Point it at a folder of documents — PDFs, slides, Word reports — and get a structured, searchable Obsidian wiki. Powered by PaddleOCR for scanned documents and complex layouts, with Claude Vision for diagrams and charts. Zero API key required.
Your desktop is full of documents — sales decks, audit reports, technical specs, scanned contracts — scattered across formats and folders. This skill collection turns that chaos into a structured Obsidian wiki with categorized pages, wikilink cross-references, and searchable source files.
Core pipeline (docs-to-wiki): point at a directory, and the orchestrator collects all documents, converts them in parallel, OCRs scanned PDFs, merges everything, then synthesizes a categorized wiki with LLM-generated pages.
Three atomic converters handle the format-level work:
- pdf-to-md — auto-detects native-text PDFs (instant via
pymupdf) vs scanned/image PDFs. Scanned documents go through PaddleOCR, a state-of-the-art OCR engine with layout analysis — it handles multi-column text, tables, mixed Chinese/English, and complex page structures that traditional OCR tools choke on - docx-to-md — lossless text + table extraction via
python-docx. Embedded diagrams and screenshots are saved to disk and described by Claude Vision - pptx-to-md — renders every slide as a high-fidelity PNG via LibreOffice. This preserves spatial relationships — flowchart arrows, side-by-side comparisons, four-quadrant charts, numbered callouts — that shape-text extraction (pandoc, markitdown) silently drops
All three converters use agent mode by default: scripts do deterministic extraction; the calling Claude Code agent describes images using its built-in Vision — no ANTHROPIC_API_KEY needed. For large jobs (>5 images), the agent spawns subagents to keep image bytes out of the main context. Tested on 8,430 slides without context overflow.
npx skills add oCOZYo/docs-to-wikiInstalls all four skills. Add -g for global install, or --skill pdf-to-md to install just one.
Single document conversion:
- "Convert this PDF to Markdown" →
pdf-to-md - "Extract this Word doc" →
docx-to-md - "Turn these slides into notes" →
pptx-to-md
Full knowledge base:
- "Build a wiki from the docs in ./corpus/" →
docs-to-wikiorchestrates the complete pipeline
| Skill | Input | Method |
|---|---|---|
| docs-to-wiki | Directory of mixed docs | Collect → Convert → OCR → Merge → Synthesize wiki |
| pdf-to-md | PDF, JPG, PNG, BMP, TIFF, WEBP | pymupdf direct + PaddleOCR fallback |
| docx-to-md | DOCX | python-docx + extracted images |
| pptx-to-md | PPTX, PPSX | LibreOffice → per-slide PNG |
# Python dependencies
pip install pymupdf python-docx requests pyyaml
# PaddleOCR — for scanned PDFs + PPTX text-heavy slides
# Option A: Cloud API (recommended, no local setup)
export PADDLEOCR_TOKEN="your_token"
export PADDLEOCR_API_URL="https://xxxx.aistudio-app.com/layout-parsing"
# Sign up: https://aistudio.baidu.com/paddleocr (free tier)
# Option B: Local deployment (MLX VLM server + PaddleOCR-VL)
pip install paddleocr[doc-parser]
# Start MLX server with PaddleOCR-VL model, then:
export MLX_SERVER_URL=http://localhost:8111/
# LibreOffice — only for pptx-to-md
# macOS: brew install --cask libreoffice
# Linux: apt install libreofficeFor backend automation outside Claude Code, pass --api-key to any script.
MIT
桌面堆满了文档 —— 销售 PPT、审计报告、技术方案、扫描合同 —— 格式各异、散落各处。这套 skill 集合把这堆杂乱的文件变成结构化的 Obsidian 知识库：分类页面、wikilink 交叉引用、可检索的源文件。
核心流水线（docs-to-wiki）：指向一个目录，编排器自动采集所有文档、并行转换、OCR 扫描件、合并输出，最后用 LLM 合成分门别类的 wiki 页面。
三个原子转换器处理格式级工作：
- pdf-to-md —— 自动区分原生文字 PDF（
pymupdf秒级提取）和扫描/图片 PDF。扫描件走 PaddleOCR —— 业界领先的 OCR 引擎，带版面分析能力，能处理多栏排版、表格、中英混排、复杂页面结构，传统 OCR 工具搞不定的它都能搞定 - docx-to-md —— 通过
python-docx无损提取文字和表格。嵌入的图表、截图保存到磁盘，由 Claude Vision 描述 - pptx-to-md —— 通过 LibreOffice 将每张幻灯片渲染为高清 PNG。保留流程图箭头方向、左右对比、四象限图表、编号标注等空间布局信息 —— pandoc、markitdown 等 shape-text 提取工具会静默丢失这些
三个转换器默认使用 agent 模式：脚本负责确定性提取；Claude Code agent 用自带的 Vision 能力描述图片 —— 不需要 ANTHROPIC_API_KEY。大批量时（>5 张图）agent 自动 spawn subagent，图片字节不进入主上下文。实测 8430 张幻灯片无上下文溢出。
npx skills add oCOZYo/docs-to-wiki安装全部四个 skill。加 -g 装到用户全局目录，--skill pdf-to-md 只装一个。
单文档转换：
- "把这个 PDF 转成 Markdown" →
pdf-to-md - "提取这个 Word 文档" →
docx-to-md - "把这些幻灯片转成笔记" →
pptx-to-md
批量建知识库：
- "把 ./corpus/ 里的文档建成 wiki" →
docs-to-wiki编排完整流水线
| Skill | 输入格式 | 方法 |
|---|---|---|
| docs-to-wiki | 混合文档目录 | 采集 → 转换 → OCR → 合并 → 合成 wiki |
| pdf-to-md | PDF, JPG, PNG, BMP, TIFF, WEBP | pymupdf 直提 + PaddleOCR 兜底 |
| docx-to-md | DOCX | python-docx + 提取嵌入图 |
| pptx-to-md | PPTX, PPSX | LibreOffice → 逐页 PNG |
# Python 依赖
pip install pymupdf python-docx requests pyyaml
# PaddleOCR —— 扫描 PDF + PPTX 文字密集页面
# 方案 A：云端 API（推荐，无需本地部署）
export PADDLEOCR_TOKEN="your_token"
export PADDLEOCR_API_URL="https://xxxx.aistudio-app.com/layout-parsing"
# 注册：https://aistudio.baidu.com/paddleocr（免费额度）
# 方案 B：本地部署（MLX VLM 服务器 + PaddleOCR-VL）
pip install paddleocr[doc-parser]
# 启动 MLX 服务器加载 PaddleOCR-VL 模型后：
export MLX_SERVER_URL=http://localhost:8111/
# LibreOffice —— 仅 pptx-to-md 需要
# macOS: brew install --cask libreoffice
# Linux: apt install libreoffice如果要在 Claude Code 外做后台自动化，给脚本加 --api-key 参数即可。
MIT