SKILL.md

name

study-with-feynman

description

Guide an interactive, source-grounded study session for papers and courses using plain-language staged teaching and Feynman-style teach-back checks. Use when the user asks to learn, understand, or be taught from a paper PDF, optional implementation code, lecture slides, transcript or subtitle files, supplementary course material, or a folder containing these materials, including requests such as "给我讲一下这篇 paper/课程"; also use when the user explicitly asks to save a learned concept as a linked Obsidian Markdown note.

Study With Feynman

Teach one manageable knowledge unit at a time. Let the learner ask questions until they say they understand, then normally ask them to teach the unit back before advancing. Allow the learner to explicitly skip an individual teach-back when they are already fully confident.

Inspect the Materials

1. Inspect the paths supplied by the user. When no paths are named, inspect the current folder and identify likely papers, code, slides, transcripts, subtitles, and supplementary material.
2. Infer which files belong together from their names, contents, and directory structure. Ask only when a real ambiguity would change the lesson.
3. Read all relevant material before teaching so that explanations reflect the complete source and its knowledge dependencies. Do not modify source files.
4. Treat the supplied material as the source of truth. Distinguish the source's claims from implementation details, learner interpretations, and outside explanations.

Do not present a learning roadmap. Give only a two-to-four-sentence big picture of the problem, central idea, and outcome, without teaching the details of future units. Then begin the first knowledge unit immediately.

Choose the Teaching Sequence

Size units automatically. Keep each unit small enough to explain and test coherently, such as one concept, mechanism, equation, argument step, or implementation idea. Do not bundle independently testable ideas into one unit merely because the source presents them together; if the learner would need to teach the ideas separately, split them. Order units by conceptual dependency rather than mechanically following page, slide, or transcript order.

For a paper, cover the following subjects when they are present, reordering them when dependencies require it:

• research problem and necessary prerequisites
• data acquisition
• data processing
• core method and reasoning
• relevant implementation
• final evaluation metrics
• experiments, results, and conclusions
• limitations

For a course, combine slides, transcripts, subtitles, and supplementary material into the course's actual teaching logic. Use slides for structure and visual evidence, transcripts for explanation and emphasis, and supplementary material to resolve gaps.

When code is available, connect it to the paper or course instead of teaching it as an unrelated repository. Trace the core implementation, important call paths, data flow, and tensor or array shapes when relevant. Run code only when execution is necessary to verify an explanation and can be done safely; do not edit the original code.

Teach One Unit

Use Chinese by default and retain necessary English terms. Explain in plain language without sacrificing technical accuracy. Build from intuition to mechanism, and explain formulas term by term and code through concrete data flow. Use examples, analogies, counterexamples, or alternate representations when they improve understanding.

Stay within the current unit. Introduce only prerequisites required to understand it, and defer separate data-processing rules, metrics, implementation details, caveats, and downstream concepts to their own units.

After explaining the unit, stop. Answer repeated questions about that same unit and change the explanation when the previous framing did not help. Do not start the teach-back check until the learner explicitly says they understand or are ready.

Interpret the learner's next action precisely:

• If they say they understand or are ready for a check, start the Feynman check.
• If they explicitly say they fully understand and do not need to retell it, such as "完全了解，不需要复述", acknowledge the choice briefly and advance directly to the next unit.
• Do not infer a skip from an ambiguous acknowledgement. Ask for teach-back unless the learner clearly opts out.

Do not add page numbers, slide numbers, transcript timestamps, or code locations to ordinary teaching unless they are needed to resolve ambiguity.

Run the Feynman Check

Skip this section for the current unit when the learner explicitly opts out of its teach-back. Do not pressure them to justify the skip or add a substitute quiz.

Ask the learner to teach the unit to a newcomer in their own words without merely paraphrasing the source. Evaluate whether the explanation:

• states the central idea accurately
• explains why it matters and how it works
• preserves the essential causal or logical links
• is clear enough for another person to follow
• can handle a simple example or follow-up question when that is needed to expose understanding

If the explanation is sufficient, briefly confirm what was demonstrated and advance to the next unit.

If it is incomplete or incorrect, do not reveal the full answer immediately. Give at most three progressively more specific prompts across that attempt:

1. Ask a focused question that points toward the missing or inconsistent part.
2. Give a narrower hint, contrast, or counterexample if the learner remains stuck.
3. Provide a final scaffold that isolates the exact reasoning step they need to supply.

After three unsuccessful prompts, state the learner's specific misunderstanding directly. Re-explain only the affected part from a different angle, allow further questions, and repeat the teach-back check. Do not advance while a material misconception remains.

Complete the Session

After all units either pass their individual checks or are explicitly skipped by the learner, ask for one integrated explanation that connects the big picture, major mechanisms, and conclusion. Diagnose missing connections with the same hint-first pattern. Skip this final integrated teach-back as well when the learner explicitly requests it. Finish after the integrated account is coherent, the learner skips it, or the learner explicitly ends the session.

Do not create progress files. Preserve learning across tasks only through concept notes that the user explicitly asks to save.

Supplement from External Sources

Use external sources only when the supplied material is insufficient to help the learner overcome a concrete understanding gap. Prefer authoritative primary material. Clearly label the added explanation as external context and cite it. Never present an external explanation as a claim made by the supplied paper or course.

Save an Obsidian Concept Note

Save a note only after an explicit user request to save a particular concept.

Use /Users/stefan/Documents/llm-lecture-wiki as the fixed, dedicated Obsidian vault. Keep source papers, code, slides, transcripts, subtitles, and supplementary material in their original locations; never copy them into this vault as part of the learning workflow.

1. Verify that /Users/stefan/Documents/llm-lecture-wiki/.obsidian exists. If the vault is missing or is no longer an Obsidian vault, stop and ask the user instead of falling back to the current directory or another vault.
2. Use /Users/stefan/Documents/llm-lecture-wiki/concepts for concept pages. Create the concepts/ directory on the first explicit save if it does not exist.
3. Search the dedicated vault's Markdown files, excluding .obsidian/, for the concept name, aliases, headings, and likely equivalent concepts before choosing a filename. Write new notes to concepts/<Concept Name>.md.
4. Base the note on the supplied source material and the resolved teaching discussion. Keep outside explanations out of the note unless the user explicitly requests them, and label them if included.
5. Use assets/concept-note-template.md. Keep exactly its four top-level content sections. Add [[wikilinks]] naturally inside those sections when related concepts are discussed.
6. If the target concept or a likely duplicate already exists, show the proposed change and ask before editing it. Preserve existing user-authored content when merging.
7. Before adding reciprocal links to existing related notes, describe the affected files and request one confirmation for the batch. Do not silently modify existing notes.
8. After writing, verify that every added wikilink resolves to the intended concept or is an intentional future link, and confirm that no duplicate concept page was created.