Document to Markdown — Agent Guide

Document to Markdown — Agent Guide

You are helping a developer convert binary or non-LLM-readable document formats into clean Markdown. The primary tool is markitdown by Microsoft — a Python CLI and library purpose-built for LLM consumption. Use fallback tools only when markitdown doesn't support the format or produces poor output.

Installation

Primary: markitdown

# Basic install
pip install markitdown

# Full install (includes OCR, all optional deps)
pip install 'markitdown[all]'

Verify installation

markitdown --help

Script Prerequisites

Scripts auto-install missing packages in interactive terminals. In non-interactive environments (LLM agents, CI), packages are installed automatically without prompting.

Quick Install (all dependencies)

pip install markitdown pdfplumber mammoth pandas tabulate openpyxl mpxj jpype1

Standalone Scripts

This skill includes ready-to-run Python scripts in the scripts/ directory.

⚠️ IMPORTANT: Treat scripts as black boxes. Run python scripts/<script>.py --help to learn usage. Do NOT read the source code of these scripts — it will pollute your context window with implementation details you don't need. The scripts are designed to be used via their CLI interface only.

When the user needs conversion, execute these directly via run_in_terminal instead of generating code from scratch.

All scripts accept --help for full usage. Paths are passed as parameters — never hardcoded.

Working directory: Run scripts from any directory using absolute paths, or from the skill's root directory using relative paths (python scripts/​convert.py ...).

When the user asks for conversion, prefer running the bundled scripts over generating inline code. Adjust parameters to match the user's specific paths and requirements.

Format → Tool Mapping

Use this table to determine the conversion approach:

Conversion Guide

Quick One-Off (CLI)

For a single file, use the CLI:

markitdown path/​to/​file.docx > output.md

This works for: .docx, .xlsx, .pptx, .pdf, .html, .epub, .csv, .xls

Programmatic (Python API)

For integration into scripts or when you need to process the output:

from markitdown import MarkItDown

md = MarkItDown()
result = md.convert("path/​to/​file.docx")
print(result.text_content)

To save with metadata frontmatter:

from markitdown import MarkItDown
from datetime import datetime
from pathlib import Path

def convert_with_metadata(input_path: str, output_path: str | None = None) -> str:
    md = MarkItDown()
    result = md.convert(input_path)

    source = Path(input_path).name
    date = datetime.now().strftime("%Y-%m-%d")
    frontmatter = f"---\nsource: \"{source}\"\nconverted: \"{date}\"\n---\n\n"
    content = frontmatter + result.text_content

    if output_path:
        Path(output_path).write_text(content, encoding="utf-8")
    return content

Per-Format Instructions

Word (.doc, .docx)

Use markitdown directly. If output has formatting issues:

1. Try mammoth for .docx — it produces cleaner HTML-to-Markdown for complex layouts
2. Try pandoc: pandoc -f docx -t markdown file.docx -o output.md
3. For legacy .doc: convert to .docx first via LibreOffice headless if markitdown fails:

   libreoffice --headless --convert-to docx file.doc
   markitdown file.docx > output.md
   

Excel (.xlsx, .xls, .csv)

markitdown renders each sheet as a Markdown table. If you need more control:

import pandas as pd

df = pd.read_excel("file.xlsx", sheet_name=None)  # all sheets
for name, sheet in df.items():
    print(f"## Sheet: {name}\n")
    print(sheet.to_markdown(index=False))
    print()

For CSV:

import pandas as pd
df = pd.read_csv("file.csv")
print(df.to_markdown(index=False))

PowerPoint (.ppt, .pptx)

markitdown extracts text slide-by-slide. If you need structured extraction with slide titles:

from pptx import Presentation

prs = Presentation("file.pptx")
for i, slide in enumerate(prs.slides, 1):
    title = slide.shapes.title
    print(f"## Slide {i}: {title.text if title else 'Untitled'}\n")
    for shape in slide.shapes:
        if shape.has_text_frame:
            for para in shape.text_frame.paragraphs:
                print(para.text)
    print()

