🌐 Language / 語言：繁體中文 · English
📋 更新紀錄 / Changelog：docs/CHANGELOG.md — 最新：v0.14.0 · 2026-09-02（新增 3 支技能：投稿選刊、研究架構圖、空間資料）
-green){kind=icon}
一組給商管、財金、社會科學研究者的 Claude Skills：從選題、找資料、讀文獻，到分析、寫作、投稿與複製包封存——量化／質化／實驗／混合方法全典範，涵蓋研究全流程。站在多位開源作者的肩膀上（見致謝）。 A curated bundle of Claude Skills for research in business, finance, and the social sciences — from framing a question, finding data and reading the literature, through analysis, writing, submission, and replication-package archiving. Quantitative, qualitative, experimental, and mixed methods. Built on the shoulders of several open-source authors (see Credits).
一組可掛到 Claude 上的「技能（Skills）」。掛上後，當你在對話中談到相關任務，Claude 會自動載入對應技能、套用該領域的專業框架與紀律，不必每次重打長提示詞。
適合對象：商管／財金／社會科學的研究者（碩、博士生與研究人員）。原本依台灣學術脈絡設計（繁體中文、APA 7、口試與審查意見回覆格式），但多數技能在任何脈絡都適用。方法典範不限量化——質化、實驗、混合方法均有對應技能。
這個包在意什麼：不只把答案生出來，更在意答案站不站得住。數字要能指回出處、引用要能查證真偽、識別假設要先講清楚、不確定就標示不確定。技能寧可回報「查無」，也不生一個看起來合理的東西——因為看起來合理的錯誤最難被發現。
- 腳本會真的跑，不只是給指引。 資料與圖表類技能附實測過的 Python／R 腳本——你電腦裝了 Python 或 R，就能實際連線抓書目、清資料、建資料表、出圖（例：
literature-matrix-builder連 CrossRef 抓 DOI 直接建 Excel 比較矩陣；r-spss-syntax-architect產出可貼上就跑的 R／SPSS／Stata 語法；management-figure直出 300dpi 出版級圖）。 - 不只本國資料，也做跨國。 除了台灣公開資料，
global-opendata-scout內建 World Bank／Eurostat／ILOSTAT／IMF／UN Data 的免金鑰、已實測端點與撈取腳本，並主動提醒跨國比較的陷阱（國家代碼三套不可混、幣別／PPP／基期、產業分類不可直接對應），讓你做他國與跨國分析不被本國資料綁死。 - 反幻覺，答案要站得住。 數字能指回出處、引用能查真偽、CrossRef 查無就報錯不憑記憶補書目——寧可回報查無，也不生一個看似合理的錯誤。
- 一條龍、由「研究大腦」分派。 從方法選擇→找資料→清理→分析→寫作→投稿→複製包封存，
research-orchestrator會判斷你在哪一階段、該叫哪幾支、什麼順序，不必自己記 38 支各做什麼。 - 四大方法範式齊全，還能交複製包。 量化／質化／實驗／混合都有對應技能；
reproducibility-architect直接把研究打包成可重跑的複製包（環境鎖定、授權資料的可重現困境、資料／程式碼／AI 使用三聲明、DOI 封存），對接頂刊資料編輯要求。
如果這裡有哪支技能幫你擋掉一個幻覺引用、省下一下午的洗資料、或讓 reviewer 2 沒那麼可怕——按顆 ⭐ 就是學術版的「被引用」：下一個需要的研究者靠它找到這裡。不用送審、不用 revise & resubmit。就當它是開源研究工具版的「按讚、訂閱、開小鈴鐺」。 🔔
⚠️ 本專案與 Anthropic 無官方關聯。部分技能需搭配外部工具或**付費資料庫（如 TEJ）**才能發揮完整功能。
新增 3 支技能（35 → 38 支），都是先前已完成、這次一併釋出的。
-
journal-submission-scout｜投稿期刊選擇＋掠奪性期刊篩查。 稿子寫好了要投哪裡？從題目與摘要找出真正會刊登這類研究的候選期刊，比對客觀指標（h-index、2 年平均被引、APC、DOAJ 收錄、審查制度），並以 Think.Check.Submit 架構做掠奪性期刊篩查——投錯一本掠奪性期刊，那篇論文再也不能投別處，這是本技能最重要的價值。附免金鑰腳本（OpenAlex／DOAJ／Crossref）。- 紀律：絕不提供 JIF、接受率、ABS／FT50／SCImago 分級——那些是專有資料，公開 API 取不到，工具寧可說查無也不憑印象生成。OpenAlex 的「2 年平均被引」與 JIF 演算法不同，絕不混稱。
-
research-framework-figure｜研究架構圖產生器。 把假說畫成投稿與口試等級的架構圖：中介、調節、被調節的中介、序列中介，以及台灣商管論文最常用的「中介＋雙控制變數框」版式（每個依變數各一組控制變數，以虛線彎箭頭接入）。輸出可再編輯的 SVG 與 PPTX——口試前可在 PowerPoint 內直接改字。- 附三個可直接執行的匿名範例，涵蓋三種版式。
-
spatial-data-architect｜空間資料分析。 區域研究／hedonic 房價／不動產：地理編碼驗證、H3 六角網格聚合、TWD97↔WGS84 座標紀律（用錯座標系不會報錯，只會安靜地把台灣移到別的地方）、空間自相關診斷、出版級空間地圖。
本版起，切線改在「框架 vs 實測答案」，而不是「整支技能」：公開拿方法論（怎麼想），私人留實測換來的目錄與對照表（答案是什麼）。因此 tej-data-scout 的資料表索引與 tej-variable-mapper 的變數種子對照表不再隨公開包發布——但兩支技能的方法論完整保留，仍可獨立完成可行性判斷與變數對映。
本版沒有新增技能（維持 35 支），修的是一個不會報錯、只會安靜降級的問題——正是最難自己發現的那種。
-
路由總管指向不存在的技能（
research-orchestrator）。 名錄裡列了 5 支不在本公開包內的技能（check-citations、journal-submission-scout、research-framework-figure、spatial-data-architect、tw-opendata-scout）。Claude 依名錄去呼叫一支不存在的技能時不會拋錯，只會退回一般回答——使用者不會看到任何警示，只會覺得「怎麼跟說明寫的不一樣」。- 已移除這些條目與對應路由行，宣告的可路由數改為本包實際的 34 個（35 支減去 orchestrator 自身）。
- 同步在維護端加了一道自動閘門：往後只要名錄與實際內容不一致，打包流程就中止，不讓壞掉的版本成形。
-
指向已移除參考檔的孤兒引用（
global-opendata-scout）。 內文指向兩份未隨本包釋出的來源目錄，症狀相同：Claude 去讀一個不存在的檔案，然後無聲降級。已改寫。- 台灣在 UN Comtrade／WITS 被併入
