Paper Digest — 單篇論文內容整理（快速吸收）

What this is (and is NOT)

/paper-review answers 「這篇可不可信、做得好不好」. /paper-digest answers 「這篇講了什麼、我怎麼最快把全部內容吸收進腦袋」. It is for papers read as content / knowledge material（綜論、技術介紹、機制整理、指引）. The two are independent and can both run on one paper (combined note). A digest additionally feeds the 複習迴圈: it generates self-test cards that go to the review web page so learning doesn't end at "note written".

Tone, language, and clinical framing are driven by ${persona} (see Configuration). Follow vault note rules (Tab indent, > [!type] callouts, ${persona.highlight_style} for cut-offs, no code blocks for clinical content).

Configuration (read FIRST, every run)

Read config.yaml in the sibling paper-review/ skill directory (shared config; copy config.example.yaml → config.yaml on first setup). Keys this skill uses: ${vault.papers_dir}, ${vault.medicine_dir}, ${vault.inbox_note}, ${vault.daily_section}, ${review_site.url}, ${review_site.push_script}, ${persona}. Substitute every ${...} with the config value; a blank value means "skip that route" (e.g. empty review_site.push_script → no card push).

Input

DOI / PMID / title / PDF path / pasted full text. Optional flags:

• --into <note path> — instead of creating a standalone note, prepend the digest as a ## 內容整理（快速吸收） section into an existing note (used by /paper-sync when both 🔬+📚 picked, so content sits above the appraisal in one file).
• --no-cards — skip review-card generation/push (default is to generate).

Phase 0 — Full text is MANDATORY (hard gate)

Acquire full text per the shared protocol ~/.claude/skills/paper-review/fulltext-acquisition.md (user-provided → ${fulltext.inbox_dir} → Zotero → OA auto-fetch → Elsevier TDM → institutional SFX resolver).

If no route yields full text → STOP. Do NOT write an abstract-only digest. Report: 缺全文：{title} — 需要 PDF 才能做內容整理. When called from /paper-sync this routes to the 缺全文 handoff (logged to ${vault.inbox_note} # 缺全文待補; user supplies PDF then says「繼續」). Abstract-only is exactly the over-simplification this skill exists to avoid.

Phase 1 — Build the digest（三層漸進揭露）

Goal: a reader can absorb all the substantive content without opening the PDF. Reorganize, don't transcribe. Every digest has THREE layers (user's Esor 三層筆記法):

1. 30 秒層 — > [!summary] 一句話 + 重點 callout (between frontmatter and first heading): one-line takeaway + 3–5 bullets of the highest-value facts (mechanisms, numbers, conclusions). End the callout with one quality pointer line: 品質未評讀（內容整理）；需可信度判斷 → /paper-review — or, if a 品質快照 already exists in the combined note, 品質 → 見上方品質快照.
2. 5 分鐘層 — each ## section OPENS with a one-line **要點**：… bold summary before its bullets, so scanning only the 要點 lines reconstructs the paper.
3. 完整層 — the full structured bullets/tables below each 要點.

Section structure — ROUTED BY PAPER TYPE (pick one, don't force-fit)

A. Empirical study（RCT / 觀察性 / diagnostic…） — default:

• ## 研究問題 / 背景 — gap, why it matters, key prior context.
• ## 方法 — design, population, intervention, measures — concise, only what's needed to read the findings (deep methods scrutiny belongs to /paper-review).
• ## 主要發現 — the core. Tabulate numeric results (effect sizes, ==cut-offs==, CIs, p). One row per finding. Pull the actual numbers out of the text.
• ## 臨床意義 / 怎麼用 — so-what for a PMR clinician / teachable points.

B. Narrative review / 機制整理 — concept-map style:

• ## 全文地圖 — how the author carves the topic into sections; one line per block.
• ## 各主題重點 — one ### per theme: core claims, key numbers, the studies it leans on.
• ## 機轉白話講解 — for the hardest mechanism(s), a Feynman-style plain-Chinese walkthrough (analogy allowed); anything beyond the paper's own content marked ⚠️ 補充. Written to be reusable when teaching residents.
• ## 臨床意義 / 怎麼用

C. Guideline / consensus:

• ## 建議條文表 — table: recommendation / 強度 / 證據等級 / 適用族群, one row per recommendation.
• ## 與前版或他版差異 — what changed vs the previous edition or competing guideline (if stated).
• ## 實務落地 — how it maps to the user's practice setting.

