Paper Review — 單篇論文深度評讀

Overview

This skill performs a 360-degree deep dive into a single academic paper, covering journal credibility, author backgrounds, publication history, historical context, reference verification, critical appraisal, and citation impact analysis. The output is a structured Obsidian note in ${vault.papers_dir} plus a daily note entry under ${vault.daily_section}.

Reviews should be clinically grounded, teachable (journal-club ready), and honest about what the evidence can and cannot support. Tone, language, specialty framing, and external-validity checks are driven by ${persona} (see Configuration).

Configuration (read FIRST, every run)

Before Phase 0, read config.yaml in this skill's directory (copy config.example.yaml → config.yaml on first setup). It defines the values this document references as ${section.key}:

• ${vault.papers_dir}, ${vault.medicine_dir}, ${vault.inbox_note}, ${vault.daily_section}
• ${fulltext.*} (inbox, Zotero, Elsevier, SFX resolver, institution) — consumed via fulltext-acquisition.md
• ${secrets.backend} — how bundled scripts read API keys
• ${persona} — language, specialty, audience, locale, reimbursement system. This is user-authored; it sets whether output is 繁體中文/English, which clinical lens, and whether coverage lookups run.

Wherever this SKILL writes ${...}, substitute the config value. A blank/placeholder value means "skip that route". The bundled scripts (grade_judge.py, argdown_lint.py) need no config.

Input

Accepts any of:

• DOI (e.g., 10.1097/PHM.0000000000002700)
• PMID (e.g., 39056432)
• Paper title (partial or full)
• PDF file path (local file)
• Pasted content (abstract or full text in the conversation)

Effort tiers (pick before starting — token budget)

• Quick (daily-note [x] checkbox): Phases 1 + Critical Appraisal + 白話導讀 only. No author/journal deep-dive, no citation analysis, no reference verification. ~1 model, cheap.
• Deep (/paper-review): all 7 phases. Token-heavy (many MCP + WebFetch calls). Before starting, offer the user a focused deep: "全跑 (Phases 1-7) 還是只跑關鍵幾關 (2 期刊/6 引用查核/appraisal)?" Respect a "skip" after any phase. Don't silently run all 7 if the user wanted a spot-check.
• Either tier: reuse journal_cache.json / author_cache.json before any WebSearch.

Delegation (deep tier — added 2026-07-07 per dispatch-protocol)

The main context keeps what is intelligence-bound: reading the paper itself ONCE, the Critical Appraisal, GRADE domain ratings, and final synthesis. Everything that is fan-out lookup work gets delegated so raw search dumps never enter main context:

• Phase 3 (author profiling) → ONE subagent per key author (parallel, same message, model: "haiku"): given name + affiliation, run the S2 author tools + WebSearch, return a ≤6-bullet profile (h-index, top papers, COI/retraction flags). Main context never sees raw search results.
• Phase 6 Steps 2-3 (reference verification) → batch the selected references into 1-2 subagents (model: "sonnet"), each returns verification-table rows only (the exact columns of Step 4). Give each agent the claim text + citation; NOT your appraisal reasoning.
• Phase 7 (citation categorization + PubPeer/social sweep) → one subagent (model: "sonnet") returns the 🟢🔴🔵🟡 categorized list + any PubPeer/news hits with links.
• Every delegation prompt includes: "If you cannot meet the criteria, say FAILED and list what you tried — do not return a best guess as verified." Failed lookups are reported as gaps, never silently dropped.
• Quick tier: no delegation (overhead exceeds benefit).

Phase 0: Full Text Acquisition

Follow the shared protocol: ~/.claude/skills/paper-review/fulltext-acquisition.md (check order, OA auto-fetch rule, Elsevier TDM, institutional SFX resolver, Zotero bookkeeping + citekey).

Review-specific policy: if NO full text is obtainable from any route, proceed with abstract-only review — mark "⚠️ Abstract-level review" at the top of Deep Review, and every section limited by this notes「abstract 未報告，需全文確認」. (Contrast: /paper-digest hard-stops.) For figures/tables when working from a Zotero PDF: point the user to the PDF in Zotero.

