
- Mark every point where you stall: a place you hedged, reached for jargon, restated the claim instead of giving the reason, or silently skipped.
- Return to the source for exactly those points, repair them, then explain again from the beginning — not from your notes.
Repeat until the explanation runs clean without jargon crutches.
Step four has a detail people skip: the second explanation starts over. Patching the broken sentence and moving on lets you keep the surrounding scaffolding that made the original collapse; restarting forces the whole chain to hold weight again.
What Feynman actually left on record
The four-step formulation is a later popularization. No writing or recording of Feynman’s has been found that lays it out — researchers tracing its origin place the earliest full statement around 2011, five decades into the afterlife of his reputation. Before that, “Feynman technique” mostly meant his diagrams.
What is on record is the standard the steps operationalize:
- The freshman-lecture test. David Goodstein, a Caltech colleague, recounted asking Feynman to explain why one family of particles behaves as it does. Feynman said he would prepare an introductory lecture; days later he returned and said he could not do it — he could not reduce it to the freshman level, which meant, in his words, that they did not really understand it. Failure to simplify was treated as evidence about the theory, not about the audience.
- The notebook of ignorance. Before his oral qualifying examination, Feynman did not review the outlines of known physics. He opened a fresh notebook titled “Notebook Of Things I Don’t Know About” and rebuilt each branch from memory, looking for raw edges and inconsistencies. The notebook was not much use for the exam; the process was the point.
- The rephrase test. In a 1966 lecture he described checking understanding by asking someone to restate what they had learned in their own language, without using the new word they had just acquired. If they could only repeat the term, they had learned a label.
The common thread: explanation is treated as a probe, not a presentation. A smooth performance tells you nothing; a failure tells you where to work.
Why producing an explanation works
Three mechanisms, each with support in the learning literature:
It is retrieval practice. Building the explanation from memory exercises the same route you will need when the knowledge is required without prompts — the effect that meta-analyses of practice testing find consistently ahead of rereading. Reading your notes before explaining converts the exercise back into recognition, which is the version with the weaker evidence.
It is self-explanation. The act of saying why step B follows from step A — rather than that it does — forces you to supply causal links the source may have left implicit. Dunlosky and colleagues’ review of ten study techniques rated self-explanation and its close relative elaborative interrogation as moderate utility: promising and generalizable enough to recommend, with less classroom evidence than practice testing and spacing. You are using them as the inspection tool, not the entire method.
It makes gaps local. Vague unease about a topic (“I don’t really get distributed systems”) cannot be acted on. A specific stall — you cannot say what happens on a hash collision — is a to-do with an address. The technique converts diffuse unease into a repair list.
Worked example: explaining a hash map
Take something you have used a hundred times and try the pass:
“A hash map gives constant-time lookup because it hashes the key and uses that to find the value.”
That sentence is where most explanations die, because it is true, complete-sounding, and hollow. Keep going as if someone kept asking why:
- What does the hash do? It maps the key to an integer.
- Then what? The integer is turned into a position in an array.
- What if two keys produce the same position? — Stall. This is collision handling, and you either know chaining from open addressing here or you do not.
- Why is lookup constant time on average? Because the number of comparisons does not grow with the number of entries — but only if the hash spreads keys evenly and the load factor stays bounded.
- What happens when the array fills? Resize and rehash, which is expensive — so how is lookup still amortized constant time?
Two minutes of that produced a concrete list: collision strategies, load factor, amortized cost of resize, what a bad hash does to a chain. Each item is a small, findable gap. Compare that with rereading the chapter on hash maps, where every paragraph looks approximately familiar and none of the specific holes announce themselves.
Then repair only the list — back to the source for those items — and run the whole explanation again from the top, without looking at what you just wrote.
A second example: why HTTPS is secure
Different domain, same pass. Start with the sentence most people would give:
“HTTPS encrypts your connection so nobody can eavesdrop.”
True, hollow. Push on it:
- What does “encrypts” mean here? A cipher turns the data into something only someone with a key can read. — Which key? How do both sides agree on it without the eavesdropper seeing it? Stall. This is the key-exchange problem, and most people stall here.
- So how do they agree? The server has a private key, and its certificate links the public key to a domain. The client verifies the certificate against a trusted root. — But if the public key is public, can’t an attacker just use it to decrypt? No — the actual session key is generated by the client and encrypted with the server’s public key; only the server’s private key can decrypt it. — Why bother with certificates at all? Because a public key over an unauthenticated channel does not prove who you are talking to; the certificate chain is what binds the key to the domain.
