Agent Skills: AI-Corrected Feynman Technique

AI-Corrected Feynman Technique

Implements [[AI纠错费曼学习法]] (/Users/zhaoliang/LocalDocuments/vaults/vault/pages/AI纠错费曼学习法.md): the classic Feynman step 4 ("teach it to someone") needs a real listener willing to push back. Claude plays that listener — the user explains, Claude corrects, repeat until nothing's left to find.

Non-negotiable ground rules

• Never explain the concept first. If Claude explains before the user does, the whole point — forcing the user to expose what they don't actually understand — is defeated. If asked to "explain X" in a context that matches this skill, prompt the user for their own explanation instead.
• Don't just praise. Every round needs a real critique pass: factual errors, hand-wavy/vague spots papered over with jargon, causal claims that don't actually hold up. If a round is genuinely correct and complete, say so plainly — but check for real gaps first, including subtle ones (wrong category/analogy, mechanism described as its effect, edge cases not covered).
• Watch for parroting. If a re-explanation is suspiciously close to Claude's own prior correction wording, that's not evidence of understanding — say so and ask for it in the user's own words/own example, don't score it as a real pass.
• Ground corrections in verifiable facts when available — prior conversation context, docs, code actually read/run — not just Claude's prior knowledge. A correction backed by "we tested this earlier and it took 6s, not ~0s" lands harder than a bare assertion.
• Score every round, 1-10, with a one-line reason. The score should track structural understanding (does the mental model hold up) more than completeness of wording.
• Continue rounds until either the user says to stop, or a round earns 10/10 with nothing left to correct — mark that round's conclusion as **结论：通过**.

Flow (single concept — first-time learning)

1. Identify the concept and where the log should live. If the concept already has a note in the current vault (e.g. pages/<Concept>.md), log there. If not, ask where it should go — don't assume a new file location. Default assumption (unless the user says otherwise): log gets written into the concept's own page under a ## 费曼纠错记录 section, not a separate log file.
2. Ask the user to explain the concept in plain language, without looking up material. Wait for their answer — do not proceed on their behalf.
3. Critique it per the ground rules above. Give the 1-10 score with a reason.
4. Ask whether to log this round now or wait, then append it to the concept page using the template below (see Log template) — never overwrite prior rounds, only append.
5. If not yet correct, ask the user to revise and re-explain (informed by the critique) and repeat from step 3, incrementing the round number.
6. On the final passing round, add **结论：通过** after that round's score line, then initialize spaced-repetition scheduling for this concept (see Review mode), since a first pass is itself the first SM-2 repetition.

Log template

Append this section to the concept's page if it doesn't already have one:

## 费曼纠错记录

参见 [[AI纠错费曼学习法]]。自己先讲、AI 纠错、修正重讲，直到讲对为止——每轮都保留，不覆盖前一轮，方便看到理解是怎么一步步修正的。

Then for each round, append (never edit/overwrite earlier rounds):

### 第 N 轮 — YYYY-MM-DD

**我的解释：**
<用户原话，逐字或整理后的转述都行，不要替用户改写立场>

**AI 纠错：**
- <具体错误/含糊点/站不住的因果关系，没有就写"无——这句复述准确">

**理解度评分：** X/10 — <一句话理由>

On the concluding round, add directly under that round's score line:

**结论：通过**

Chain mode (learning chain)

Use when the user wants to work through a set of related concepts in order — e.g. "把这些概念做成学习链一步步学"、"继续那条学习链" — rather than just one concept in isolation. Chain mode is the single-concept flow above, run repeatedly over an ordered list, with progress tracked in a dedicated chain note so it survives across sessions.

Building a chain

1. If the user hasn't already given (or previously agreed to) an ordered, categorized concept list, help them build one: group concepts into stages by dependency (foundational concepts a later concept's explanation would otherwise have to re-derive come first), not just topic similarity.
2. Write the chain to its own note: pages/<链名> 学习链.md. One ## section per category, each concept as an unchecked checklist item linking to its own page:

---
aliases: []
created: YYYY/MM/DD
tags:
  - ai
  - learning-chain
title: <链名> 学习链
---

# <链名> 学习链

<一两句话说明这条链是什么、覆盖范围来自哪里>

## <分类1>

- [ ] [[概念A]]
- [ ] [[概念B]]

## <分类2>

- [ ] [[概念C]]

3. Don't duplicate per-round Feynman content into the chain note — it stays a thin index. The actual record of understanding lives on each concept's own page under ## 费曼纠错记录, per the single- concept flow above.

Running a chain

1. Find the resume point. Read the chain note top to bottom; the next concept to work on is the first unchecked - [ ] item. (Checkbox state is the source of truth for "where we are" — if it's ever out of sync with a concept page's own **结论：通过**, trust the concept page and fix the checkbox.)
2. Run that concept through the single-concept flow (steps 2–6 above), same ground rules — no shortcuts because it's part of a chain.
3. On that concept's pass, edit the chain note: flip its checkbox to - [x]. Then ask whether to continue immediately to the next concept or stop here for today.
4. On stop, leave the chain note as-is (the flipped checkbox is already the correct resume marker) — no other bookkeeping needed. A later "继续那条学习链" just re-enters at step 1.
5. Don't jump ahead to a later concept while an earlier one in the chain is still unchecked, unless the user explicitly asks to skip it — the ordering encodes a dependency, not just a preference.

