The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes external tools for document processing. Specifically, scripts/ingest.py uses subprocess.run to call docling and pandoc. The shell scripts scripts/pdf-to-md.sh and scripts/read-paper.sh execute docling via a hardcoded path (/Users/nealcaren/.local/bin/docling). These operations are legitimate and necessary for the skill's purpose of converting academic papers to markdown.
[EXTERNAL_DOWNLOADS]: scripts/ingest.py performs network requests to fetch metadata from Crossref (api.crossref.org) and Open Library (openlibrary.org). Additionally, scripts/rag.py uses the sentence-transformers library, which may download embedding models from official Hugging Face repositories. These sources are well-known and reputable academic or technology services.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted PDF and EPUB files provided by the user, which introduces an attack surface for indirect prompt injection.
Ingestion points: Files are converted to markdown and stored in library/markdown/ before being read by sub-agents.
Boundary markers: The template in scripts/reading-agent-prompt.md uses <paper_content> tags and explicit instructions for the sub-agent to treat the content as data and ignore embedded instructions.
Capability inventory: The skill has capabilities to read files, write reading notes, and execute shell commands for conversion.
Sanitization: Content is processed through converters (docling, pandoc) which may filter some malicious formatting.

lit-synthesis