PDF (.pdf)

markitdown handles most PDFs. For table-heavy PDFs, pdfplumber gives better results:

import pdfplumber

with pdfplumber.open("file.pdf") as pdf:
    for page in pdf.pages:
        text = page.extract_text()
        if text:
            print(text)
        tables = page.extract_tables()
        for table in tables:
            # Convert to Markdown table
            header = "| " + " | ".join(str(c) for c in table[0]) + " |"
            sep = "| " + " | ".join("---" for _ in table[0]) + " |"
            rows = "\n".join("| " + " | ".join(str(c) for c in row) + " |" for row in table[1:])
            print(f"{header}\n{sep}\n{rows}\n")

Visio (.vsd, .vsdx)

No direct Python converter exists. Use this cascade:

1. Best: Ask the user to export from Visio as PDF or SVG, then convert that
2. Automated: Use LibreOffice headless to convert to PDF, then use markitdown or PyMuPDF:

   libreoffice --headless --convert-to pdf diagram.vsdx
   markitdown diagram.pdf > output.md
   

3. OCR route: Convert to image, then use markitdown with OCR support:

   libreoffice --headless --convert-to png diagram.vsdx
   pip install 'markitdown[all]'
   markitdown diagram.png > output.md
   

Always warn the user: Visio conversion is lossy. Complex diagrams with connectors and swimlanes will lose structural information. Suggest Mermaid diagrams as the target representation if the user needs to preserve diagram logic.

MS Project (.mpp, .mspdi, .mpx)

Natively supported via convert.py --tool mpxj. Requires Java 11+ (auto-detected).

python scripts/​convert.py project.mpp -o project.md

Output includes: project properties, resources table, full task table (WBS, dates, durations, predecessors, resources), and summary statistics. The script uses MPXJ's UniversalProjectReader which handles .mpp (all versions), .mspdi (XML), and .mpx formats.

If Java is not available, ask the user to:

1. Install Java 11+: choco install temurin17 (Windows) or apt install openjdk-17-jdk (Linux)
2. Or export from MS Project as CSV, then: python scripts/​convert.py project.csv -o project.md

Images (.png, .jpg, .jpeg, .gif, .bmp, .tiff)

Requires the full install: pip install 'markitdown[all]'

markitdown uses OCR to extract text from images. Best results with:

• High-resolution images (300+ DPI)
• Clean, printed text (not handwriting)
• Good contrast between text and background

markitdown scan.png > output.md

If OCR quality is poor, suggest the user try preprocessing the image (increase contrast, deskew) or use a dedicated OCR tool like Tesseract directly.

Audio (.mp3, .wav, .m4a)

Requires the full install: pip install 'markitdown[all]'

markitdown transcribes audio files using speech-to-text. Useful for meeting recordings, voice notes, and dictation.

markitdown meeting.mp3 > transcript.md

For long audio files (> 30 min), warn the user that conversion may take significant time and the transcript should be reviewed for accuracy. For higher-quality transcription, suggest OpenAI Whisper directly:

pip install openai-whisper
whisper meeting.mp3 --output_format txt

Jupyter Notebook (.ipynb)

markitdown extracts notebook cells natively — code cells, markdown cells, and outputs.

markitdown analysis.ipynb > analysis.md

Fallback with nbconvert for more control over output:

jupyter nbconvert --to markdown analysis.ipynb

ZIP Archive (.zip)

markitdown extracts the ZIP contents and converts each supported file inside:

markitdown documents.zip > contents.md

The output includes all convertible files from the archive with headers indicating the source file within the ZIP. Non-convertible files are listed but skipped.

Outlook Email (.msg)

markitdown extracts email headers (From, To, Subject, Date) and body content:

markitdown email.msg > email.md

Attachments are listed but not automatically converted. If the user needs attachment content, suggest extracting the .msg file first and then converting each attachment separately.

RSS/​ATOM Feed (.xml)

markitdown can parse RSS and ATOM feed XML files, extracting entries as markdown:

markitdown feed.xml > feed.md

Note: This only works with RSS/​ATOM feed XML, not arbitrary XML files. For arbitrary XML, the LLM can read it directly — no conversion needed.

