The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external document files (PDF/EPUB) which are converted to markdown and indexed for semantic search. The LLM processes search results containing this untrusted text, which could contain malicious instructions.
Ingestion points: The scripts/ingest.py script reads user-provided PDF and EPUB files and converts them to markdown files in library/markdown/.
Boundary markers: No specific delimiters or instructions are provided to the agent to treat the indexed text chunks as untrusted content when they are retrieved via rag.py search or rag.py context.
Capability inventory: The skill utilizes subprocess.run to execute conversion tools and performs file write operations to the local library and bibliography files.
Sanitization: There is no evidence of text sanitization or filtering to remove potential prompt injection sequences from the documents before they are indexed or returned to the agent.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to fetch metadata and models.
Metadata APIs: scripts/ingest.py fetches bibliographic data from api.crossref.org and openlibrary.org.
Embedding Model: scripts/rag.py utilizes sentence-transformers which downloads the all-MiniLM-L6-v2 model from Hugging Face on the first run.
[COMMAND_EXECUTION]: The scripts/ingest.py script uses subprocess.run to call external utilities for file conversion.
Evidence: It executes docling for PDF-to-markdown conversion and pandoc for EPUB-to-markdown conversion. These calls use list-based arguments rather than shell strings, reducing the risk of shell injection.

zotero-rag