Phase 1: Paper Identification & Metadata

Use Semantic Scholar MCP as primary (more reliable), PubMed MCP as supplement:

Semantic Scholar MCP (primary):

• search_papers or get_paper_details → citation count, fields of study, TL;DR, DOI
• Semantic Scholar's paper ID is needed for later phases (citations, references, authors)

PubMed MCP (supplement — may be flaky, session errors common):

• search_articles → find PMID if not already known
• get_article_metadata → MeSH terms, publication dates, abstract
• If PubMed MCP fails (session terminated / invalid content), fall back to WebSearch site:pubmed.ncbi.nlm.nih.gov "{title}" to get PMID

Extract and confirm: title, all authors (with affiliations), journal, DOI, PMID, study type, received/accepted/published dates, keywords/MeSH terms.

After this phase, show the user a brief summary — confirm this is the right paper before proceeding with the deep dive.

Phase 2: Journal Credibility

Assess the journal's trustworthiness and standing:

• Impact metrics: WebSearch for current Impact Factor, SJR ranking, CiteScore
• Indexing: Confirm presence in PubMed/MEDLINE, Scopus, Web of Science
• Peer review process: Single-blind, double-blind, or open review
• Reputation in field: Where does this journal sit in the specialty hierarchy? (e.g., for PM&R: AJPMR > Archives PMR > J Rehabil Med > smaller specialty journals)
• Red flags: WebSearch "{journal name}" retraction watch and "{journal name}" predatory to check for controversy, inclusion in Beall's list, or unusual retraction rates
• Retraction/correction check (this paper) — SINGLE authoritative check, done once here: mcp__zotero__scite_check_retractions on the DOI catches retractions, expressions of concern, and corrections in one call. This is the canonical paper-level integrity check; Phase 4 reuses this result instead of re-searching. Only if scite is unavailable, fall back to one WebSearch "{DOI}" retraction OR "expression of concern" OR erratum.

Output: 2-4 bullet points summarizing journal credibility with specific metrics.

Phase 3: Author Profiling

Focus on first author and corresponding author (often the senior PI). For other authors, brief mentions only.

Delegate this phase (see §Delegation) — one parallel model: "haiku" agent per key author. Each agent runs, and returns only a ≤6-bullet profile from:

• search_authors + get_author_details → h-index, total citations, paper count
• get_author_top_papers → their most-cited works (shows research trajectory)
• WebSearch → institutional affiliation, lab/research group, recent grants
• Check for COI: look at the paper's COI declaration section + WebSearch for industry ties
• Red flags: WebSearch "{author name}" retraction to check history
• Any tool failure → the agent reports "FAILED: " for that bullet, not a guess

Output: a brief profile for each key author — credentials, expertise match to this paper's topic, and any concerns.

Phase 4: Publication Timeline

From metadata already gathered:

• Received → Revised → Accepted → Published dates
• Calculate review duration (received to accepted). Flag if unusually fast (<30 days for a non-letter/editorial) or slow (>12 months)
• Reuse the Phase 2 scite retraction/correction result — do NOT re-search. Note here only if a correction/erratum post-dates publication (relevant to the timeline).
• Check if there's a related commentary or editorial in the same journal issue

Output: timeline with any notable observations.

Phase 5: Historical Context & Background

This phase explains why this paper exists — the research landscape that motivated it:

• Read the paper's Introduction carefully for the narrative arc
• Use Semantic Scholar get_related_papers and get_paper_references to map the intellectual lineage
• Identify:
  • The specific research gap or controversy this paper addresses
  • 3-5 key prior works that directly led to this study
  • The field's trajectory — is this an emerging area, a mature debate, or a paradigm shift?
• For each key prior work, get a one-line summary via get_paper_details

Output: a narrative (in bullet form) tracing the intellectual path from prior work to this paper.