490「Other Asia, nes」 的陷阱警告仍完整保留——「不知道就會踩」的警告屬公共利益，一律公開。
- 台灣在 UN Comtrade／WITS 被併入
-
內容同步與新增參考檔。 補上 2026-08-20 之後累積的修訂，涵蓋
academic-journal-polisher、q1-journal-polisher、phd-milestone-tracker、qualitative-thematic-coder、r-spss-syntax-architect、public-disclosure-scout、academic-slides、research-method-selector等。- 新增
tej-data-scout/references/tej-access-channels.md：TEJ 三條取得管道（Pro 桌面端／tejapi／TQuant-Lab）對照，含一個常見誤解——TEJ Pro 校園帳號不含 API 授權，兩套系統不通用——以及金鑰與帳號的安全紀律。
- 新增
-
新增
docs/CHANGELOG.md，往後的變更集中記錄於此。
本版新增兩道學術誠信防線——都是「不做會出事、做了沒人看得見」的那種功能。
-
撤稿查核，在建檔當下就攔截（
literature-matrix-builder）。 引用到已撤稿的論文是實質學術風險，但既有的引用工具只查「文獻存不存在」「格式對不對」，不查「這篇是不是已經被撤稿」。本版讓每次add自動反查 Crossref 的撤稿通知，在文獻進庫的當下就攔下來——而不是等寫進文獻回顧才發現。- 已用陽性／陰性對照實測：已知被撤論文測得出，三篇正常論文皆不誤報。只驗前者會做出一個到處亂報的工具。
- 查核失敗 ≠ 沒被撤稿：工具標「
⚠️ 查核失敗：原因」而非「未發現」。查不到與確認沒事是兩回事，這個區別攸關風險判斷。 - 實務細節：Elsevier 慣例是把撤稿通知放在同一個 DOI（原文被替換成撤稿聲明），若照直覺去找「另一個 DOI」會查不到。
-
雙盲投稿的身分資訊清除（
thesis-consistency-audit）。.docx／.pptx／.xlsx是壓縮檔，作者身分藏在五個肉眼看不到的地方，而 Word 內建的「檢查文件」不一定清得掉後兩項——這是 desk reject 的常見原因：docProps/core.xml建立者、最後修改者docProps/app.xmlCompany（學校）、Manager（指導教授）docProps/custom.xml自訂屬性（常含計畫編號）word/comments.xml每則註解都帶作者姓名- 追蹤修訂 每個
w:ins／w:del都帶w:author與時間戳
- 紀律：註解與追蹤修訂含有內容、不只是中繼資料，故預設只報告不刪除，要處理必須明確加旗標；清除前一律自動備份。追蹤修訂採「作者匿名化」而非刪除，避免改變文件內容。
- 誠實邊界：工具只處理檔案中繼資料。正文自我引用、致謝、基金計畫編號、檔名含姓名——這些才是最常見的雙盲破功點，必須人工檢查。
- v0.12.0 — 台灣代碼陷阱：第二例，並升級為通則（
global-opendata-scout）。 繼 World Bank／OECD 完全沒有台灣資料之後，本版實測確認第二個體系的處理方式不同但同樣會坑人：UN Comtrade／WITS 把台灣併入代碼490「Other Asia, not elsewhere specified」。實測結果——查490回傳 218 筆，查常被誤用的158回傳 0 筆卻不報錯。查錯代碼會靜默得到空結果，極易誤判成「這個資料源沒有台灣資料」而放棄整個來源。本版把它從個案寫成通則：遇到任何聯合國體系下的國際資料庫，預設假設「台灣不會用標準 ISO 碼出現」，先查該庫怎麼處理台灣再決定研究設計，並列出三種常見處理型態（完全沒有／併入其他代碼／用非標準名稱如 "Chinese Taipei"）。使用 490 時論文須揭露的三件事也一併寫明。
近期把重心放在資料涵蓋的廣度（跨國）、取得的穩健度與文件前處理——研究流程最前段、也最容易「髒進髒出」的幾塊。
- v0.11.0 — 跨國／國際比較資料（
global-opendata-scout）。 研究要做他國或跨國比較時，內建 World Bank／Eurostat／ILOSTAT／IMF／UN Data 的免金鑰、已實測端點與撈取腳本，另附「如何找到任一國家官方統計機構」的五步方法論。核心價值不在找到數字，而在主動點出跨國資料的可比性陷阱（國家代碼三套、幣別／PPP／基期、會計年度、產業分類 ISIC／NACE／NAICS 不可直接對應、涵蓋率遺漏造成選擇偏誤）。台灣資料不涵蓋（World Bank／OECD 皆無台灣），走官方來源。 - v0.10.0 — 複雜揭露文件前處理（
text-analytics-architect）。 語料若是版面複雜的揭露文件（10-K、年報、ESG 永續報告、掃描檔），pdftotext直抽會把表格壓平、多欄交錯、頁尾混進正文——髒進髒出，污染後面的斷詞與情緒分析。新增 Step 0 指引：先用版面感知抽取轉成保留語意結構的乾淨文字（並驗抽取品質、保 source map、遮罩 PII）再分析。融合 KingsleyOWO/Semark（Apache 2.0）的語意化文件處理概念，只取概念不取依賴。 - v0.9.0 — 動態網站抓取升級階梯（
public-disclosure-scout）。requests抓不到 JS 動態頁時的由輕到重升級階梯（先找背後 API → 官方批次 → 無頭瀏覽器）。融合 unclecode/crawl4ai（Apache 2.0）「把網頁轉成 LLM 可讀結構」概念，只取概念不取依賴，並明訂 🚫 禁用其 stealth 反偵測功能——學術研究抓不到就升級到官方管道／申請／人工，不是躲過偵測。 - v0.8.0 — 讀文獻線（
literature-matrix-builder+bilingual-paper-reader）。 把幾十篇 PDF 變成有結構的東西:文獻比較矩陣(PDF→DOI→CrossRef 免金鑰→APA 7→20 欄 Excel,CrossRef 查無 DOI 直接報錯不憑記憶補書目),與單篇論文雙欄精讀(逐段中譯+五色標記+離線閱讀器)。兩支天然接力,精讀結論回填矩陣綜整欄。
一貫紀律不變:融合外部開源都只取概念、不吞依賴、遇 copyleft 就規避、用了就在 NOTICE.md 與各技能 ATTRIBUTION.md 誠實致謝。
完整版本歷史見各 Releases。技能總數：35。 安裝:整個 repo 的 ZIP 無法直接當單一技能上傳;請到
dist/下載你要的個別.zip(見下方安裝說明)。
- 資料由你自己抓。 資料類技能一律假設你（或你的機構）擁有合法訂閱／授權帳號，由你自己登入下載。技能只教「在哪找、怎麼判斷、怎麼分析」，不代抓資料、不散布任何資料庫的專屬目錄。
- 教方法，重識別假設。 找到資料後，技能會建議最恰當的研究設計與估計方法（先講清楚識別假設會不會成立，再談跑哪個模型），並產出可重現的 R／SPSS／Stata 語法。
- 資料庫中立、歡迎擴充。 框架適用於任何資料庫（TEJ 只是內建範例）。歡迎社群依
docs/ADD_A_DATABASE.md新增 WRDS/Compustat、CSMAR、World Bank 等 profile。
| 研究階段 | 技能 | 一句話功能 | 觸發詞（示例） | 授權 |
