Turn any concept into a self-contained interactive HTML study page: two parables from unrelated domains teach one idea indirectly, you pick the mechanism they secretly share before the term is ever named, then a staged set of instant-feedback multiple-choice questions walks you from comprehension → common gaps → analogy discrimination → transfer, capped with flashcards for active recall. Every question locks on your first click and tells you right/wrong + why on the spot.
Like Socrates, it doesn't hand you the answer — it reflects your own understanding back at you.
A Claude Code / openclawmp skill. The skill ships Chinese-first for the openclawmp marketplace, but the generated study pages are language-adaptive — a Chinese concept yields a Chinese page, an English concept an English one. 中文说明见 README.zh-CN.md.
"AI explains, you nod" is the most common illusion of competence — recognizing ≠ understanding. Re-reading quietly reinforces that illusion; forced choice with immediate correction breaks it. Socratic Mirror makes every interactive stage a multiple-choice question whose distractors are built from the real misconceptions for that concept — so a wrong pick is caught and explained the instant you make it, and you can't advance a stage until you've answered every question in it. Grounded in well-supported learning principles: analogical transfer (two isomorphic stories make you abstract structure, not surface), the testing effect with corrective feedback (graded retrieval beats rereading), and active recall (the flashcards).
Note: this version has no free-text input — it's recognition-with-feedback by design. It deliberately does not claim the self-explanation effect (which requires the learner to produce a long explanation).
Two stories (same machine, unrelated surfaces) → Predict (pick the shared mechanism before it's named — two stories beat one because you're forced to strip the surface and grab the structure) → Reveal (the term, finally named; both stories mapped back) → Check (comprehension MCQs) → Gaps (MCQs probing where people most often get it backwards) → Analogy (pick the analogy that's truly isomorphic; the feedback shows where the wrong ones break and where the right one has edges) → Active recall (flashcards) + Transfer (apply the concept to a brand-new scenario).
Every MCQ gives instant per-question feedback, locks on first answer (commit, then see the rationale), and gates the stage on all answered — not all correct (the why already teaches a wrong pick). Progress auto-saves in the browser.
Why two parables? One story binds the concept to its surface details ("comparative advantage is the chef story"). Two structurally-identical stories plus a "what do they share?" question force the learner to abstract the transferable structure — the classic analogical-transfer lever.
- Two-parable contrast + Predict — identify the shared mechanism across two unrelated stories before the term is named, so you abstract structure instead of memorizing surface.
- All-MCQ, instant feedback — every interactive stage is multiple-choice; each wrong pick names the misconception on the spot. Distractors are concept-specific, not throwaway.
- Stage gates — you must answer every question in a stage to move on.
- Self-contained — one HTML file, works offline, no API key, no dependencies.
- Keyboard & screen-reader friendly —
←/→walk the stages, the stepper is a real ARIA tablist, options are an ARIA radiogroup, cards flip withEnter/Space. - Print to PDF —
⌘/Ctrl+Pexpands every stage (your answers + the correct one + the rationale) into a clean, archivable worked sheet. - Language-adaptive — the study page follows the concept's language.
- Resumable — per-concept progress in
localStorage.
openclawmp install skill/socratic-mirror --target-dir ./skillsOr use it as a Claude Code skill — drop the folder into your skills directory.
Trigger with /learn <concept> (also /feynman, /parable, or natural language like "teach me X" / "help me really understand X"). It asks at most 1–2 questions only if the concept is too broad, then writes ./<concept>.html — open it in a browser and work through the seven stages.
| Path | What |
|---|---|
SKILL.md |
The skill's instructions + 7-step workflow (written in Chinese — it targets the openclawmp audience) |
references/ |
The pedagogy + authoring rules the skill loads at runtime (Feynman method as a question-picking lens, parable method, anti-AI-slop, the HTML data contract) |
assets/tutor-template.html |
The single-file interactive study page (vanilla JS, warm-graphite theme) |
.metadata.json |
openclawmp asset manifest |
The two parables do the encoding (get the idea in via structure, not surface); the staged MCQs do the diagnosis (catch where you'd actually trip and fix it on the spot). The Feynman four-step is still here — but as the lens for choosing what to test (explain-simply → comprehension, find-gaps → gaps, analogize → analogy, plus transfer), not as a free-text exercise. The discipline that replaces the old write-first lock: you answer every question, you commit before you see the rationale, and wrong answers are taught, not just marked.
MIT © Li Xuan (krux3009)