slide-verify

醫學簡報「驗證 → 勘誤 → 重排 → 拆頁 → 白話改寫」一條龍 Claude Skill

A one-stop Claude Skill for medical slide decks: verify → correct → reorder → split → plain-language rewrite

---

📖 專案簡介 (About The Project)

slide-verify 是一個 Claude Skill：把醫學簡報的文字內容丟給它，它會用 PubMed 等資料庫逐條查證每個引用文獻和數字「是不是真的、有沒有引錯」，然後幫你修正錯誤、把投影片排成更順的順序、把塞太滿的頁面拆開，最後還能把艱澀的句子改寫成白話。

slide-verify is a Claude Skill: feed it the text content of a medical slide deck, and it checks every citation and number against PubMed and other databases to see whether they are real and correctly used. It then fixes the errors, reorders the slides into a smoother flow, splits overloaded pages, and can finally rewrite dense sentences into plain language.

📌 注意：輸入的不是 .pptx 檔案本身，而是從簡報萃取出來的文字（例如 Markdown 檔，或直接複製的投影片文字）。如果你手上還是 .pptx，請先用我們的姊妹專案 PPT-to-MD 把簡報忠實轉成 Markdown，再把結果交給 slide-verify。

📌 Note: the input is not the .pptx file itself — it is the text extracted from the deck (e.g. a Markdown file, or slide text copied directly). If all you have is a .pptx, first convert it faithfully to Markdown with our sister project PPT-to-MD, then hand the result to slide-verify.

💡 為什麼要用這個專案？ (Why Choose This?)

市面上其實已經有不少開源的「文獻查核」工具，它們做的事情大同小異：拿你的參考文獻清單，去 CrossRef、Semantic Scholar 等學術資料庫比對，回答**「這篇文獻存在嗎？是不是 AI 捏造的？」**

There are already quite a few open-source citation-checking tools out there, and they all do roughly the same thing: take your reference list, match it against academic databases like CrossRef and Semantic Scholar, and answer "does this paper exist, or did an AI make it up?"

但真實簡報裡最危險的錯，往往不是假文獻，而是真文獻被用錯了——文獻確實存在、連結也是真的，所以只查「存在不存在」的工具會直接放行。這正是 slide-verify 與眾不同的地方：

But the most dangerous errors in real decks are usually not fake papers — they are real papers used wrongly. The paper exists and the link works, so any tool that only checks existence waves the error straight through. This is exactly where slide-verify stands apart:

比較項目 What we compare	一般文獻查核工具 Typical citation checkers	slide-verify
文獻存在嗎？ Does the paper exist?	✅	✅
這篇真文獻真的說過這個數字嗎？ Did this real paper actually say this number?	❌ 只比對書目 metadata only	✅ 追進原文逐條核對 goes into the paper itself
產出是什麼？ What do you get back?	一張錯誤清單 an error list	一份修好、排順、拆頁、可白話化的完整簡報 a fully corrected, reordered, split, optionally plain-language deck
醫學領域專用？ Medical-domain aware?	❌ 泛學術 general academic	✅ PubMed 優先、懂灰色文獻（指引／海報／法規）PubMed-first, handles grey literature (guidelines / posters / regulatory docs)
需要安裝程式環境？ Needs a code environment?	多數要 Python／API key（usually yes）	❌ 純 Claude Skill，clone 進資料夾就能用 pure Claude Skill — just clone into a folder

具體來說，slide-verify 專抓這五種「文獻是真的、說法是錯的」，全部來自真實簡報的實跑校準：

Concretely, slide-verify hunts these five kinds of "real paper, wrong claim", all calibrated from real-world deck audits:

錯誤型態 Error type	白話說就是 In plain terms
🎭 張冠李戴 Wrong drug	拿別人（競品）試驗的成績，講成自己藥的成績 Quoting a competitor drug's trial results as if they were your drug's
🔢 挑錯試驗 Wrong trial	簡報引的是 A 試驗，數字卻拿自條件不同、比較漂亮的 B 試驗 The deck cites trial A but the headline number comes from trial B, run under different conditions with prettier results
🔢 計畫當結果 Plan quoted as result	引用的是研究「計畫書」（只有預期值），卻被講成已經做出來的結果 Citing a study protocol — which only states expectations — as if it reported actual results
🎭 冠錯名字 Wrong name	真實的論文被貼上一個根本不存在的研究代號 A real paper labeled with a trial acronym that doesn't exist
🎭 作者寫錯 Wrong author	作者的名與姓被顛倒等書目錯誤 Bibliographic slips like an author's given name and surname swapped

除此之外還有兩個底線保證：絕不捏造——坐實數字只認「原始全文、PubMed/Crossref、出版社頁面」三層來源，查不到就老實標記，不會編一個給你；reference 永不脫鉤——不管怎麼重排、拆頁，每一條參考文獻都跟著它支撐的內容走。

On top of that, two hard guarantees: nothing is ever made up — only three source tiers (original full text, PubMed/Crossref, publisher pages) can confirm a number, and anything unverifiable is flagged honestly instead of invented; and references never detach from content — however slides are reordered or split, every citation travels with the exact claim it supports.

一句話總結：別的工具問「這篇文獻是真的嗎？」，slide-verify 問「這篇真文獻，真的說過這個數字嗎？」

In one sentence: other tools ask "is this paper real?" — slide-verify asks "did this real paper actually say this number?"

🆕 最新更新 (What's New)

• README 全面改版：重新編排為八大段落、繁中／英文逐段對照，並新增與 GitHub 上同類查核工具的比較表，新手一眼就能看懂這個專案的定位。

• Skill v2.1：拆頁後投影片一律重新連續編號（不再用 1-a / 1-b）；每頁可加一句「🔗 銜接語」讓前後頁自然接續；「📎 待補文獻列表」成為固定產出的尾段，分成「需全文／需讀圖／灰色文獻」三區，一眼看出還缺什麼。

• 勘誤帳本 v1.3：會議海報常在 1–2 年後刊出正式論文——查證時會順手追蹤並自動升級改引全文；灰色文獻的數字可用官方新聞稿或政府網頁坐實。

• Fully restructured README: Now organized into eight sections with paragraph-by-paragraph Traditional Chinese / English, plus a comparison table against similar citation-checking tools on GitHub — a newcomer can see this project's positioning at a glance.

• Skill v2.1: Slides are renumbered consecutively after splitting (no more 1-a / 1-b); each slide can carry a one-line "🔗 bridge sentence" to connect it to the next; a "📎 Pending References" list is now a fixed final section, grouped into needs full text / needs figure reading / grey literature so you can see at a glance what is still missing.

• Verify-ledger v1.3: Conference posters are often published as peer-reviewed papers 1–2 years later — the skill now tracks these and upgrades the citation to the full paper automatically; numbers from grey literature can be confirmed via official press releases or government pages.

🚀 快速開始 (Getting Started)

1. 把這個 repo 複製到你的 Claude skills 目錄：

   Clone this repo into your Claude skills directory:

   # Claude Code
   git clone https://github.com/kau10082/slide-verify.git ~/.claude/skills/slide-verify

2. 確認以下環境需求（本 skill 不需要執行程式碼，以 Claude 本身為運算引擎）：

   Make sure the following are available (no code execution needed — Claude itself is the engine):

   • PubMed MCP（必備 / required）— 查證文獻存在與 metadata / verify paper existence and metadata
   • web_search / web_fetch — Crossref、出版社頁面、灰色文獻 / Crossref, publisher pages, grey literature
   • Google Drive MCP（選用 / optional）— 讀取你上傳的全文 PDF / read full-text PDFs you upload
   • Consensus MCP（選用 / optional）— 橫向共識，僅作導引 / cross-checking consensus, guidance only

3. 重新啟動 Claude Code（或重新整理 skills），確認清單中出現 slide-verify 即完成。

   Restart Claude Code (or refresh skills) and confirm slide-verify appears in the skill list.