|---|---|---|---|---|
| ① 選題・文獻・資料 | research-orchestrator |
研究大腦總管，替你分派合適的子技能 | 不知道從何著手、綜合任務 | 原創 |
research-method-selector |
方法論適配：判量化/質化/實驗/混合＋Q1 過程套模；含新手小白引導模式 | 該用什麼方法、不知道要研究什麼 | 原創 | |
phd-researcher |
文獻分析、方法論逆向、研究缺口、PRISMA、後設分析 | 文獻分析、系統性回顧、meta-analysis | 混合 🔒 | |
literature-matrix-builder |
文獻語料庫與比較矩陣：PDF→DOI→CrossRef→APA 7→Excel 橫向比較表 | 文獻矩陣、文獻整理、建文獻庫、APA7、DOI 查書目 | 原創 | |
bilingual-paper-reader |
單篇論文雙欄精讀：逐段中譯＋五色預先標記＋離線閱讀器（標記可持久化/匯出） | 精讀論文、論文翻譯、中英對照、畫重點做筆記 | 原創 | |
tej-data-scout |
資料可行性偵察＋研究設計/估計方法建議（以 TEJ 為範例，資料庫中立） | 選題、資料可行性、這題能不能做 | 原創 | |
public-disclosure-scout |
免費官方公開揭露偵察（MOPS 重大訊息/年報/裁罰）＋事件研究事件源整備 | 公開資訊觀測站、MOPS、免費資料、事件源 | 原創 | |
global-opendata-scout |
跨國/國際公開統計偵察：World Bank/Eurostat/ILOSTAT/IMF/UN Data 免金鑰端點＋撈取腳本＋可比性陷阱 | 跨國、國際比較、World Bank、他國統計、cross-country | 原創 | |
multi-source-data-integrator |
多源嚴謹結合：實體解析、跨源值調解、來源譜系、三角驗證、合併損耗對帳 | 多源結合、跨源、統編對接、三角驗證 | 原創 | |
tej-variable-mapper |
把 Compustat／CRSP 變數定義映射到 TEJ 對應欄位 | 變數對應、TEJ 欄位、Compustat | 原創 | |
tej-data-wrangler |
TEJ 原始 Excel/CSV 清理、遺漏值分析、格式標準化 | 資料清理、遺漏值、格式標準化 | 原創 | |
| ② 分析・語法・量表 | survey-research-architect |
問卷研究全流程：設計、先驗檢定力抽樣、發放回收、CMV 攻防 | 問卷設計、樣本數、CMV | 原創 |
interview-method-designer |
深度訪談設計：三層大綱、理論抽樣與飽和、倫理知情同意 | 訪談大綱、訪幾個人、飽和 | 原創 | |
experiment-design-architect |
實驗設計：組間/組內、counterbalancing、情境實驗、操弄檢核 | 實驗設計、vignette、操弄檢核 | 原創 | |
r-spss-syntax-architect |
依假說生成可重現的 R／SPSS／Stata 語法（含 SEM/PLS 軌） | R 語法、SPSS 語法、PLS-SEM | 原創 | |
causal-inference-architect |
因果識別策略：現代交錯 DiD、IV、RDD、合成控制、事件研究圖 | DiD、內生性、識別策略 | 原創 | |
text-analytics-architect |
文字資料變研究變數：主題模型、情緒語調、LLM 標註信效度 | 文字探勘、LDA、LLM 標註 | 原創 | |
ob-hrm-scale-adaptor |
量表跨文化改編（合規版）＋測量恆等性檢定語法 | 量表改編、測量恆等性、lavaan | 原創 | |
qualitative-thematic-coder |
Braun & Clarke 主題分析，深度訪談逐字稿編碼 | 主題分析、逐字稿編碼、質性 | 原創 | |
management-figure |
出版級統計圖：倒 U 轉折點、係數森林圖、交互作用、邊際效果 | 出版級圖、forest plot、交互作用圖 | 📎 MIT | |
| ③ 寫作・潤飾 | academic-journal-polisher |
台灣學術環境文句潤飾，杜絕 AI 慣用語 | 潤飾、去 AI 味、學術文句 | 原創 |
q1-journal-polisher |
Q1–Q4 國際期刊英文潤飾＋APA 7＋模擬審查 | 期刊潤飾、投稿前、APA 7 | 原創 | |
nstc-grant-writer |
國科會計畫申請書寫作：結構化撰寫＋審查人視角自評 | 國科會、NSTC、計畫書、預期成果 | 原創 | |
response-letter-craftsman |
投稿修訂（R&R）逐點回覆信＋Response to Reviewers | 審查回覆、response letter、R&R、逐點回覆 | 原創 | |
| ④ 投稿前品管 | thesis-consistency-audit |
六維度一致性稽核：假設↔迴歸表、樣本數、文字↔表格、引用 | 一致性稽核、論文對帳、投稿前檢查 | 🔒 CC BY-NC-SA |
reproducibility-architect |
複製包架構：可重現專案結構、環境鎖定、授權資料可重現困境、資料/程式碼/AI 使用聲明、DOI 封存 | 複製包、可重現、資料可用性聲明、AI 使用揭露 | 原創 | |
q1-journal-reviewer |
模擬 ABS 3*/4* 匿名審查委員的批判 | 審稿、peer review、審查意見 | 原創 | |
citation-verifier |
揪出 AI 捏造假文獻、孤兒引用、引用不貼合主張 | 引用查核、假文獻、孤兒引用 | 原創 | |
| ⑤ 口試・簡報 | academic-pptx |
學術簡報內容與結構標準：行動式標題、論證式編排 | 口試簡報、conference talk | 📎 MIT |
academic-slides |
Beamer 風格、零依賴單檔 HTML 學術簡報（含 KaTeX） | 學術投影片、Beamer、HTML 簡報 | 📎 MIT | |
academic-deck-animator |
簡報動畫引擎：HTML+Canvas 粒子／原生 PPTX 進場動畫、逐步揭示 | PPT動畫、粒子特效、動態簡報、進場動畫 | 原創 | |
academic-poster |
研討會學術海報：A0/A1、傳統三欄或 Better Poster 版式 | 海報、poster、A0、研討會海報 | 原創 | |
defense-qa-coach |
口試答辯教練：委員提問題庫＋擬答框架＋模擬攻防 | 口試、答辯、模擬口試、追問 | 原創 | |
| ⑥ 博士修業（模板） | phd-milestone-tracker |
博士修業里程碑與 deadline 追蹤（規則為範例，請換成你系所的） | 修業里程碑、資格考期限、畢業時程 | 原創 |
qual-exam-coach |
學科資格考備考教練（可依你的考科調整） | 資格考、備考、記憶卡、模擬考 | 原創 |
授權圖例：原創＝作者本人著作（採根 LICENSE）；📎＝改作／收錄他人開源作品（保留原授權與姓名標示）；🔒＝含非商業（NC）條款。詳見 NOTICE.md。