Phase 6: Reference Verification (Key Claims Only)

This is the most distinctive part of this review. The goal is to check whether the paper accurately represents its sources — a common and often undetected problem in the literature.

Step 1: Identify key claims

Read through the paper and identify 5-10 claims that are central to the paper's argument. These are typically in the Introduction (establishing the rationale) and Discussion (comparing results).

If the paper has >20 references, ask the user which claims or sections to focus on. Don't auto-select all.

Adaptation for Systematic Reviews / Umbrella Reviews

For SR/meta-analysis/umbrella reviews, Reference Verification takes a different form:

• Instead of checking individual citation accuracy, assess the quality of included studies
• Check whether the search strategy was comprehensive (databases, date range, grey literature)
• Verify the quality assessment tool used (AMSTAR-2 for SRs, Cochrane RoB for RCTs)
• Flag if conclusions are stronger than the quality of included evidence supports
• Check for overlap in primary studies across included SRs (umbrella reviews)
• Run the actual PRISMA 2020 checklist against this paper (not just "did they cite PRISMA") — walk the 27 items (title/abstract/introduction/methods/results/discussion/funding sections; full item list at prisma-statement.org if needed) and mark each Reported / Partial / Missing. Report only the gaps (don't pad a table of 27 ✅ rows) — missing protocol registration, incomplete search strategy reporting, or no risk-of-bias-across-studies assessment are the highest-yield flags. This is a reporting-completeness check on the paper's OWN methods section, separate from AMSTAR-2 (which judges methodological rigor).

Step 2: CrossRef existence gate (anti-hallucination, do this FIRST)

Before any semantic comparison, run a cheap deterministic existence check per selected reference — this catches fabricated DOIs and "right DOI, wrong paper" swaps that a purely semantic read misses (ported from htlin222/robust-lit-review's crossref.py, the same anti-hallucination gate DeepTRACE found necessary — LLM-written citations run 40–80% accuracy unaudited):

• WebFetch https://api.crossref.org/works/{doi} (no key needed). 404 / no message → does not exist — flag ❌ fabricated/unresolvable DOI, do not proceed to semantic comparison for it.
• If it resolves, compare the returned title against the reference's title as cited in the paper (simple token-overlap read, not exact-match — journals reformat titles). Low overlap → flag ⚠️ DOI resolves to a different paper before trusting the citation at all.
• No DOI given for a reference → note "no DOI, existence unverified" and rely on Semantic Scholar/PubMed title search as the fallback existence check.
• Hard rule — existence pass is never a claim-level pass: machines catch nonexistence, not misquotation. A CrossRef pass only unlocks Step 3; it must never shortcut or substitute the semantic comparison. If CrossRef itself is down, mark "existence gate skipped" on affected rows and still run Step 3 — degrade to explicit uncertainty, never to a silent pass.

Step 3: Verify each claim (semantic layer, only for references that passed Step 2)

Delegate Steps 2-3 as a batch (see §Delegation): split the selected references across 1-2 model: "sonnet" agents; each returns only completed Step-4 table rows (+ explicit "FAILED" rows where a lookup died). Main context assembles the table and judges borderline verdicts.

For each key claim and its cited reference:

• Use Semantic Scholar get_paper_details or PubMed MCP get_article_metadata to get the reference's abstract
• Compare what the paper says the reference found vs. what the reference actually says
• Classification:
  • ✅ Accurate — faithful representation
  • ⚠️ Minor discrepancy — slightly overstated, simplified, or missing nuance
  • ❌ Misrepresented — the reference doesn't support the claim, says something different, or is taken out of context
• If a reference looks suspicious at abstract level, try to get full text (same acquisition strategy, but only for suspicious ones)

Step 4: Output verification table

#	Reference	CrossRef	Claim in Paper	Reference Actually Says	Verdict
1	Author2020	✅ exists, title matches	"X causes Y"	"X correlates with Y in subset"	⚠️ overstated

Critical Appraisal (integrated after Phase 5-6)

Step 1 — Route by study design (MANDATORY, do not use a generic checklist)