D. 技術 / 方法學論文:

• ## 這個技術是什麼 / 解決什麼問題
• ## Step-by-step protocol — restated so it could be followed without the PDF.
• ## 適用時機與限制 — when to reach for it, failure modes, alternatives.

Sections common to ALL types

• ## 重要圖表重述 — restate what each key figure/table shows in words. If a figure is essential, flag it for /figure-remap rather than embedding blindly.
• ## 與既有認知的對照 — run vault_search on the note's core concepts; where an existing ${vault.medicine_dir} / ${vault.papers_dir} note says something this paper updates, refines, or contradicts, list 舊認知（[[note]]）→ 本篇. This is the knowledge-delta layer — also the candidate list for a later /note-supplement. Nothing to compare → one line「vault 無相關既有筆記」, don't pad.
• ## 概念 / 名詞整理 — teaching layer: define and connect the concepts/terms a learner needs. Link related vault notes [[NoteName]].
• ## 自我測驗 — see Phase 2.
• ## Reference — the paper itself (`Author 2026, Journal` + doi); pivotal citations it leans on.

Citation discipline: every claim in the digest comes from THIS paper's full text. Outside context added to explain a concept → mark ⚠️ 補充.

Phase 2 — Self-test cards（主動回憶層）

The user's known failure mode is over-organizing and under-recalling — the digest must end with retrieval practice, not just structure.

1. Write ## 自我測驗 in the note: 3–5 questions as folded callouts —

   > [!question]- Q1：{題目——偏臨床決策/機轉理解，不是背數字}
   > {答案，2-4 行，含關鍵數值與理由}
   

   Question quality bar: answerable from this paper alone; tests understanding ("為什麼選 X 而不是 Y", "什麼情況下這結論不適用") over recall of trivia; one question may target the paper's single most exam/practice-relevant number.
2. Push to the 學習中樞 (unless --no-cards OR ${review_site.push_script} is blank): write the cards to a JSON file in the scratchpad — [{"citekey","note","title","question","answer","tags":["topic",...],"deck":"論文","source":"paper"}, ...] (note = vault filename without .md; title = 中文短標; deck 固定 "論文"、source 固定 "paper") — then:

   python ${review_site.push_script} <cards.json>
   

   The push script is idempotent (card_id = citekey + question hash, INSERT OR IGNORE) and prints the pushed count. Report:「已推 N 張複習卡 → ${review_site.url}」. Push failure (or blank push_script) → note still stands; report it and leave the JSON path for a manual retry — never block the digest on the card push.

Output

• Standalone (default): write ${vault.papers_dir}{中文短標}.md. Frontmatter tags: [research/digest, …topic], citekey, doi, aliases for the English title if useful. Return the note filename to the caller.
• --into <path>: insert the ## 內容整理（快速吸收） section (the body above, minus its own frontmatter) directly after the target note's > [!summary] callout / before its first appraisal heading, so content reads first and the /paper-review appraisal follows. Cards still get pushed.
• One line to today's daily note ${vault.daily_section} is handled by /paper-sync (don't double-write when called from it). Standalone manual runs: add the daily-note line yourself.

Notes

• This is a new-note creation task → normally show a draft for review. Exception when invoked by /paper-sync: write directly (the user already opted in by pressing 📚 and confirming the batch).
• Heavy (full-text + synthesis). Run on a capable model; for a /paper-sync batch, one subagent per paper.
• Don't duplicate /paper-review's appraisal. If the user really wants both, that's the combined note — keep the digest descriptive and leave judgement to the appraisal section.

Self-Check (before finalizing)

• 結構用對 paper type（A/B/C/D），沒有硬套 empirical 模板
• 三層齊：summary callout（含品質指標一行）/ 每節 要點 行 / 完整內容
• 主要發現有實際數值（不是「有顯著差異」）
• ## 與既有認知的對照 跑過 vault_search（或標明無相關筆記）
• ## 自我測驗 3-5 題、folded callout、偏理解型
• 複習卡已推（或回報失敗 + JSON 路徑）