關於引用查核：另有優秀開源工具
PHY041/claude-skill-citation-checker（比對 CrossRef／Semantic Scholar／OpenAlex）。因其上游未附授權（預設保留一切權利），本 repo 不重製其程式碼，建議直接前往取用，與本包citation-verifier搭配。
⚠️ 不能把整個 repo 的 ZIP(Code → Download ZIP)當成一個技能上傳——那裡面有 35 個SKILL.md、深層巢狀,Claude 一次只吃一個技能。也不能只給網址讓 claude.ai 自己抓;技能上傳是「上傳檔案」。請用個別的.zip。
claude.ai(逐支安裝)
- 到
dist/點你要的技能.zip(例literature-matrix-builder.zip)→ 右側 Download。 - claude.ai → 頭像 → Settings → Capabilities,開啟 Code execution(含腳本的技能需要)。
- Settings → Skills → Add / Upload → 選剛下載的
.zip→ 完成。要幾支重複幾次(平台一次一支)。上傳後跨裝置自動同步。
Claude Code:clone 後把 skills/<名稱>/ 資料夾複製到 ~/.claude/skills/(不需打包)。詳見 dist/README.md 與 docs/INSTALL.md。
其他 AI 代理(Codex 等):技能的「內容」(SKILL.md 指引+scripts+references)是可攜的 markdown 與 Python;技能的「自動載入」機制是 Claude 專屬。在 Codex/其他代理要用,見 docs/USE_WITH_OTHER_AGENTS.md。
範例 1 — 發想期問「這題能不能做」＋方法建議
你：我想做「董事會性別多元化對創新產出（專利數）的影響」，TEJ 做得起來嗎？ Claude（
tej-data-scout）：拆構念、給「變數 × 資料對照表」，標示董監資料【直接】、專利數【外部】（建議 TIPO）；判定為公司-年 panel，因專利為計數且過度離散 → 建議負二項固定效果；再交棒/r-spss-syntax-architect產語法。
範例 2 — 投稿前一致性稽核
你：這是我的迴歸章節（.docx），投稿前幫我抓內部矛盾。 Claude（
thesis-consistency-audit）：逐一比對假設↔迴歸表、樣本數、文字↔表格數字、APA 引用，列出會被審查委員圈起來的矛盾點與修正建議。
範例 3 — 模擬 Q1 期刊審查
你：用 ABS 3* 審查委員的角度狠一點審這篇 introduction 與 hypotheses。 Claude（
q1-journal-reviewer）：以匿名審查口吻指出理論貢獻、假設推導斷點、identification 威脅，並給 Major/Minor Revision 的具體要求。
本 repo 是合輯（collection），採逐資料夾授權：原創技能採根 LICENSE；改作／收錄他人作品的技能，各資料夾內保留原作者 LICENSE 與 ATTRIBUTION.md。完整對照見 NOTICE.md。🔒 標示 Non-Commercial 的技能僅供非商業學術用途。
- Zara Zhang —
academic-slides（MIT） - Yuan Yizhe（Yuan1z0825） —
nature-skills，management-figure的出版級繪圖底子（MIT） - Cheng-I Wu（Imbad0202） —
academic-research-skills，phd-researcher的系統性回顧／後設分析模組（CC BY-NC 4.0） - qqfly1to19 —
awesome_proofreading_auto，thesis-consistency-audit的稽核框架理念（CC BY-NC-SA 4.0） - PHY041 —
claude-skill-citation-checker，推薦搭配的引用查核工具（本 repo 未重製其程式碼） academic-pptx內容準則參考 Barbara Minto《Pyramid Principle》、Naegle (2021) PLOS Comput Biol 等公開學術實務。
若你是上述任一作者、對收錄或標示有任何意見，歡迎開 issue，我會立即配合調整。
歡迎 issue 與 PR，特別是新增資料庫 profile（見 docs/ADD_A_DATABASE.md）。請先讀 CONTRIBUTING.md。本專案與 Anthropic 無官方關聯；技能輸出僅供研究輔助，最終學術判斷與責任在使用者本人。phd-milestone-tracker／qual-exam-coach 內的規則為範例模板，務必以你所屬系所公告為準。部分技能需付費資料庫（TEJ）。
A bundle of Claude Skills covering the full research workflow. Once installed, Claude auto-loads the relevant skill when your conversation touches a matching task, applying that field's framework and discipline — no need to re-type long prompts.
Audience: researchers in business, finance, and the social sciences (master's/PhD students and faculty). Built around the Taiwanese academic context (Traditional Chinese, APA 7, oral-defense and reviewer-response conventions), though most skills work anywhere. Not limited to quantitative work — qualitative, experimental, and mixed-methods skills are included.
What this bundle optimizes for: not just producing an answer, but whether the answer holds up. Numbers should trace back to a source; citations should be checkable; identification assumptions get stated before models get run; uncertainty gets labelled as uncertainty. These skills would rather report "not found" than return something that merely looks plausible — because a plausible-looking error is the hardest kind to catch.