Identify the design first, then appraise domain-by-domain with the matching tool. Each domain gets 🟢/🟡/🔴 + a one-line justification, then an overall risk-of-bias judgement.

study_type	Tool	Domains to walk through
RCT	Cochrane RoB 2	randomization process / deviations from intended interventions / missing outcome data / outcome measurement / selective reporting
Cohort / case-control / cross-sectional	ROBINS-I (interventions) or Newcastle-Ottawa (etiologic)	confounding / selection / classification of exposure / missing data / outcome measurement / selective reporting
SR / MA	AMSTAR-2 critical domains + GRADE per outcome	protocol registration / search adequacy / RoB of included studies / meta-analysis methods / publication bias / heterogeneity handling
Diagnostic accuracy	QUADAS-2	patient selection / index test / reference standard / flow & timing
Prediction model	PROBAST	participants / predictors / outcome / analysis (EPV, validation)
Guideline	AGREE II (精簡: rigor of development + editorial independence)	evidence base / recommendation-evidence link / COI management / update policy
Narrative review	SANRA + cherry-picking check	justification / literature search description / referencing of key statements / are contradicting studies acknowledged?
Case series / report	JBI checklist	consecutive inclusion / clear criteria / complete reporting

Observational studies additionally: read the bundled causal-appraisal.md (35 Hernán + Users' Guide doctrines, folded into this skill — no external dependency) and check the applicable ones explicitly (confounding by indication, immortal time bias, selection into study, time-zero misalignment, collider bias…). Name the doctrine when flagging a problem — teachable output.

Step 2 — Statistical & spin review

• Statistical significance ≠ clinical significance: for each primary outcome, compare the effect size against the MCID of the instrument (search vault/textbook-md for the MCID of common ${persona.specialty} scales; cite the source). p<0.05 but < MCID → say so explicitly.
• RCT extras: ITT vs per-protocol (which was used, does it matter here), multiplicity correction for multiple outcomes/timepoints, fragility index when computable from a 2×2 (events small → quote it), sample size calculation vs actual enrollment.
• Trial registration check (RCTs; prospective SRs → PROSPERO): WebFetch the ClinicalTrials.gov / registry record and compare registered primary outcome vs published primary outcome. Outcome switching or unreported registered outcomes → 🔴 flag. No registration for a modern RCT → 🔴.
• Spin detection: put the abstract conclusion side-by-side with the actual results. Flag when the conclusion (a) claims benefit off a non-significant primary outcome via secondary/subgroup emphasis, (b) generalizes beyond the studied population, or (c) omits harms.
• Argument-structure check (does the conclusion follow from the paper's own premises?) — run it through the bundled argdown_lint.py so the verdict is a deterministic gate, not an impression:
  1. Read the paper's conclusion sentence → claim + claim_type (causal / hard_endpoint / consistency / general).
  2. List each primary/secondary finding meant to support it as a premise, TAGGING its type (direct_rct / association / surrogate_outcome / single_study / subgroup / secondary_outcome / mechanistic / expert_opinion).
  3. python argdown_lint.py argmap.json — it flags the illicit jumps deterministically (surrogate→hard-endpoint, association→causal, single-study→"consistently shows", subgroup/secondary-only→benefit). Exit 1 = a gap exists. The LLM does the semantic tagging; the script does the logic. Write the flagged gap(s) inline in this section; when the linter returns clean (exit 0), skip silently (don't pad every paper).
• Funding & COI: funding source, sponsor role in design/analysis, author COI vs the direction of conclusions.

