The Agent Skills Directory

[PROMPT_INJECTION]: The skill is designed to handle untrusted third-party documents and implements advanced defensive measures to mitigate indirect prompt injection. It uses a randomized 16-character hexadecimal nonce (secrets.token_hex(8)) to generate unique, unpredictable block delimiters (e.g., <page-{nonce}>) for every session. This ensures that content within the PDF cannot forge delimiters to 'escape' its data block. Furthermore, the pdf_guard_text function sanitizes the text by neutralizing tag-shaped strings (e.g., <instructions>) using Unicode substitution (‹), preventing them from being interpreted as agent commands while maintaining readability.
[EXTERNAL_DOWNLOADS]: The skill utilizes standard PDF and image processing libraries including pypdfium2, pillow, pymupdf (fitz), and pypdf. It relies on the platform's native manage_packages tool for any necessary installations. These are well-known, reputable packages used for their intended purpose.
[DATA_EXFILTRATION]: No unauthorized data transmission or sensitive file access was detected. The skill uses a local cache directory (.cache/pdf-explore/) for performance, which is standard behavior. Network operations are restricted to internal platform calls (host.llm) for processing document content.

pdf-explore