- The scripts actually run — not just guidance. Data and figure skills ship tested Python/R scripts: if your machine has Python or R, they genuinely fetch bibliographic data, clean data, build tables and render figures (e.g.
literature-matrix-builderpulls DOIs from CrossRef straight into an Excel comparison matrix;r-spss-syntax-architectemits paste-and-run R/SPSS/Stata syntax;management-figureoutputs 300 dpi publication-grade figures). - Not just domestic data — cross-country too. Beyond Taiwan's public data,
global-opendata-scoutships key-free, tested endpoints and fetch scripts for World Bank / Eurostat / ILOSTAT / IMF / UN Data, and proactively flags cross-country comparability traps (three country-code schemes that mustn't be mixed, currency/PPP/base-year, non-mappable industry classifications) — so international and cross-country analysis isn't tied to home-country data. - Anti-hallucination — answers that hold up. Numbers trace back to a source; citations are checkable; a CrossRef miss fails loudly instead of filling a bibliography from memory — it would rather report "not found" than return a plausible-looking error.
- End-to-end, dispatched by a "research brain." From method selection → finding data → cleaning → analysis → writing → submission → replication-package archiving,
research-orchestratorworks out which stage you're in and which skills to call in what order — you don't have to memorize what all 38 do. - All four method paradigms, plus a replication package. Quantitative, qualitative, experimental and mixed methods are all covered;
reproducibility-architectpackages the study into a re-runnable replication bundle (environment locking, restricted-data reproducibility, data/code/AI-use statements, DOI archiving) that meets top-journal data-editor requirements.
If a skill here caught a hallucinated citation, saved you an afternoon of data-wrangling, or made Reviewer 2 a little less terrifying — a ⭐ is the academic equivalent of a citation: it's how the next researcher finds this. No peer review, no revise-and-resubmit required. Think of it as the "like, subscribe, and hit the bell" of open-source research tools. 🔔
⚠️ Not affiliated with Anthropic. Some skills require external tools or a paid database (e.g., TEJ) for full functionality.
Adds 3 skills (35 → 38), all previously completed and released together here.
-
journal-submission-scout— target-journal selection and predatory screening. Your draft is ready; where do you send it? Finds journals that actually publish this kind of work, compares objective indicators (h-index, 2-year mean citations, APC, DOAJ listing, review model), and screens for predatory venues using the Think.Check.Submit framework — submit to a predatory journal and that paper can never go anywhere else, which is this skill's most important job. Ships with key-free scripts (OpenAlex / DOAJ / Crossref).- Discipline: never reports JIF, acceptance rates, or ABS / FT50 / SCImago tiers. Those are proprietary and unavailable through public APIs; the tool says "not found" rather than inventing a plausible answer. OpenAlex's 2-year mean citation count is not JIF and is never presented as such.
