The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill utilizes the marker-pdf library for high-quality OCR, which involves downloading approximately 2.5GB of pre-trained models from Hugging Face during its initial run. This is a standard requirement for the library's functionality and is explicitly mentioned in the skill documentation.
[PROMPT_INJECTION]: As the skill extracts text from external documents, it inherently creates a surface for indirect prompt injection where a document could contain instructions intended to manipulate the agent.
Ingestion points: Local documents and PDF files processed through scripts/extract_marker.py and scripts/extract_pymupdf.py.
Boundary markers: The skill does not implement specific delimiters or 'ignore' instructions when outputting extracted text, relying on the agent's internal safety layers.
Capability inventory: The scripts are functionally restricted to text extraction, image saving, and metadata retrieval; they do not perform arbitrary code execution or network operations with the extracted data.
Sanitization: Extracted content is printed as raw text or markdown without sanitization, which is typical for OCR/extraction tools.
[SAFE]: The Python scripts provided are well-structured and use standard APIs for their respective libraries. No obfuscation, persistence mechanisms, or unauthorized data exfiltration patterns were detected. The skill provides clear guidance on managing disk space for large model dependencies.

ocr-and-documents