The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, including user-supplied raw datasets in 'data_curator.md', and prior drafts or research notes in 'intake.md'. Boundary markers are present in the form of explicit instructions to use '[UNVERIFIED]', '[EVIDENCE GAP]', or '[NEEDS DATA]' placeholders for any unsupported content, and a strict no-fabrication policy. The capability inventory includes the execution of shell commands for document compilation (LaTeX, Pandoc) and file manipulation via a 'docx' skill. Sanitization is addressed through mandatory pre-delivery checks using 'scripts/privacy_scan.py' to redact local paths and secrets, 'scripts/verify_citations.py' to detect fabricated references, and 'scripts/check_docx_fields.py' to prevent metadata leakage in Word documents.- [COMMAND_EXECUTION]: The 'references/latex-guide.md' file provides specific shell commands for compiling manuscripts using Pandoc and LaTeX toolchains (e.g., 'latexmk', 'xelatex'). These are standard utilities for the skill's primary purpose of academic document preparation.- [DATA_EXFILTRATION]: While the skill accesses potentially sensitive user data (manuscripts and datasets), it includes a dedicated privacy protection guide in 'references/privacy-and-confidentiality.md' and a scanning script 'scripts/privacy_scan.py' designed to detect and redact local filesystem paths, machine identifiers, and secrets before delivery. This represents a security best practice for handling user content.

food-paper