Step 3 — Grade and summarize

• GRADE (per main outcome where applicable) — recompute the certainty with the bundled grade_judge.py, don't let a holistic LLM impression stand. The model rates the domains; the script computes the authoritative final grade (its value, not your gut-call label, goes in the note):
  1. Starting level: RCT-dominant evidence → high; observational-dominant → low.
  2. Rate each of the 5 domains — risk_of_bias / inconsistency / indirectness / imprecision / publication_bias — as not_serious (0) / serious (−1) / very_serious (−2), one-sentence justification each.
  3. Observational + no downgrade applied → add upgrades (large_effect RR>2 or <0.5 → +1, RR>5 or <0.2 → +2; dose_response; opposing_confounding). The script enforces the GRADE rule that upgrades apply only to observational evidence with no downgrade — don't pre-filter, just list them.
  4. Write the ratings to grade_input.json and run python grade_judge.py grade_input.json. The printed FINAL CERTAINTY is authoritative and goes into the 品質快照 + Critical Appraisal. If your holistic impression disagrees with the computed value, that's a signal to re-examine a domain rating, not to override the number. (Deterministic core ported from htlin222/robust-lit-review's grade_judge.py.)
• Limitations: authors' own AND unacknowledged ones.
• Fill the note's ## Critical Appraisal section (domain table) and the 品質快照 callout (see note template) — the snapshot is the 10-second verdict, the section is the evidence.

Phase 7: Citation & Impact Analysis

Evaluate how the paper has been received by the academic community and public:

Academic citations:

• scite first: mcp__zotero__scite_enrich_item (or scite_enrich_search) → supporting / mentioning / contrasting citation counts. A non-trivial contrasting count is the signal to dig — deep-read only the contrasting/critical citers instead of sampling blindly.
• get_paper_citations → list of citing papers
• Categorize the top 5-10 most relevant citing papers:
  • 🟢 Supporting: builds on or confirms findings
  • 🔴 Contradicting: presents opposing evidence
  • 🔵 Extending: takes the work in a new direction
  • 🟡 Methodological critique: questions the methods or analysis
• Note total citation count and trend (increasing? plateaued?)

Public/social discussion (part of the Phase-7 delegated agent, see §Delegation):

• WebSearch: "{title}" OR "{DOI}" site:pubpeer.com
• WebSearch: "{title}" OR "{DOI}" twitter OR reddit OR blog
• Look for news coverage, especially in medical/science news outlets
• Note any notable commentary, criticism, or endorsement
• WebSearch/WebFetch failure or zero hits → state "PubPeer/social sweep: no hits / FAILED ()" in the note — an unrun sweep must not read as "no criticism found"

Output: impact assessment relative to field norms and paper age.

Output: Obsidian Note

Write to ${vault.papers_dir}{FirstAuthorYear} - {Short Title}.md

If a note for this paper already exists (check by DOI or PMID), merge the deep review sections into the existing note rather than creating a duplicate. Show the user a diff of changes before writing (per CLAUDE.md rules). Brand-new note → show the full draft before writing (CLAUDE.md tiered approval — new notes are draft-first, same as merges are diff-first).

Note template — 三塊結構

筆記分為三塊邏輯：(1) 讀完能用 (2) 可不可信 (3) 連結與來源

Quick review（[x] checkbox 觸發）只寫前兩塊 + Resource。 Deep review（/paper-review 觸發）加入 # Deep Review 段落。

## 白話導讀 一律要寫（quick 與 deep 都寫）：放在 ## Summary 之後、## Clinical Application 之前， 用一個 > [!abstract] callout 把整篇用繁體中文白話「講一遍」，讓使用者不必讀完深度評讀就能掌握全篇在說什麼。 這是「導讀／意譯」不是逐句翻譯。外框固定三段（一句話 / 為什麼做 / 所以呢），中間「主體」依 study_type 分流：

• RCT：逐組講清楚——實驗組 vs 對照組各自的 protocol（具體介入、劑量、頻率、療程；對照是 sham 還是 standard care）， 且每個 outcome 要給兩組的實際數值與組間差異（mean±SD / Δ / 95% CI / p），不可只寫「有顯著差異」。
• SR/MA：逐 outcome 講——每個 outcome 的納入族群定義（PICO）+ pooled effect（SMD/RR/OR + 95% CI、I²、納入篇數/人數）。
• Narrative review：逐組別/主題講——他把主題分成哪幾組，每組整理了什麼重點（一組一行）。
• 觀察性研究（cohort/case-control/cross-sectional）：逐組講——各組的定義、各組的結果（關鍵數值）、 以及作者推測的原因/機轉（明確標示為 inference，不是因果證實）。

