The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/convert_pdf.py script executes the external xberg CLI tool using subprocess.run. The implementation is secure as it passes arguments as a list and does not invoke a shell, mitigating shell injection risks. The script also validates the executable path before execution.
[EXTERNAL_DOWNLOADS]: The instructions in SKILL.md disclose that the PaddleOCR backend (used for Korean OCR) may download models during its first execution. This is a standard behavior for the OCR engine and is documented as a network dependency requiring user awareness.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a data processing pipeline that ingests untrusted PDF files and outputs Markdown intended for agent/LLM consumption, creating an inherent attack surface for indirect prompt injection.
Ingestion points: External PDF files provided by the user are read by the xberg tool via scripts/convert_pdf.py.
Boundary markers: The script implements a defense-in-depth measure by using randomized private tokens (secrets.token_hex(16)) for page markers during extraction to prevent document content from spoofing page structure.
Capability inventory: The skill can execute the xberg subprocess and write files to the local file system.
Sanitization: While the script validates the structural integrity and page counts of the parser's output, it does not sanitize the text content within the Markdown body, which is passed through from the PDF source.

pdf-to-markdown