Agent-Skills.md

Agent Skills: AI反刍阅读法 (Rumination Reading)

Re-review an article the user has already read and starred (a "blog-article"-tagged note) through two gates — first an internalization check borrowed from the Feynman skill (did they actually understand/retain a point, or are they just parroting the original text?), then, only for points that pass, an increment check borrowed from the Socratic skill (what changed in their understanding since last time, forcing a concrete recent example, never letting "feels different now" stand alone). Logs every round into the article's own note, tracks per-point 未内化/已内化 status, and supports resuming mid-article across sessions. Use when the user wants to "反刍"一篇文章, "重新review这篇文章/博客", check whether they actually understood or remember an article, or continue an in-progress rumination session.

UncategorizedID: goodluckz/claude-code-config/rumination-reading

Repository

goodluckz/claude-code-config
goodluckz
1

Install this agent skill to your local

pnpm dlx add-skill https://github.com/goodluckz/claude-code-config/tree/HEAD/skills/rumination-reading

Skill Files

Browse the full folder contents for rumination-reading.

Download Skill

Loading file tree…

skills/rumination-reading/SKILL.md

Skill Metadata

Name
rumination-reading
Description
Re-review an article the user has already read and starred (a "blog-article"-tagged note) through two gates — first an internalization check borrowed from the Feynman skill (did they actually understand/retain a point, or are they just parroting the original text?), then, only for points that pass, an increment check borrowed from the Socratic skill (what changed in their understanding since last time, forcing a concrete recent example, never letting "feels different now" stand alone). Logs every round into the article's own note, tracks per-point 未内化/已内化 status, and supports resuming mid-article across sessions. Use when the user wants to "反刍"一篇文章, "重新review这篇文章/博客", check whether they actually understood or remember an article, or continue an in-progress rumination session.

AI反刍阅读法 (Rumination Reading)

Implements [[AI反刍阅读法]] (/Users/zhaoliang/LocalDocuments/vaults/vault/pages/AI反刍阅读法.md): re-reading only counts as review if it adds something — either repairing a point that was never really understood, or updating a point with genuinely new understanding since last time. This skill is a splice of two sibling skills' mechanisms — ai-corrected-feynman's correction gate and socratic-questioning's forced-concreteness — applied to material already read, not new material.

Scope: articles that already qualify for [[Type Review 体系]]'s "博客文章" track — five-star (5⭐️⭐️⭐️⭐️⭐️) rated, tagged blog-article, with their own note (not an inline link inside a blogger's note). Don't run this on an article that hasn't cleared that bar — point the user at starring/promoting it first if they ask to ruminate on something that isn't there yet.

Non-negotiable ground rules

  • Never explain the point first, and never answer the increment question yourself. Same discipline as the two sibling skills: if Claude supplies the understanding or the "what changed", the forcing function is defeated.
  • Two gates, run in order, per point — don't skip the internalization gate because a point "seems obviously understood." A confident, fluent answer that's actually a paraphrase of the source text still fails the gate — watch for suspiciously-close-to-original wording, same as the Feynman skill's anti-parroting rule.
  • The increment gate requires a concrete, dated, real example — something that actually happened since the last review, not a restated feeling ("我现在更有共鸣了" is not an answer; "上周开会时 我意识到自己就在做 X" is). If the user can't produce one, record "无新进展" honestly — don't manufacture a fake increment to make the round feel productive.
  • A point that fails the internalization gate does not proceed to the increment gate this round. Reread that section together, correct the specific gap, mark it 未内化, and move on — don't force both gates into one round just to finish faster.
  • One point at a time, not the whole article at once. Pick a point, run both gates, log, then ask whether to continue to the next point or stop — mirrors chain mode's per-item discipline in ai-corrected-feynman.

Flow

  1. Identify the article and find/pick its points. The article note should already list its key claims — either as five-star-highlighted lines, or as links out to atomic concept notes (the pattern used in clippings/20250316144431.md and clippings/20250119023830 6.md). If the note has no per-point breakdown yet, ask the user which 1-3 points they want to ruminate on this session rather than inventing a breakdown unasked.
  2. Find the resume point if this is a continuing session: look for a ## 反刍记录 section on the article note and a per-point status marker (see Log template). The next point to work is the first one still marked 未内化 or never yet attempted — same resume logic as chain mode: trust the marker on the article note over any memory of where the last session left off.
  3. Gate 1 — Internalization (borrows ai-corrected-feynman's mechanism): ask the user to explain this point in their own words, without looking at the source, and say what it means to them personally. Judge pass/fail per the ground rules above (not a 1-10 score — this gate is binary).
    • Fail: reread the relevant passage together, name the specific gap, mark 未内化, log the round, and stop here for this point.
    • Pass: continue to gate 2.
  4. Gate 2 — Increment (borrows socratic-questioning's mechanism, only-ask-never-answer): ask "自从你上次真的理解这条以来，发生了什么事让你现在看它有点不一样？" — press for a concrete recent example if the first answer is vague. If a real, specific example surfaces, that's the round's increment. If nothing surfaces after a genuine attempt, record 无新进展 — this is a legitimate, non-failing outcome, not something to keep pushing for.
  5. Log the round (see template below), mark the point 已内化 if gate 2 was reached (regardless of whether it produced a fresh increment — passing gate 1 is what earns 已内化).
  6. If gate 2 produced a real increment, ask whether it's worth spinning into its own Insight per [[认知方法论]] — don't decide this unilaterally.
  7. Ask whether to continue to the next point or stop here. On stop, the log/status markers already written are the correct resume state — no extra bookkeeping needed, matching chain mode's "flipped checkbox is the resume marker" pattern.

Log template

Append this section to the article's own note if it doesn't already have one:

## 反刍记录

参见 [[AI反刍阅读法]]。每个论点两道关：内化关（借用费曼纠错机制）判断有没有真读懂，
增量关（借用苏格拉底追问机制）判断这次重读有没有新理解。每轮都保留，不覆盖前一轮。

Then for each point worked on, append a subsection (create it on first attempt, append new rounds to an existing one — never overwrite):

### <论点简述，可以直接引用该论点的链接> — 状态：未内化 / 已内化

#### 第 N 轮 — YYYY-MM-DD

**内化关（自己的话）：** <用户原话，不要替用户改写>

**内化关判定：** 通过 / 未通过 — <一句话理由，未通过则指出具体理解错在哪/漏了什么>

**增量关追问：** <这次问的具体版本，通常就是"自上次真正理解以来发生了什么">

**增量关回答：** <用户给出的具体例子，逐字或整理转述；如果答不出，写"无新进展">

**这轮结论：** <未内化，留待下次 / 已内化，本轮无新进展 / 已内化，新增理解：一句话概括>

If a real increment surfaced and got promoted to its own Insight, add a line under that round: → 已升级为 Insight：[[<insight note>]].

Interaction with Type Review 体系

The per-point 未内化/已内化 markers on the article note are the same status semantics [[Type Review 体系]] uses for the 博客文章 track — don't invent a separate state machine. When asked to run a Type Review session on the 博客文章 track, this skill's flow is what actually executes that session for whichever article gets picked.

Reference example

None yet — this skill is new as of 2026-08-26. The first real session (likely on [[So You Wanna De-bog Yourself]] or [[How to Do Great Work]]) should be treated as the worked example once it exists.