Batch Processing

Directory Batch Conversion Script

When the user needs to convert multiple files, provide this pattern:

from markitdown import MarkItDown
from pathlib import Path
from datetime import datetime

SUPPORTED = {".docx", ".doc", ".xlsx", ".xls", ".csv", ".pptx", ".ppt", ".pdf", ".html", ".epub", ".png", ".jpg", ".jpeg", ".gif", ".bmp", ".tiff", ".mp3", ".wav", ".m4a", ".ipynb", ".zip", ".msg"}

def batch_convert(input_dir: str, output_dir: str | None = None):
    md = MarkItDown()
    input_path = Path(input_dir)
    output_path = Path(output_dir) if output_dir else input_path

    if output_dir:
        output_path.mkdir(parents=True, exist_ok=True)

    converted = []
    failed = []

    for file in input_path.rglob("*"):
        if file.suffix.lower() not in SUPPORTED:
            continue
        try:
            result = md.convert(str(file))
            out_file = output_path / file.relative_to(input_path).with_suffix(".md")
            out_file.parent.mkdir(parents=True, exist_ok=True)

            frontmatter = f"---\nsource: \"{file.name}\"\nconverted: \"{datetime.now().strftime('%Y-%m-%d')}\"\n---\n\n"
            out_file.write_text(frontmatter + result.text_content, encoding="utf-8")
            converted.append(str(file))
        except Exception as e:
            failed.append((str(file), str(e)))

    print(f"Converted: {len(converted)} files")
    for f in converted:
        print(f"  ✓ {f}")
    if failed:
        print(f"Failed: {len(failed)} files")
        for f, err in failed:
            print(f"  ✗ {f}: {err}")

CLI Batch (one-liner)

for f in docs/​*.docx; do markitdown "$f" > "${f%.docx}.md"; done

PowerShell:

Get-ChildItem docs\*.docx | ForEach-Object { markitdown $_.FullName | Set-Content ($_.FullName -replace '\.docx$', '.md') -Encoding UTF8 }

Output Quality Tips

After conversion, apply these cleanup steps when output quality matters:

1. Strip excessive whitespace: Collapse 3+ consecutive blank lines to 2
2. Normalize headings: Ensure heading levels are sequential (no jumping from # to ###)
3. Add frontmatter: Include source, converted date, and optionally format fields
4. Handle images: markitdown with [all] uses OCR to describe images as alt-text. Without OCR, images appear as empty image references — warn the user about this
5. Table cleanup: markitdown sometimes produces tables with misaligned columns — advise using a Markdown formatter if tables look broken

When the User Asks for Help

• "Convert this document to Markdown" → Identify the file format from the extension. Run python scripts/​convert.py <path> -o <output> via terminal. Use --tool to force a specific converter if markitdown output is poor. Always verify markitdown is installed first (pip install markitdown).

• "I need to batch-convert a folder of documents" → Run python scripts/​batch_convert.py <dir> -o <output_dir> -r via terminal. Use --dry-run first to preview. Ask what formats to include (-f docx,pdf) and whether to recurse into subdirectories.

• "The converted output looks bad / has formatting issues" → Read the troubleshooting guide. Check if the format has a better fallback tool in the mapping table. For tables, suggest pdfplumber (PDF) or pandas (Excel). For complex Word layouts, suggest mammoth.

• "How do I convert a Visio diagram?" → Explain that no direct converter exists. Offer the cascade: manual export → LibreOffice headless → OCR. Suggest Mermaid as a better representation for diagrams that need to stay editable.

• "How do I convert an .mpp file?" → Explain mpxj option (requires JVM) or suggest exporting from MS Project as CSV/​XML first. Provide the appropriate parsing code.

• "Can I convert images/​scanned documents?" → Yes, run python scripts/​convert.py scan.png --ocr -o output.md. Requires pip install 'markitdown[all]' for OCR. Quality depends on image resolution and text clarity.

• "Can I transcribe audio to Markdown?" → Yes, run python scripts/​convert.py meeting.mp3 -o transcript.md. Requires pip install 'markitdown[all]' for speech-to-text. For long files (> 30 min), warn about processing time. For better quality, suggest OpenAI Whisper.

• "Convert a Jupyter notebook?" → Run python scripts/​convert.py notebook.ipynb -o notebook.md. markitdown extracts code cells, markdown cells, and outputs natively.

• "What about password-protected files?" → markitdown does not handle encrypted/​password-protected files. The user must remove protection first. For PDFs, suggest qpdf --decrypt if the password is known.

For detailed tool comparisons, see tools-reference.md. For code examples by format, see examples.md. For troubleshooting common issues, see troubleshooting.md.

Quick Reference

Companion Skills

• For the reverse direction (Markdown → documents): use markdown-to-document
• For template pattern extraction before generating documents: use markdown-to-document (analyze_template.py)