The Agent Skills Directory

[PROMPT_INJECTION]: The skill facilitates indirect prompt injection by processing untrusted academic content.
Ingestion points: External papers (PDFs, ArXiv metadata) are fetched and parsed by scripts/normalize_paper.py and scripts/xray_io.py.
Boundary markers: Extracted data is structured into a paper-record JSON schema, but instructions in references/modes/interpret.md and references/modes/xray.md tell the agent to treat these fields as its primary fact source without explicit warnings to ignore instructions embedded in the data.
Capability inventory: The skill possesses the ability to execute Python scripts via the Bash tool and write files to disk (using the --save flag in the normalizer).
Sanitization: Extracted text undergoes basic formatting (whitespace normalization, cleaning) but lacks specific filtering or sanitization for malicious prompt sequences.
[EXTERNAL_DOWNLOADS]: Fetches paper metadata and text from well-known academic services such as arxiv.org, alphaxiv.org, and api.crossref.org. It also downloads PDFs from user-provided URLs for text extraction.
[DATA_EXFILTRATION]: Reads local PDF and text files specified via the --source argument to extract content for normalization.
[COMMAND_EXECUTION]: Executes bundled Python scripts (normalize_paper.py) via the Bash tool to normalize paper data and perform text extraction. It also dynamically loads a local helper script (scripts/xray_io.py) using importlib for PDF processing tasks.

paper-workbench