-
research-framework-figure— conceptual model diagrams. Turns hypotheses into submission- and defense-grade framework figures: mediation, moderation, moderated mediation, serial mediation, and the dual-control-box layout common in Taiwanese business scholarship (one control set per dependent variable, joined by dashed curved arrows). Outputs editable SVG and PPTX, so you can fix wording in PowerPoint the night before a defense.- Includes three runnable anonymised examples covering three layouts.
-
spatial-data-architect— spatial analysis. For regional studies, hedonic pricing, and real estate: geocoding validation, H3 hexagonal aggregation, TWD97 ↔ WGS84 coordinate discipline (mixing them raises no error — it just quietly relocates Taiwan), spatial autocorrelation diagnostics, and publication-grade maps.
From this release the boundary is framework vs. measured answers, not whole skills: the public bundle carries the methodology (how to reason about a problem), while catalogues and lookup tables that only exist because someone measured them stay private. Accordingly, the TEJ table index and the variable seed-mapping table are no longer shipped — but both skills keep their full methodology and remain independently usable for feasibility assessment and variable mapping.
No new skills in this release (still 35). What it fixes is a failure that never raises an error and degrades silently — the hardest kind to notice on your own.
-
The router pointed at skills that aren't here (
research-orchestrator). Its catalogue listed 5 skills not included in this public bundle (check-citations,journal-submission-scout,research-framework-figure,spatial-data-architect,tw-opendata-scout). When Claude follows the catalogue and calls a skill that does not exist, nothing throws — it just falls back to a generic answer. You get no warning, only a vague sense that the behaviour doesn't match the documentation.- Those entries and their routing lines are removed; the declared routable count is now the bundle's actual 34 (35 skills minus the orchestrator itself).
- An automated gate was added on the maintenance side: if the catalogue and the actual contents ever diverge again, the build aborts rather than shipping a broken bundle.
-
Orphaned pointers to removed reference files (
global-opendata-scout). The text referenced two source catalogues not shipped with this bundle — same symptom: Claude reads a file that isn't there and quietly degrades. Rewritten.- The warning that Taiwan is folded into
490"Other Asia, nes" in UN Comtrade / WITS is fully retained. Traps you cannot avoid without knowing about them are a public good and stay public.
- The warning that Taiwan is folded into
-
Content sync and a new reference. Picks up revisions accumulated since 2026-08-20 across
academic-journal-polisher,q1-journal-polisher,phd-milestone-tracker,qualitative-thematic-coder,r-spss-syntax-architect,public-disclosure-scout,academic-slides, andresearch-method-selector.- Adds
tej-data-scout/references/tej-access-channels.md: a comparison of the three ways to obtain TEJ data (Pro desktop /tejapi/ TQuant-Lab), including a common misconception — a TEJ Pro campus account does not include API access; they are separate systems — plus key-handling discipline.
- Adds
-
Adds
docs/CHANGELOG.mdas the single place for release notes from now on.
This release adds two research-integrity guardrails — the kind of feature nobody notices until its absence costs you something.
-
Retraction checking at ingestion time (
literature-matrix-builder). Citing a retracted paper is a real academic risk, yet existing citation tools check whether a reference exists and whether its format is right — not whether it has been retracted. Everyaddnow queries Crossref for retraction notices and flags the paper as it enters your library, rather than after it has already reached your literature review.- Verified with both positive and negative controls: a known-retracted paper is detected; three ordinary papers produce no false alarm. Testing only the positive case would have shipped a tool that cries wolf.
- A failed check is not a clean bill of health: the tool reports "
⚠️ check failed: reason" rather than "none found." Not knowing and knowing-it's-fine are different states, and the difference matters for risk. - Practical detail: Elsevier's convention is to place the retraction notice at the same DOI (the original article is replaced by the notice), so looking for "a different DOI" finds nothing.
-
Blinding your submission files (
thesis-consistency-audit)..docx／.pptx／.xlsxfiles are archives, and author identity hides in five places you cannot see — and Word's own "Inspect Document" does not reliably clear the last two. This is a common cause of desk rejection:docProps/core.xml— creator, last-modified-bydocProps/app.xml— Company (your university), Manager (your supervisor)docProps/custom.xml— custom properties, often a grant numberword/comments.xml— every comment carries its author's name- Tracked changes — every
w:ins／w:delcarriesw:authorand a timestamp
- Discipline: comments and tracked changes contain content, not just metadata, so the tool reports without deleting by default; removal requires an explicit flag, and a backup is always made first. Tracked-change authors are anonymised rather than stripped, so the document itself is unchanged.
- Honest boundary: the tool handles file metadata only. Self-citation in the body text, acknowledgements, grant numbers, and your name in the filename are the far more common ways blinding fails — those still need a human pass.
- v0.12.0 — The Taiwan country-code trap: a second case, generalised into a rule (
global-opendata-scout). After confirming that World Bank and OECD contain no Taiwan data at all, this release verifies a second international system that handles Taiwan differently — and just as treacherously: UN Comtrade / WITS folds Taiwan into code490, "Other Asia, not elsewhere specified." Verified live: querying490returns 218 records; querying158(a code often mistaken for Taiwan) returns 0 records without raising an error. A wrong code yields a silent empty result, which is very easily misread as "this source has no Taiwan data" — leading researchers to abandon an otherwise usable dataset. This release turns the observation into a general rule: for any UN-system international database, assume by default that Taiwan will not appear under a standard ISO code; check how that particular database handles Taiwan before designing around it. Three common patterns are documented (absent entirely / folded into another code / listed under a non-standard name such as "Chinese Taipei"), along with the three things a paper must disclose when using code 490.
Recent work focused on data coverage (cross-country), data-acquisition robustness, and document preprocessing — the earliest stages of research, and the ones most prone to "garbage in, garbage out."