🛠️ 基本使用方式 (Usage)

把簡報的文字內容貼進對話（可以是上游工具 PPT-to-MD 轉出的 Markdown，或直接複製的投影片文字），然後用自然語言下指令：

Paste the text content of your deck into the chat (Markdown from the upstream converter PPT-to-MD, or text copied straight from slides), then ask in natural language:

• 「幫我驗證／勘誤這份簡報」 — "Verify / proofread this deck for me"
• 「查證這些 reference 對不對」 — "Check whether these references are correct"
• 「把這份簡報重排成比較順的順序」 — "Reorder these slides into a smoother flow"
• 「這張投影片內容太多，幫我拆成幾張」 — "This slide is overloaded — split it up"
• 「把描述改白話一點（數字不動）」 — "Rewrite in plain language (don't touch the numbers)"

⚠️ 本 skill 只吃文字，不吃 .pptx 二進位檔——請先用上游工具（如 PPT-to-MD）把簡報轉成文字。 流程順序固定：① 驗證勘誤 → ② 重排 → ③ 拆頁 → ④ 白話改寫（選用），不可顛倒。

⚠️ This skill takes text only, not .pptx binaries — convert your deck to text first (e.g. with PPT-to-MD). The pipeline order is fixed: ① verify & correct → ② reorder → ③ split → ④ plain-language rewrite (optional).

📂 目錄結構 (Repository Structure)

slide-verify/
├── README.md                   # 本說明文件（雙語）/ this guide (bilingual)
├── LICENSE                     # MIT 授權條款 / MIT license
├── SKILL.md                    # 主要 skill 定義檔 / main skill definition
└── reference/
    └── verify-ledger.md        # 勘誤紀錄的鎖定 schema / locked schema for the correction ledger

• SKILL.md：skill 的大腦。定義整個工作流程（驗證→重排→拆頁→白話）、五大失真型態、證據來源層級，以及輸出格式。

• reference/verify-ledger.md：勘誤帳本的「鎖定 schema」。規定錯誤怎麼分類（🔧 機械錯、🎭 張冠李戴、🔢 數字不符、👻 幽靈引用、📄 灰色文獻）、旗標怎麼標（🟢🟡⚪🔴⚫）、狀態怎麼轉換。

• SKILL.md: The brain of the skill. It defines the whole pipeline (verify → reorder → split → rewrite), the five distortion patterns, the evidence-source tiers, and the output format.

• reference/verify-ledger.md: The locked schema for the correction ledger — how errors are classified (🔧 mechanical, 🎭 misattribution, 🔢 wrong number, 👻 ghost citation, 📄 grey literature), how flags work (🟢🟡⚪🔴⚫), and how statuses transition.

🤝 貢獻指南 (Contributing)

非常歡迎任何形式的幫忙！不論你是發現了新的失真型態、想改進判定規則，還是單純覺得哪段說明看不懂，都歡迎告訴我們：

Contributions of any kind are very welcome! Whether you've found a new distortion pattern, want to improve a verification rule, or simply think a section is confusing:

1. 回報問題：直接開一個 Issue，描述你遇到的狀況即可，不用拘泥格式。

2. 修改內容：先 Fork 這個專案 → 建立你的分支（git checkout -b feature/my-idea）→ 提交修改 → 發 Pull Request。

3. 如果你有真實簡報的實跑案例（去識別化後），特別歡迎分享——本 skill 的規則全部來自實跑校準。

4. Report an issue: Just open an Issue describing what you ran into — no strict format required.

5. Submit changes: Fork the repo → create a branch (git checkout -b feature/my-idea) → commit → open a Pull Request.

6. Real-world (de-identified) deck audit cases are especially welcome — every rule in this skill came from real audit calibration.

📄 授權條款 (License)

本專案採用 MIT License：你可以自由使用、修改、散布（包含商業用途），只需在你的副本中保留原始的授權與版權聲明。詳見 LICENSE。

This project is licensed under the MIT License: you are free to use, modify, and distribute it (including commercially), as long as you keep the original license and copyright notice in your copies. See LICENSE for details.