markitdown

name	markitdown
description	Convert local files into token-efficient Markdown with Microsoft MarkItDown. Use when the user asks to convert, extract, normalize, read, or ingest PDF, Word, PowerPoint, Excel, Outlook, HTML, CSV, JSON, XML, EPUB, ZIP, image, or audio content as Markdown, and by default when supported non-Markdown files enter the Beats PM intake staging lane.

MarkItDown Intake

Use the repo wrapper so conversion stays local, repeatable, and bounded:

python3 system/scripts/markdown_intake.py "<input-file>" --output "<output-file>.md"

Workflow

1. Preserve the source file. Accept local files only; never pass an untrusted URL or enable remote fetching.
2. Run python3 system/scripts/markdown_intake.py --check. If MarkItDown is unavailable, explain that Python 3.10+ is required and ask before installing system/requirements-markitdown.txt.
3. Convert supported staged files to a sibling .md file unless the user names another local destination. Never overwrite an existing file without the user's explicit instruction.
4. Verify that the Markdown file exists and is non-empty, then use that file as the primary extraction and reasoning input. Keep the original as evidence.
5. Report scanned PDFs, image-only pages, complex tables, or unsupported media as conversion limitations. Do not enable plugins, LLM vision, Azure services, or other billable/networked processing without explicit approval.

Intake Defaults

• Automatically convert .pdf, .docx, .pptx, .xlsx, .xls, .msg, .html, .htm, .csv, .json, .xml, and .epub files during file intake.
• Keep screenshots and pasted images on the existing visual-extraction path unless the user explicitly requests Markdown conversion.
• Convert ZIP, image, or audio files only when explicitly requested because expansion, OCR, or transcription may be expensive or incomplete.
• Treat existing .md files as ready. Convert .txt only when normalization is requested.
• Reject inputs larger than the wrapper's safety limit instead of silently consuming them.

MarkItDown output is optimized for LLM analysis and structure preservation, not pixel-perfect reproduction. Prefer the original file for visual or formatting QA.