The Agent Skills Directory

[SAFE]: The skill demonstrates secure orchestration by using xml_escape boundary markers when processing user messages and external PDF content. It implements a strict synthesis hierarchy that prioritizes verbatim user excerpts over extracted data to prevent document-based prompt injection.
[PROMPT_INJECTION]: The skill mitigates indirect prompt injection via a mandatory evidence chain analysis. Ingestion points include user messages and external PDFs via pdf-toolkit. Boundary markers include xml_escape applied to all external data interpolated into LLM steps. Capability inventory includes reading via pdf-toolkit and writing to memory/pdf-intel.md via the restricted memory_save tool. Sanitization is enforced through xml_escape and length-limiting truncate filters throughout the pipeline.

meta-pdf-intelligence