Convert PDF to Markdown

When to use this skill

Trigger this skill any time there is a .pdf file that needs to be understood or processed — for example, a user attaches a PDF and asks questions about it, wants a summary, wants specific data or tables pulled out, or wants multiple PDFs in a folder processed together. PDF is a layout/print format, not reliably readable as plain text, so always convert it to Markdown first using the script in this skill rather than trying to open or parse the file directly.

This skill only supports .pdf — that's MarkItDown's only PDF-family format, so there's no legacy format to worry about here (unlike Word's .doc or Excel's .xls).

Mixed file types: When the user references a folder or set of documents containing multiple supported file types (.pdf, .docx, .xlsx), this skill handles only .pdf files. The agent MUST also invoke the sibling skills in parallel:

• convert-word-to-md for any .docx files
• convert-excel-to-md for any .xlsx files

Never process a folder and silently skip a supported file type. All three skills must be invoked together when mixed types are present.

Setup (once per environment)

Before the first conversion in a given environment, follow references/setup.md step by step to ensure Python, pip, markitdown, and pymupdf (for image extraction) are installed. Do this proactively rather than guessing whether the environment is ready — the script itself will also fail with a clear pointer back to that file if a dependency turns out to be missing, so it's safe to just try the conversion first if you're reasonably confident setup was already done.

Usage

The conversion script lives at scripts/convert_pdf_to_md.py.

Output structure: MarkItDown's PDF converter extracts text and tables only — it has no concept of embedded images at all. This script separately extracts real embedded images via PyMuPDF and writes a self-contained folder per document:

<name>/
    img/
        page001_img001.<ext>
        page002_img001.<ext>
        ...
    <name>.md

Because MarkItDown's PDF text does not preserve reliable per-page markers, there's no safe way to know exactly where inline an image belongs. Rather than risk misplacing images next to the wrong paragraph, the script appends a ## Extracted Images section at the end of the Markdown, with a ### Page N subheading per page that has images — read this section separately from the main body text. If the document has no embedded images, no img/ folder or Extracted Images section is created.

Single file:

python scripts\convert_pdf_to_md.py "C:\path\to\document.pdf"

This creates a document\ folder next to the source file (containing document.md and, if present, document\img\). To control the destination folder explicitly:

python scripts\convert_pdf_to_md.py "C:\path\to\document.pdf" -o "C:\path\to\output_folder"

A folder of PDFs (batch mode):

python scripts\convert_pdf_to_md.py "C:\path\to\folder"

Add --recursive to also include subfolders:

python scripts\convert_pdf_to_md.py "C:\path\to\folder" --recursive

Each .pdf found gets its own <name>\ output folder next to it by default. Pass -o "C:\path\to\output_parent" to collect all the generated <name>\ folders under a separate parent directory instead (subfolder structure is preserved when combined with --recursive).

After conversion, read the resulting .md file(s) to perform the actual analysis the user asked for — the script's job is only to produce accurate Markdown (and images), not to interpret the content.

Deciding where output goes

Default — always output next to the source file. The <name>/ folder is created in the same directory as the source .pdf. This is the required default for every case. Do NOT override it unless the user explicitly asks for a different location.

Only use -o when the user explicitly provides an output path (e.g., "save the output to C:\output", "put the results in D:\work"). Do NOT pass -o based on the agent's current working directory, the session state folder, or any implied location.

If the source file path cannot be fully resolved — for example, the user provides only a filename with no directory, or the path is ambiguous — use ask_user to confirm the full absolute path before running the

…