The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from research papers and academic metadata APIs.
Ingestion points: PDF/EPUB content is converted and read by the agent; metadata is fetched via scripts/ingest.py from Crossref and Open Library.
Boundary markers: The sub-agent prompt in SKILL.md uses <paper_content> XML-style tags and an explicit instruction to ignore embedded commands within the text.
Capability inventory: scripts/ingest.py and bash scripts execute docling and pandoc via subprocess.run; scripts/rag.py manages a local vector database; file operations for notes and git commits.
Sanitization: scripts/ingest.py contains a _sanitize function that cleans metadata strings for use in citation keys.
[COMMAND_EXECUTION]: The skill uses subprocesses to run document converters.
Evidence: scripts/ingest.py calls subprocess.run to execute docling and pandoc with structured argument lists.
Evidence: scripts/pdf-to-md.sh and scripts/read-paper.sh execute the docling utility.
[EXTERNAL_DOWNLOADS]: The skill connects to academic metadata providers.
Evidence: scripts/ingest.py uses urllib.request to fetch publication data from api.crossref.org and openlibrary.org. These are well-known academic services.

reading-agent