醫學名詞、藥名、參數、量表保留英文，關鍵數值用 ==highlight==，其餘用白話中文。 abstract-only 時就依摘要能取得的範圍寫，缺的數值標「abstract 未報告，需全文確認」，並沿用 ## Summary 的 ⚠️ 標記。

---
tags:
  - source/journal
  - med/{specialty}      ← 只用已存在的 med/ tag，新 tag 需更新 registry
citekey: "{BBT citation key}"
doi: "{DOI}"
zotero: "zotero://select/items/@{citekey}"
journal: "{Journal}"
year: {YYYY}
authors: "{All authors}"
study_type: "{RCT/SR+MA/cohort/etc.}"
pmid: "{PMID}"
review_depth: "quick"   ← 或 "deep"
---

> [!info] 品質快照
> **設計**：{RCT / cohort / SR+MA…}｜**n**：{樣本數}｜**RoB**（{RoB 2/ROBINS-I/AMSTAR-2…}）：{🟢低/🟡中/🔴高}
> **GRADE**：{High/Moderate/Low/Very Low}（主要 outcome）｜**Registration**：{✅一致 / 🔴 outcome switching / ❌未註冊 / n.a.}
> **一句話 verdict**：{可信度＋最關鍵的 1-2 個弱點，例：可信度中等——隨機化健全但 open-label、primary outcome 未達 MCID}

# Key Points

## Summary

> ⚠️ Based on abstract only  （或 > ✅ Full text reviewed）

- 一句話重點，==highlight== 最重要數值

## 白話導讀

> [!abstract] 整篇講給你聽（繁體中文白話）
> - **一句話**：這篇到底在說什麼
> - **為什麼做**：背景動機、想補的研究缺口
> - **主體**（依研究類型，見下方規則填對應結構）
> - **所以呢**：臨床/實務上的意義 + 最重要的 1-2 個限制

> [!example]- 主體結構範例（依 study_type 擇一，實際只留對應那種）
> **RCT** — 逐組講清楚 + outcome 實際展現：
> - 組別 protocol：實驗組（介入、劑量/頻率/療程）vs 對照組（sham/standard care 具體是什麼）
> - 每個 outcome：實驗組 vs 對照組 的==實際數值與組間差異==（不只寫「有顯著」；給 mean±SD / Δ / 95% CI / p）
>
> **SR/MA** — 逐 outcome 講：
> - 每個 outcome：納入的==族群定義==（誰、什麼介入、什麼比較）+ pooled effect（==SMD/RR/OR + 95% CI==、I² 異質性、納入幾篇/幾人）
>
> **Narrative review** — 逐組別/主題講：
> - 他把主題切成哪幾個組別/段落，==每一組整理了什麼重點==（一組一行）
>
> **觀察性研究（cohort / case-control / cross-sectional）** — 逐組講：
> - 各組的==定義==（暴露/非暴露、case/control 怎麼分）
> - 各組的結果（==關鍵數值==）
> - 作者推測的原因/機轉（標明這是 inference，非因果證實）

## Clinical Application

- **適應症：**（具體 condition、severity）
- **Protocol：**（full text 時補充：頻率、pulse 數、coil 類型、治療天數、時機）
- **禁忌/注意：**
- **實務 tips：**
- **在地適用性 / external validity：** 研究族群能否外推到 `${persona.locale}` 的
	`${persona.specialty}` 病人。**給付/自費查詢為 OPTIONAL**：只在 `${persona.reimbursement_system}`
	非空、且使用者問到給付或介入明顯涉及給付決策時才查（用你設定的給付資料庫工具）；否則略過此 bullet。

# Paper Details

## Methods

- 研究設計、收案條件、介入方式
- （full text 時補充：statistical methods、sample size calculation、具體 protocol）

## Critical Appraisal