- What is the session key actually used for? Symmetric encryption of the rest of the connection, because symmetric ciphers are much faster than the asymmetric operations used only during setup.
The gap list this time: asymmetric vs symmetric roles, certificate verification, forward secrecy (if you go further — how does TLS 1.3 avoid reusing a compromised long-term key?), and what an attacker in the middle can still observe (timing, sizes, SNI). Repair, then explain again from scratch. Notice the second pass always goes deeper than the first — that is the technique working, not a sign that the first pass was wasted.
Simplifying without dumbing down
The failure mode of step two is producing a summary in easier words. Plain language is not the goal by itself; the mechanism must survive the simplification.
Tests that an explanation has to pass:
- Causal, not declarative. “It uses a hash function” is a label. “It converts the key to an index so the array slot can be computed directly instead of searched” says why the trick works.
- Conditions stated. If your explanation is true only under assumptions — uniform hashing, a warm cache, an idealized network — say which. An explanation that silently drops its conditions is a claim that will fail in production.
- Analogies defended. Analogies are load-bearing only if you can also say where they break. “The hash is like a shelf number for your books” stops being useful the moment someone asks about two books with the same number; if you cannot extend it to collisions, it is decoration.
- Predictions possible. You understand it if you can say what should happen when something changes — bigger inputs, a slower link, a different distribution. If you cannot predict, you have memorized a description.
The freshman-lecture standard does not mean removing the hard parts. It means the hard parts have to be reachable by steps a beginner could follow, which is harder than using the term and moving on.
Formats that make the test real
The constraint is producing the explanation with no escape hatch. Increasingly demanding versions:
- Write it — a paragraph or a page, from a blank file.
- Record it — speaking catches the pauses and hedges that writing lets you edit around.
- Whiteboard it — diagrams force structure; you cannot fake a data flow with prose.
- Teach it live — a colleague or a study partner will ask the follow-up you were avoiding.
- Answer unprompted questions — after explaining, generate the three hardest questions someone could ask and answer them. The third question is usually where the real gap is.
If you can only pick one, write from a blank file: it is slow enough to be honest and requires no scheduling with anyone else.
Failure modes
- Teaching vocabulary. If the explanation is mostly terms defined in terms of other terms, nothing was explained. Apply the rephrase test: no protected words.
- “It depends” as a stop sign. Sometimes the honest answer is contextual — but you must be able to say what it depends on and how the outcome shifts. “It depends” with no variables is a place you stopped thinking.
- Reading your notes first. The gap-finding pass is a recall test. Notes turn it into a copying exercise.
- Skipping the second pass. Repairing three gaps and never rebuilding the chain leaves you with a patched explanation held together by the notes you used to make it.
- One pass and done. The first clean explanation often reveals the next layer of holes — you now notice subtleties you lacked the vocabulary for earlier. Clean once means understood-so-far, not finished.
- Confusing fluency with understanding. Being able to narrate a process smoothly while it runs is not the same as handling it when it breaks. Explanation identifies the gap; it does not close it alone.
Timing: how long a session should take
The first pass on a moderately familiar concept takes five to fifteen minutes of writing. A clean explanation of something you know well takes less. The repair pass, where you go back to the source, is variable — it is the part that expands as you discover more.
If your first pass is taking an hour, the concept is probably too large: break it into two or three sub-concepts, explain each, then explain how they connect. “Explain how TCP works” is a session; “explain the three-way handshake” is a pass. The technique works at the granularity where you can finish a first pass without running out of patience.
A useful stopping rule for a session: stop when your gap list has three concrete items you could go repair right now. More than three and the list stops being actionable; fewer and you probably did not push hard enough.
Combining it with the rest of the system
The technique slots neatly into a study loop rather than replacing one:
- Encode once — read or work through the material with attention.
- Explain from memory immediately — this is your first retrieval pass, and it produces the gap list.
- Repair the list at the source.
- Explain again from scratch, days later, not the next minute — spacing the retrieval is where retention compounds, and the same research that supports active recall supports the delay between attempts.
- Interleave — after two or three concepts, explain them in one session in mixed order, so you also practice deciding which ideas apply, not just narrating each in its own lane.
Where it stops being the right tool: raw factual recall, which spaced repetition schedules more efficiently than repeated explanation; and procedural skill — you cannot explain your way to being able to do. For those, describe it once to find the gaps, then go practice, and come back to explain what surprised you.
The measure of the method is not the elegance of the final explanation. It is the number and quality of the gaps the first failing pass exposed — because that is the only part of the process that tells you, precisely and without flattery, what you did not know.