- v0.11.0 — cross-country / international-comparison data (
global-opendata-scout). For studies that need other-country or cross-country comparisons, it ships key-free, tested endpoints and fetch scripts for World Bank / Eurostat / ILOSTAT / IMF / UN Data, plus a five-step method for locating any country's official statistics office. Its core value isn't finding numbers but proactively flagging cross-country comparability traps (three country-code schemes, currency/PPP/base-year, fiscal-year differences, non-mappable industry classifications ISIC/NACE/NAICS, coverage gaps causing selection bias). Taiwan isn't covered (neither World Bank nor OECD includes it) — use official domestic sources. - v0.10.0 — complex-disclosure preprocessing (
text-analytics-architect). When the corpus is layout-heavy disclosure (10-Ks, annual reports, ESG/sustainability reports, scans), naivepdftotextflattens tables, scrambles columns and mixes footers into the body — garbage in, garbage out, poisoning downstream tokenization and sentiment. A new Step 0 says: run layout-aware extraction into clean, structure-preserving text first (verify extraction quality, keep a source map, mask PII), then analyze. Adopts the concept from KingsleyOWO/Semark (Apache 2.0) — concept only, no dependency. - v0.9.0 — dynamic-scraping escalation ladder (
public-disclosure-scout). Whenrequestscan't reach a JS-rendered page, a lightest-to-heaviest escalation ladder (find the underlying API first → official batch files → headless browser only if needed). Adopts the "turn web pages into LLM-readable structure" concept from unclecode/crawl4ai (Apache 2.0) — concept only, no dependency — and explicitly forbids its stealth / anti-detection features: scholarly scraping that can't reach a page escalates to official channels / a request / manual work, not to evading detection. - v0.8.0 — the literature-reading line (
literature-matrix-builder+bilingual-paper-reader). Turn dozens of PDFs into something structured: a comparison matrix (PDF→DOI→CrossRef, no key→APA 7→20-column Excel; CrossRef miss fails loudly, never fills bibliography from memory) and side-by-side close reading of a single paper (paragraph translation + five-colour marking + offline reader). The two hand off naturally.
The standing discipline holds: external open source is adopted as concept only, never as a bundled dependency; copyleft is avoided; and every borrowing is credited honestly in NOTICE.md and each skill's ATTRIBUTION.md.
Full version history in Releases. Total skills: 35. Install: the whole-repo ZIP is not a single installable skill — grab the individual
.zipyou want fromdist/(see Install below).
- You fetch your own data. Every data skill assumes you (or your institution) hold a legitimate subscription/license and download the data yourself. Skills only teach where to look, how to judge feasibility, and how to analyze — they never fetch data for you and never redistribute any database's proprietary catalog.
- Teach method, foreground identification. After data is located, skills recommend the appropriate research design and estimator (is the identifying assumption plausible first, model choice second) and generate reproducible R/SPSS/Stata syntax.
- Database-agnostic and extensible. The framework applies to any database (TEJ is just the built-in example). Contribute profiles for WRDS/Compustat, CSMAR, World Bank, etc. via
docs/ADD_A_DATABASE.md.
| Stage | Skill | What it does | License |
|---|---|---|---|
| ① Ideation・Literature・Data | research-orchestrator |
A router "research brain" that dispatches the right sub-skill | Original |
research-method-selector |
Methodological-fit advisor (quant/qual/experiment/mixed) + Q1 process templates + beginner guidance mode | Original | |
phd-researcher |
Literature analysis, methodology reverse-engineering, research gaps, PRISMA systematic reviews, meta-analysis | Mixed 🔒 | |
literature-matrix-builder |
Literature corpus + comparison matrix: PDF→DOI→CrossRef→APA 7→Excel synthesis table | Original | |
bilingual-paper-reader |
Side-by-side close reading of one paper: paragraph-aligned translation + five-colour pre-marking + offline reader (persistent highlights, note export) | Original | |
tej-data-scout |
Data-feasibility scouting + research-design/estimator advice (TEJ as example; database-agnostic) | Original | |
public-disclosure-scout |
Free official public-disclosure scouting (MOPS filings/annual reports/sanctions) + event-study event source | Original | |
global-opendata-scout |
Cross-country/international open-stats scout: key-free World Bank/Eurostat/ILOSTAT/IMF/UN Data endpoints + fetch scripts + comparability traps | Original | |
multi-source-data-integrator |
Rigorous multi-source integration: entity resolution, cross-source reconciliation, data lineage, triangulation, merge-loss accounting | Original | |
tej-variable-mapper |
Maps Compustat/CRSP variable definitions to TEJ fields | Original | |
tej-data-wrangler |
Cleans raw TEJ Excel/CSV: missing values, outliers, formatting | Original | |
| ② Analysis・Syntax・Scales | survey-research-architect |
End-to-end survey research: design, a-priori power & sampling, fielding plan, CMV defenses | Original |
interview-method-designer |
In-depth interview design: layered protocols, theoretical sampling & saturation, ethics | Original | |
experiment-design-architect |
Experimental design: between/within, counterbalancing, vignettes, manipulation checks | Original | |
r-spss-syntax-architect |
Generates reproducible R/SPSS/Stata syntax from hypotheses (incl. SEM/PLS lane) | Original | |
causal-inference-architect |
Causal identification: modern staggered DiD, IV, RDD, synthetic control, event-study plots | Original | |
text-analytics-architect |
Text-as-data: topic models, tone/sentiment, LLM-annotation validity discipline | Original | |
reproducibility-architect |
Replication packages: reproducible project structure, environment locking, restricted-data reproducibility, data/code/AI-use statements, DOI archiving | Original | |
ob-hrm-scale-adaptor |
