Office Document Assistant

---

name: office-document-assistant description: Read, extract, summarize, and compare office documents including PDF, Word, Excel, and PowerPoint. Use when a user provides .pdf/.doc/.docx/.xls/.xlsx/.ppt/.pptx files and asks for summaries, key point extraction, page-by-page outlines, field extraction, table explanation, or multi-document comparison. Prefer the bundled extraction script for deterministic text extraction; for PDFs, fall back to OCR when embedded text is missing.

---

Office Document Assistant

Read, extract, summarize, and compare common office documents:

PDF

Word (.docx, .doc)

Excel (.xlsx, .xls)

PowerPoint (.pptx, .ppt)

Use this skill when the user wants the contents of a document explained, summarized, searched, or extracted into a simpler structure.

When to Use

Use this skill when the user:

uploads a .pdf / .doc / .docx / .xls / .xlsx / .ppt / .pptx

asks to summarize a document

asks to extract dates, amounts, contacts, conclusions, specifications, risks, or action items

asks for page-by-page / slide-by-slide structure

asks what a spreadsheet or slide deck is saying

asks to compare two or more documents after extracting their text

When Not to Use

Do not position this skill as a high-fidelity layout or visual analysis system.

It is not ideal for:

precise preservation of original layout, formatting, or pagination

detailed chart / diagram / image interpretation

password-protected or encrypted files

OCR-heavy image understanding beyond basic text recovery

advanced spreadsheet analytics or formula auditing

tracked-changes / redline reconstruction in Office documents

Core Workflow

1. Confirm the document path. 2. Run the bundled script: - python3 {skill_dir}/scripts/extract_office_text.py --json 3. Inspect the JSON fields: - type - extraction - warning - truncated - text 4. Separate clearly in your response: - directly extracted content - your summary / inference based on that content 5. If extraction is empty or weak: - for PDF, check OCR availability first - for legacy Office formats, check conversion tools 6. If the user asks for a summary, default to: - one-sentence overview - 3–8 key points - extra sections only when clearly present (dates, people, risks, data, conclusions, contacts) 7. If the user asks for extraction, prefer structured fields over long prose.

Supported Formats and Strategy

PDF

First extract embedded text with pypdf.

If extracted text is too short, fall back to OCR.

OCR prefers chi_sim+eng, then chi_sim, then eng.

OCR pipeline requires both pdftoppm and tesseract.

If an official first-class PDF tool is exposed in the environment and the task is high-value or multi-PDF, you may prefer that tool; otherwise use this skill's script.

Word

.docx: extract paragraphs and tables directly.

.doc: try antiword, then catdoc, then LibreOffice conversion to .docx.

Excel

Extract sheet names and the first rows of each sheet.

Best for quickly understanding workbook structure and core fields.

When explaining, focus on what each sheet represents, key columns, important figures, and obvious anomalies.

PowerPoint

Extract slide text from shapes.

Extract speaker notes when present.

Summaries should usually be slide-by-slide or theme-based, not a giant raw dump.

Tools and Dependencies

Document clearly what is required versus optional.

Required runtime

python3

Required Python packages

pypdf — embedded text extraction from PDFs

python-docx — .docx extraction

openpyxl — .xlsx extraction

python-pptx — .pptx extraction

Optional but strongly recommended system tools

poppler-utils — provides pdftoppm for PDF → image conversion before OCR

tesseract-ocr — OCR engine

tesseract-ocr-chi-sim — Simplified Chinese OCR language pack

libreoffice — conversion fallback for legacy .doc, .xls, .ppt

antiword — direct .doc extraction fallback

catdoc — additional .doc extraction fallback

What each tool is used for

pypdf: try text-layer extraction from PDFs first

pdftoppm: rasterize PDF pages when OCR is needed

tesseract: recover text from scanned/image PDFs

python-docx: read paragraphs and tables from .docx

openpyxl: read sheets and rows from .xlsx

python-pptx: read slide text and notes from .pptx

libreoffice: convert older Office formats into newer parseable formats

antiword / catdoc: lightweight extraction options for .doc

Minimum useful setup

If only modern documents matter, the minimum practical setup is:

python3

Python packages: pypdf, python-docx, openpyxl, python-pptx

Recommended full setup

For the most robust behavior across real-world files, install:

python3

Python packages: pypdf, python-docx, openpyxl, python-pptx

system tools: poppler-utils, tesseract-ocr, tesseract-ocr-chi-sim, libreoffice, antiword, catdoc

Dependency check

Use the bundled checker to quickly see what is missing in the current environment:

python3 {skill_dir}/scripts/check_deps.py

Common Commands

python3 {skill_dir}/scripts/extract_office_text.py "/path/to/file.pdf" --json
python3 {skill_dir}/scripts/extract_office_text.py "/path/to/file.docx" --json
python3 {skill_dir}/scripts/extract_office_text.py "/path/to/file.xlsx" --json
python3 {skill_dir}/scripts/extract_office_text.py "/path/to/file.pptx" --json

Useful flags:

# limit PDF pages scanned/extracted
python3 {skill_dir}/scripts/extract_office_text.py "/path/to/file.pdf" --page-limit 10 --json
limit rows per sheet when probing spreadsheets
python3 {skill_dir}/scripts/extract_office_text.py "/path/to/file.xlsx" --row-limit 30 --json
cap output text size
python3 {skill_dir}/scripts/extract_office_text.py "/path/to/file.pdf" --max-chars 30000 --json

Output Style

Default to a compact answer:

one-sentence summary

3–8 key points

then expand only if the user asks for:

- detailed summary - page-by-page / slide-by-slide notes - field extraction - document comparison

Failure Handling

If PDF text is empty, suspect scanned pages or missing OCR tools.

If Chinese OCR is weak, check whether tesseract-ocr-chi-sim is installed.

If .doc / .xls / .ppt extraction fails, check libreoffice, antiword, and catdoc.

If tables look messy, explain that this is text-first extraction rather than full layout reconstruction.

If a file is encrypted or unreadable, say so plainly and stop guessing.

References

Read these only when needed:

references/capabilities.md — capability boundaries and what each format can/can't do well

references/troubleshooting.md — dependency checks and common failure modes