Learn Loop結構化學習迴圈

Get taught and quizzed on a concept, then write the note yourself while it verifies sources and files it into your Obsidian vault先教後考,筆記你親手寫,它負責查證來源、挑洞、歸檔進 Obsidian

繁中User-invoked使用者叫用
Read SKILL.md on GitHub在 GitHub 看 SKILL.md

Learn Loop is a structured learning loop that teaches you a concept from verified primary sources, quizzes you on it, and then has you write the permanent note yourself while it checks your draft against the sources. It never writes the note for you. This skill is invoke-only — it never auto-triggers just because a concept comes up in conversation; it has to be called explicitly.

Install

npx skills add https://github.com/leoluyi/skills -g -a learn-loop -y

Update later with:

npx skills update learn-loop

Source

Invoke it with /learn-loop <concept> in Claude Code, $learn-loop in Codex, or by explicitly asking to run the learn-loop process.

What it does

The skill runs on a 先教後考 + 來源查證 mechanism: it teaches the concept from sources it has verified (primary sources outrank secondhand write-ups, and it will say so rather than fabricate a citation), then switches into examiner mode and asks retrieval questions you answer in your own words. Only once your answers hold up does it move to distillation.

At distillation, the roles flip. You close the explanation and write the note from memory, in your own words — the skill's only job at that point is to poke holes: where the draft disagrees with the sources, where it's vague, where it isn't atomic. It never ghostwrites the note. The act of writing the distillation from memory is the learning itself, so that step is never delegated back to the skill — it acts strictly as a practice partner and fact-checker, not an author.

When to use

Reach for it when you want to genuinely learn a new technical concept and crystallize it into a permanent Obsidian note, not just skim an explanation.

When not to

Skip it for reorganizing an already-distilled document into Diátaxis blocks (use knowledge-doc-writing instead), or for a quick one-off explanation with no note to keep.

How it works

The loop runs six steps in order, one at a time, waiting for your response after each (they don't have to happen in one sitting):

  1. Capture the question — opens a working note in the vault's 00-inbox/, titled after the concept, and records what you want to understand and why.
  2. Ground — sweeps recently captured material (notes, PDFs, screenshots) related to the concept, treats brought-in material as primary source or researches it via web search, and cross-references the vault for concepts you already know as anchors. Produces a literature note with source links.
  3. Teach, then test — explains the concept concisely from verified sources tied to your existing anchors, then asks 3-5 retrieval questions and waits. Gaps send it back to step 2 for more grounding.
  4. Distill — you write the note from memory with the explanation closed; the skill only critiques for accuracy, vagueness, and atomicity, and can propose a claim-style title.
  5. Promote & connect — decides evergreen vs. reference filing, adds Templater frontmatter, wikilinks to the anchors found in step 2, and flags which MOC it belongs under. Your words are authoritative; the skill only polishes formatting.
  6. Schedule a revisit — adds the new note to the vault's weekly-review retrieval queue so you restate the claim from memory before checking it again later.

If the machine has no vault, it falls back to a portable temp-vault package you merge back by hand, with a MERGE.md checklist and a zip archive instead of step 6.

The skill only writes the distilled note itself — it does not reorganize existing documents. Once a note exists and later needs restructuring into Diátaxis blocks, that hands off to knowledge-doc-writing, not to another learn-loop run.

Related skills

  • knowledge-doc-writing — takes over once a note is distilled and needs reorganizing into tutorial/how-to/reference/explanation blocks; learn-loop only handles the interactive learning and initial write, never the restructuring.
  • obsidian-vault — general-purpose search/create/organize operations over the same vault; reach for it directly when you just want to find or file a note, without the six-step teach-then-quiz loop.

Learn Loop 是一套結構化的學習迴圈:先用查證過的一手來源教你一個概念、考你回想題,再讓你憑記憶親手寫下永久筆記,過程中它負責挑洞比對來源。它從不代寫筆記。這個技能只在明確叫用時啟動——對話裡提到某個概念不會自動觸發,必須明確呼叫。

安裝

npx skills add https://github.com/leoluyi/skills -g -a learn-loop -y

之後更新:

npx skills update learn-loop

原始碼

叫用方式:Claude Code 用 /learn-loop <要學的概念>,Codex 用 $learn-loop,或直接明說「跑 learn-loop 流程學 X」。

它做什麼

這個技能運作在「先教後考 + 來源查證」的機制上:先用查證過的來源教你概念(一手來源優先於二手轉述,查不到就明說,絕不捏造連結),接著切換成考官模式,問你幾道回想題,你要用自己的話回答。只有答對了,才會進入 distill(精煉)階段。

到了 distill 階段,角色互換。你要闔上剛才的解釋,憑記憶用自己的話寫下筆記——技能此時唯一的工作是挑洞:草稿哪裡跟來源不符、哪裡含糊、哪裡不夠原子化。它從不代寫筆記。憑記憶寫下 distillation 這個動作本身就是學習,所以這一步永遠不會外包回技能去做——它扮演的是練習夥伴與事實查核者,不是代筆者。

何時使用

當你想真正學懂一個新技術概念,並把它結晶成 Obsidian vault 裡的長期筆記時使用,而不只是想瀏覽一段解釋。

何時不要

不要用於把已經 distill 完的文件重整成 Diátaxis 區塊(改用 knowledge-doc-writing),也不要只是想要一次性的口頭解釋、沒有筆記要留存的情況。

運作方式

整個迴圈依序跑六個步驟,一次一步,每步都等你回應(不需要一次跑完,可以分好幾次進行):

  1. 擷取問題——在 vault 的 00-inbox/ 開一份工作筆記,以概念命名標題,寫下你想搞懂什麼、為什麼現在關心這件事。
  2. 紮根(Ground)——掃過最近累積的相關素材(筆記、PDF、截圖),你帶來的素材視為一手來源,沒有的話就上網研究,同時比對 vault 裡既有的相關概念作為錨點。產出一份附來源連結的文獻筆記。
  3. 先教後考——用查證過的來源,扣著既有錨點,簡潔講解概念,接著問 3-5 道回想題後停下等你作答。答不出來的部分會退回步驟 2 補材料。
  4. 精煉(Distill)——闔上解釋,你憑記憶寫下筆記;技能只針對正確性、含糊處、原子性挑洞,並可以提議一個主張式(claim-style)標題。
  5. 歸位與連結——判斷歸入 evergreen 還是 reference,補上 Templater frontmatter、連到步驟 2 找到的錨點 wikilink,並指出該掛在哪個 MOC 下。內容以你的文字為準,技能只潤飾格式。
  6. 排入複習——把新筆記加進 vault 的 weekly review 複習佇列,之後複習時你要先憑記憶複述一次主張,再對照筆記核對。

如果這台機器沒有 vault,會改走可攜式的 temp-vault 備援方案,產出 MERGE.md 併回清單和一個 zip 包,取代步驟 6,由你事後手動併回真正的 vault。

這個技能只負責把 distill 出來的筆記寫進去,不會重整既有文件。等筆記寫好之後如果要重構成 Diátaxis 區塊,那是交給 knowledge-doc-writing 去做,不是再跑一次 learn-loop。

相關技能

  • knowledge-doc-writing——筆記 distill 完、需要重整成 tutorial/how-to/reference/explanation 區塊時接手;learn-loop 只管互動學習與初次寫入,從不做重整。
  • obsidian-vault——同一個 vault 的通用搜尋/建立/整理操作;只想找筆記或單純歸檔,不需要六步驟先教後考流程時直接用它。