The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to use uvx opencite to download and execute an external tool from PyPI for PDF conversion as described in references/review-procedure.md.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of external dependencies at runtime, including Python packages pdf2image and pillow via uv run, and system-level utilities via brew install poppler or apt install poppler-utils.
[DYNAMIC_EXECUTION]: The skill uses python -c to execute a dynamically provided Python script for processing PDF pages in references/review-procedure.md.
[COMMAND_EXECUTION]: The procedure involves executing shell commands such as pdftoppm and opencite for document processing.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided grant proposals (PDF, Markdown, LaTeX) and possesses extensive shell and package management capabilities, creating an exploitable attack surface.
Ingestion points: Document intake described in references/review-procedure.md (PDF, Markdown, LaTeX).
Boundary markers: None identified in the instructional procedure; the skill relies on "fresh-context" isolation which is an architectural boundary rather than a prompt-level delimiter.
Capability inventory: Execution of uvx, uv run, python -c, pdftoppm, brew, and apt commands in references/review-procedure.md.
Sanitization: No explicit validation or filtering of ingested document content is performed before processing.

grant-review