大模型网摘deepseek.wang
中文

AI Usage Guides · 2026-11-08

Feynman-Style Self-Testing with DeepSeek: Two Copy-Paste Templates

The real test of learning something is not "having read it" but "being able to explain it and get the questions right." DeepSeek handles both well: the R-series deep thinking mode is suited to breaking down hard concepts, while the V-series general chat is more than enough for everyday Q&A and quiz-style self-testing. This piece gives you two templates you can use as-is — one that forces it to explain a concept until you truly get it, and one that quizzes you in return. One ground rule first: deep thinking models don't respond to roleplay flourishes like "you are a gentle, patient teacher"; they respond to clear constraints and structured requirements. And for simple questions, there's no need to trigger deep thinking at all.

Step 1: Use the "explain it to me" template

The core of the Feynman technique is "explain it in plain language, as if to a beginner." In practice, you can have DeepSeek play the person you're explaining to, feed it your understanding, and let it poke holes. But for many complex concepts, the more efficient order is to have it explain first — using strict output structure, not a pile of adjectives.

text
Please explain [concept name, e.g., Bayes' theorem] to me.

Requirements:
1. First, in 2-3 sentences, what problem does it solve;
2. Give a real-life example and walk through the core logic;
3. Point out 2 of the most common misconceptions and why they're wrong;
4. End with a "30-second elevator explanation," as if you had half a minute in an interview.

My background is [your level, e.g., high school math / just finished intro probability]; match the examples to that level.

If you switch on the R-series deep thinking mode, the breakdown tends to be more organized, especially for math, algorithms, and long logical chains. The rule of thumb is simple: turn on thinking when the concept is nested with "whys"; for purely memorization material (terms, dates, categories), just ask with the V-series directly.

Step 2: Use the "quiz me" template

Once it's explained, immediately move to the second half of the Feynman loop: being tested. Having DeepSeek generate questions beats flipping through a textbook yourself, as long as you pin down question types, difficulty, and grading rules.

text
Based on the topic I just studied, [topic], give me 5 self-test questions:

Requirements:
1. Mix: 2 concept-discrimination questions, 2 application questions, 1 open explanation question;
2. Ask one question at a time; wait for my answer before the next;
3. After each answer, first point out which parts are right and which are wrong, then give the correct reasoning;
4. After all 5, give me a summary: which types I got wrong, what the underlying knowledge gaps are, and which parts I should go back and re-study.

Grading: no politeness — incomplete answers lose points, and vague phrasing counts as wrong.

The beauty is in item 4: the quiz ends with an automatic "error diagnosis" that tells you exactly what to go back and patch, no guessing required.

Run two loops, not one

Getting a bunch wrong on the first pass is normal — don't jump topics yet. Feed the wrong answers back and have it re-explain those gaps and generate variant questions. Two or three rounds until your accuracy climbs, and then you've actually got it.

text
I got questions [X] and [X] wrong. Please:
1. For just these two gaps, re-explain each with a fresh example;
2. Generate 2 variant questions per gap, slightly easier than before, so I can solidify the basics first.

Three rules for toggling modes

DeepSeek's open weights make local deployment possible, so this learning workflow can become part of a private setup; and with low API pricing, long high-frequency self-testing sessions won't stress the budget. The templates are above — pick a topic you have an exam on next week and run one loop tonight. You'll find that "did I actually learn it" finally gets a measurable answer.