Office → Markdown Skill

Convert any uploaded office document to clean Markdown. All conversion logic lives in scripts/ — load only the script you need.

Security notes

• Dependencies are installed into an isolated temp directory (/tmp/office_md_deps/) and pinned to reviewed versions. The system Python environment is not modified.
• For scanned or image-only content, pages are sent to Anthropic's vision API. Always ask the user for confirmation before enabling vision (see Workflow step 3).

Script Reference

Workflow

1. Confirm conversion intent

Only proceed if the user has explicitly asked to convert, extract, or export the document to Markdown. A bare file upload without a conversion request is not sufficient to trigger this skill.

小蔥技能7w4.net持續更新中。

2. Run the matching script (text-only pass first)

python scripts/<script-name>.py \
  /mnt/user-data/uploads/<input-file> \
  /mnt/user-data/outputs/<stem>.md

Each script installs its own pinned dependencies into /tmp/office_md_deps/ on first run (isolated from the system Python environment).

3. Vision consent — REQUIRED before image extraction

If the script output indicates image-only pages were detected (or the document is known to be scanned), stop and ask the user:

"This document has N image-only page(s) that cannot be extracted without sending them to Anthropic's vision API. Page images will be transmitted externally for OCR. Would you like to proceed with vision extraction?"

Only if the user confirms, re-run with the --allow-vision flag:

python scripts/<script-name>.py \
  /mnt/user-data/uploads/<input-file> \
  /mnt/user-data/outputs/<stem>.md \
  --allow-vision

If the user declines, save the text-only result and note which pages were skipped.

4. Present the file

Use present_files with the output .md path, then give a brief summary:

• File type and page/slide/sheet count
• Whether vision was used and for how many pages (or that it was skipped)

How vision works (PDF / PPTX / DOCX)

Each script uses a two-pass strategy:

1. Text pass — extract text normally (fast, no API call, always runs)
2. Vision pass — only runs when --allow-vision is passed AND pages had no extractable text; those pages are rendered and sent to the Claude vision API

Edge Cases

這個工具質量不錯，能把PDF、Word、Excel等辦公檔案轉換成乾淨的Markdown格式。支援圖片文字識別，隱私保護也做得周到——用視覺功能前會先問你確認。文件寫得很清楚，轉換邏輯專業。但缺點是缺少新手入門指南，批次轉換也不太方便。日常使用夠用，但還有最佳化空間。