Review mode (SM-2 spaced repetition)

Passing a concept once doesn't mean it's retained. Review mode schedules and runs spaced repetition on concepts that already have **结论：通过**, using the classic SM-2 algorithm (Wozniak), tracked directly in each concept page's frontmatter — no external app needed.

Frontmatter fields (per concept page)

review_due: YYYY-MM-DD     # next date this concept is due for review
review_interval: N          # current interval in days
review_ease: N               # ease factor × 100, e.g. 250 = EF 2.50 (SM-2 default start)
review_reps: N               # consecutive successful repetitions (resets to 0 on a failed review)

Initializing on first pass

The moment a concept earns **结论：通过** for the first time, this is SM-2's first repetition. Set: review_reps: 1, review_interval: 1, review_due = tomorrow's date, and review_ease from the quality of that final passing round (map the 1-10 Feynman score to SM-2 quality 0-5, roughly quality = round(score / 2)), via the EF update formula below with a starting EF of 2.50.

After writing the frontmatter, also create the SP task for this concept — see SP sync.

SM-2 update formula

On every review (first pass included), given quality q (0-5) and prior EF (default 2.50 if none):

EF' = EF + (0.1 - (5 - q) * (0.08 + (5 - q) * 0.02))
EF' = max(EF', 1.3)

if q < 3:
    review_reps = 0
    review_interval = 1
else:
    review_reps += 1
    if review_reps == 1: review_interval = 1
    elif review_reps == 2: review_interval = 6
    else: review_interval = round(review_interval * EF')

review_ease = round(EF' * 100)
review_due = today + review_interval days

A q < 3 result resets the repetition streak — treat it as evidence the understanding didn't actually stick, not just a bad day.

Running a review session

Trigger on things like "今天有什么要复习的"、"开始复习"、"review".

1. Find due concepts. Prefer sp_get_tasks filtered on project: "费曼复习" (and filter: "today" or "all" as appropriate) — it already aggregates due concepts across every vault, so this is normally faster than scanning each vault's pages/*.md for review_due <= today. Fall back to scanning pages directly if SP is unavailable, a vault's concepts aren't synced yet, or the user scopes the request to one vault/chain by name.
2. Per due concept, run a short recall check — lighter than first-time learning: ask for a quick explanation (a few sentences, not a full teach-back), same anti-parroting and real-critique rules apply. This is not a fresh multi-round drill by default; it's "can you still produce this on demand right now."
3. Judge quality 0-5 from how the recall went:
   • 5 — correct, fluent, no hesitation, nothing to correct
   • 4 — correct with a minor imprecision, self-caught or trivially fixed
   • 3 — correct after one real clarifying correction
   • 2 — recalled the shape but got a real mechanism wrong; needed real correction (reps reset)
   • 0-1 — couldn't reconstruct the core mechanism at all (reps reset)
4. Apply the SM-2 update above and write the new review_due/review_interval/review_ease/ review_reps into that concept's frontmatter. Then reschedule the SP task to the new review_due — see SP sync.
5. If quality < 3, don't stop at the quick check — the gap is real, so run that concept through a full correction round appended to its ## 费曼纠错记录 (same as first-time learning steps 3-5), not just a frontmatter update, since the understanding needs actual repair, not just rescheduling.
6. Log the review outcome on the concept page under a ## 复习记录 section (separate from ## 费曼纠错记录, which stays for the original learning rounds):

## 复习记录

| 日期 | 质量(0-5) | 结果 | 新间隔 |
|------|-----------|------|--------|
| YYYY-MM-DD | q | 通过 / 需要重新纠错（见费曼纠错记录第N轮） | N天 |

Append one row per review, oldest first — never overwrite prior rows.

SP sync (cross-vault review inbox)

Concept pages live scattered across vaults (AI-vault, survey-vault, etc.), so review_due alone can't be scanned in one place. SP (Super Productivity, via the sp-mcp tools) is that one place — a task per concept, all filed under a single shared project so the SP planner view is a cross-vault review inbox. SP frontmatter in each concept page stays the source of truth for SM-2 state (review_ease/review_reps/review_interval); the SP task is just a scheduling pointer to it.

• Project: 费曼复习, shared across all vaults — do not create a per-vault project.
• Tag: 费曼复习 (via sp_create_tag if sp_get_tags doesn't list it yet).
• Title: 复习: <概念名> (the concept's page title / filename without extension).
• dueDay: the concept's current review_due.
• notes: the concept page's path relative to its vault, prefixed with the vault name, e.g. AI-vault/pages/Tokenizer.md — this is what lets a review session jump back to the right file.

Creating (on first pass — see Initializing on first pass): call sp_create_task with the fields above.

Rescheduling (on every later review — see Running a review session): call sp_update_task or sp_schedule_task with task: "复习: <概念名>" and the new dueDay. If the title match is ambiguous (SP returns multiple candidates — this can happen if the same concept name exists in more than one vault), disambiguate using each candidate's notes path against the concept page you're actually updating, then retry with the resolved task id.

If SP isn't running (sp-mcp tools error out), don't block the Feynman/review flow on it — finish the frontmatter update, tell the user the SP sync failed, and move on; do not silently skip mentioning it.

Reference example

/Users/zhaoliang/LocalDocuments/vaults/survey-vault/pages/dlt.md's ## 费曼纠错记录 section is a worked example — 4 rounds on dlt (6/10 → 8/10 → 9/10 → 10/10 通过), including a correction that was verified against a benchmark run earlier in that conversation rather than asserted from memory alone.