feynman-learning

Feynman Learning Method

Restructure the content of any subject into one complete Feynman learning loop. The core object is not the textbook, but the learner's current explanation. The learner must go through "explain first, expose gaps, rebuild on their own, rephrase, transfer and apply," ultimately leaving behind a "Feynman Learning Record" and the goal for the next round.

First load /stage-design, and follow its new classroom creation, roster, page-by-page persistence, and audio acceptance processes. This Skill only specifies the teaching method, page brief, and interaction rhythm.

Scope of application

Use for the following requests:

• Use the Feynman method to learn or teach a specific concept;
• Transform courseware, lesson plans, videos, or materials into Feynman-style learning activities;
• Design a class centered on teach-back, self-explanation, probing questions, and transfer.

Do not use for:

• Popular-science classes whose main knowledge thread is "what is the Feynman Learning Method";
• Cases where you only want to make learning strategies a parallel objective of a concept class, in which case use /learning-to-learn；
• Brand style replication or direct PPT import, in which case use /style-clone or /pptx-import. Imported materials can still serve as the factual source for this Skill.

Core principles

1. The learner speaks first; the AI intervenes afterward. Before presenting authoritative definitions, textbook explanations, or standard processes, first collect the learner's own explanation. Without an initial explanation, there is no Feynman loop.
2. Fix only the smallest gap at a time. From the explanation, identify the 1–2 currently most critical breakpoints; do not dump the entire textbook on the learner all at once.
3. Probing questions take priority over answering for them. First use the fewest possible questions to help the learner complete the causal chain on their own. Only when they are still stuck after multiple probes and the user permits it should necessary facts or scaffolds be provided.
4. Understanding must withstand rewriting and transfer. Being able to recite terminology does not equal understanding; the learner must be able to remove terminology, explain the boundaries of analogies, and re-explain in unfamiliar situations.
5. Materials only provide facts and do not preempt the initial explanation. Materials uploaded by the teacher are used to determine conceptual boundaries and verify facts, not as the standard answer to be displayed directly on the first page.

Page content red lines

The following words belong to planning language and may only be written in dialogue, briefs, or roster design; they must not appear as student-visible titles, labels, or sections:Teach-back、Jargon Challenge、Transfer、v1/v2/v3、概念定界、缺口诊断、类比破坏测试、费曼循环、概念变化轨迹。

Student-visible titles must be written as executable actions, for example:

• "First explain it to someone who has never learned it";
• "At which step did you get stuck?";
• "Replace these few words and explain it again";
• "Where will this analogy fail?";
• "Try it in a new situation."

The teacher's and student agent's oral narration, probing questions, and discussion go into narration / actions, and are not written as static page text such as "the teacher says" or "Xiaobai says."

First determine three constraints

Before the page plan, clarify:

1. concept boundary: which learnable core concept this round only explains, and explicitly what it does not cover;
2. audience: to whom the learner will explain it, for example a middle school student who has not learned the concept or a layperson;
3. teacher constraints: the depth of learning, and whether the AI is allowed to give the answer after probing questions fail.

Only when the request is insufficient to determine these constraints, use ask_user to fill them in.

Feynman loop and page mapping

The cycle can span pages 7–9. Do not mechanically chase a page count, but the following evidence chain must not be missing. The first page should merge the conceptual boundary and the first explanation into a single interactive task, avoiding revealing the standard content before the initial explanation.

1. First explain it to a layperson (interactive)

• Briefly state this round's conceptual boundary and target audience, without giving the standard definition;
• Ask the learner to complete the first explanation in their own words without consulting materials;
• Prefer to provide text or voice input. If the component cannot save free input, ask the learner to first write down or say their own version, and have the teacher's voiceover remind them to keep it for later comparison.

2. Find the smallest sticking point (interactive or slide)

• Break the initial explanation into the necessary conceptual elements;
• Identify only 1–2 of the most critical skipped steps, term substitutions, incorrect causality, or omissions;
• Let the learner judge for themselves “where I am stuck,” not just show an AI score.

Write the concept checklist into the brief and turn it into questions the student can answer; do not render internal labels such as “Gap 1” and “Gap 2.”

3. Use questions to fill in the causal chain (interactive, or a slide ending with a discussion action)

• Each question targets only one identified gap;
• The novice agent keeps asking “why” and “what would happen if the conditions changed”;
• The brief states what the question targets and what evidence of understanding should stop it;
• Do not give the answer directly before the teacher constraints are met.

4. Explain it again (interactive)

• Give a few structural hints and ask the learner to generate a second version of the explanation;
• The hints must point to the causal structure and must not restate the standard answer;
• Keep the comparable differences between the first and second versions for the final learning record.

5. Remove the terms just used (interactive)

• Select the terms the learner actually used from their second explanation;
• Ask them to explain again without using these words, or give counterexamples that expose rote memorization;
• Do not preemptively show a blacklist of terms unrelated to the learner's expression.

6. Break your own analogy (interactive)

• Have the learner propose an analogy, then separately explain “where it is similar” and “where it is not similar”;
• Force out at least one point where the analogy fails;
• The page does not preemptively give all failure points; the agent's follow-up questions help the learner discover them.

7. Switch to an unfamiliar situation (interactive or quiz)

• Transfer the concept to a new situation that has not been used earlier;
• You may restrict use of the concept name, forcing the learner to rely on the mechanism for explanation;
• It is best to add a counterexample, boundary condition, or changed assumption to test whether they truly grasp the structure.

8. Synthesize the Feynman learning record (slide or interactive)

The final record uses the learner's own words and includes at least:

1. what I can explain now;
2. how I understood it at first;
3. where I found myself stuck;
4. my new version of the explanation;
5. my analogy and its failure points;
6. new problems I can solve;
7. what I still cannot explain, and the goal for the next round of learning.

It is not a celebration page saying “you learned it,” but the starting point for the next cycle.

Roster

• Teacher / Coach: Keep the order of explaining first and giving feedback after, decide when to provide scaffolding, and pull the discussion back to the core concept.
• Novice / beginner agent: As the target of explanation, genuinely express that you do not understand and ask about causes, serving as the main gap detector.
• Optional misconception agent: Present a common but incorrect explanation to help the learner distinguish it; do not let it reveal the standard answer in advance.

Scaffolding and acceptance

In the order of /stage-design create the class:create_stage → set_roster → page by page according to the approved plan generate_scene → list_scenes → read_stage acceptance. When a page needs its narration supplemented or rebuilt, use generate_actions, and ensure that the final speech actions all have audio; after modifying the narration text, call generate_tts。

Before completion, check page by page:

• the initial explanation occurs before any standard content;
• each page can state what it externalized, what it exposed, or what it reconstructed;
• diagnosis focuses only on 1–2 minimal gaps;
• follow-up questions are not surreptitiously replaced with directly answering for the learner;
• term stripping targets words the learner has used themselves;
• the analogy test produces at least one failure point;
• transfer leaves the original case and includes a boundary or counterexample;
• static pages have no planning labels, pedagogical jargon, or character lines;
• the final record includes both what has been understood, remaining gaps, and goals for the next round.

Capability boundaries

Generative pages cannot dynamically diagnose every sentence a learner inputs in real time. Use the three paths truthfully:

1. interactive collect explanations and support learner self-checking with checkpoints;
2. the roster agent and teacher narration simulate Socratic questioning;
3. when sentence-by-sentence diagnosis and continuous preservation of explanation versions are truly needed, run the loop in the workbench conversation, using the class pages as the activity skeleton.

Do not describe static or pre-generated pages as capable of real-time scoring.