Cross-cultural scale adaptation (copyright-compliant) + measurement-invariance syntax | Original | |
qualitative-thematic-coder |
Braun & Clarke thematic analysis for interview transcripts | Original | |
management-figure |
Publication-grade figures: inverted-U turning points, coefficient forest plots, interaction & marginal-effects plots | 📎 MIT | |
| ③ Writing・Polishing | academic-journal-polisher |
Prose polishing for Taiwanese academic writing; removes AI-tells | Original |
q1-journal-polisher |
Q1–Q4 journal English polishing + APA 7 + mock review | Original | |
nstc-grant-writer |
NSTC (Taiwan) grant-proposal writing + reviewer-lens self-assessment | Original | |
response-letter-craftsman |
Point-by-point R&R response letters + Response to Reviewers | Original | |
| ④ Pre-submission QA | thesis-consistency-audit |
Six-dimension consistency audit: hypotheses↔tables, sample sizes, text↔tables, citations | 🔒 CC BY-NC-SA |
q1-journal-reviewer |
Simulates an ABS 3*/4* anonymous reviewer's critique | Original | |
citation-verifier |
Catches AI-fabricated references, orphan citations, claim–citation mismatches | Original | |
| ⑤ Defense・Slides | academic-pptx |
Academic slide content & structure standards: action titles, argument-driven decks | 📎 MIT |
academic-slides |
Beamer-style, zero-dependency single-file HTML academic slides (KaTeX) | 📎 MIT | |
academic-deck-animator |
Presentation animation engine: HTML+Canvas particles / native-PPTX entrance animations | Original | |
academic-poster |
Conference posters: A0/A1, classic 3-column or Better-Poster layouts | Original | |
defense-qa-coach |
Defense Q&A coach: committee question bank + answer frameworks + mock drills | Original | |
| ⑥ PhD milestones (templates) | phd-milestone-tracker |
PhD milestone & deadline tracker (rules are a template — replace with your program's) | Original |
qual-exam-coach |
Qualifying-exam prep coach (adaptable to your subjects) | Original |
Legend: Original = authored by this project (root LICENSE); 📎 = adapted from / includes third-party open-source work (original license & attribution preserved in-folder); 🔒 = includes a Non-Commercial (NC) clause. See NOTICE.md.
On citation checking: there's an excellent open-source tool,
PHY041/claude-skill-citation-checker(checks CrossRef/Semantic Scholar/OpenAlex). Because its upstream ships no license (all rights reserved), this repo does not reproduce its code — please get it from the original repo and pair it with this bundle'scitation-verifier.
⚠️ The whole-repo ZIP (Code → Download ZIP) is not a single installable skill — it holds 35 nestedSKILL.mdfiles, and Claude ingests one skill at a time. You also can't just give claude.ai a URL; skill upload is a file upload. Use the individual.zipfiles.
claude.ai (one skill at a time)
- In
dist/, click the skill.zipyou want (e.g.literature-matrix-builder.zip) → Download on the right. - claude.ai → avatar → Settings → Capabilities, enable Code execution (needed for skills with scripts).
- Settings → Skills → Add / Upload → pick the downloaded
.zip→ done. Repeat per skill (the platform takes one at a time). Uploads sync across devices.
Claude Code: clone, then copy the skills/<name>/ folder into ~/.claude/skills/ (no packaging needed). See dist/README.md and docs/INSTALL.md.
Other AI agents (Codex, etc.): a skill's content (SKILL.md guidance + scripts + references) is portable Markdown and Python; the auto-loading mechanism is Claude-specific. To use these with Codex or another agent, see docs/USE_WITH_OTHER_AGENTS.md.
1 — "Is this topic feasible?" + method advice
You: I want to study board gender diversity → innovation output (patent counts). Can TEJ support this? Claude (
tej-data-scout): decomposes the constructs, gives a variable×data table (board data = direct; patent counts = external, suggest TIPO), identifies it as a firm-year panel, and — since patents are over-dispersed counts — recommends a negative-binomial fixed-effects model, then hands off to/r-spss-syntax-architectfor runnable syntax.
2 — Pre-submission consistency audit
You: Here's my results chapter (.docx). Find internal contradictions before I submit. Claude (
thesis-consistency-audit): cross-checks hypotheses↔tables, sample-size flow, text↔table numbers, and APA citations; lists the contradictions a reviewer would circle, with fixes.
3 — Mock Q1 journal review
You: Review this introduction and hypotheses as a tough ABS 3* reviewer. Claude (
q1-journal-reviewer): in an anonymous-reviewer voice, flags gaps in theoretical contribution, breaks in hypothesis logic, and identification threats, with Major/Minor-Revision-level requests.
This repo is a collection under per-directory licensing: original skills follow the root LICENSE; adapted/included third-party skills keep their own LICENSE and ATTRIBUTION.md in-folder. Full map in NOTICE.md. Skills marked 🔒 Non-Commercial are for non-commercial academic use only.
Deep thanks to the open-source authors this work stands on:
- Zara Zhang —
academic-slides(MIT) - Yuan Yizhe (Yuan1z0825) —
nature-skills, the publication-grade plotting foundation ofmanagement-figure(MIT) - Cheng-I Wu (Imbad0202) —
academic-research-skills, the systematic-review / meta-analysis module ofphd-researcher(CC BY-NC 4.0) - qqfly1to19 —
awesome_proofreading_auto, the audit-framework concept behindthesis-consistency-audit(CC BY-NC-SA 4.0) - PHY041 —
claude-skill-citation-checker, a recommended companion tool (its code is not reproduced here) academic-pptx's content guidelines draw on Barbara Minto's Pyramid Principle and Naegle (2021, PLOS Comput Biol).
If you are any of these authors and have any concern about inclusion or attribution, please open an issue — I will adjust immediately.
Issues and PRs welcome — especially new database profiles (docs/ADD_A_DATABASE.md). Please read CONTRIBUTING.md. Not affiliated with Anthropic; skill outputs are research aids — final academic judgment and responsibility rest with the user. The phd-milestone-tracker / qual-exam-coach rules are templates — verify against your own institution. Some skills require a paid database (TEJ).