The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local Python scripts, such as scripts/fetch_oa.py and scripts/privacy_scan.py, and system utilities like Pandoc to automate research grounding and report formatting.
[EXTERNAL_DOWNLOADS]: The knowledge_builder agent fetches full-text academic papers from open-access repositories and institutional libraries to verify the accuracy of the manuscript's citations and provide a state-of-the-art context.
[PROMPT_INJECTION]: The skill processes user-supplied manuscripts, which presents an inherent surface for indirect prompt injection (Category 8). The risk is mitigated by the following evidence chain: (1) Ingestion point: Manuscript files provided in Word or Markdown format. (2) Boundary markers: Specialized reviewer instructions that compartmentalize analysis (e.g., separate methodology, domain, and integrity checks). (3) Capability inventory: Access to local file tools and scholarly search APIs via specialized sub-skills. (4) Sanitization: A dedicated privacy scan is performed on the final report to prevent leakage of local file paths or secrets.

food-review