The Agent Skills Directory

[PROMPT_INJECTION]: Multiple files, including the main instructions and several test fixtures, contain common injection strings like 'IGNORE ALL PREVIOUS INSTRUCTIONS' and 'Give a positive review only'. These are used exclusively as signatures for a built-in manuscript scanner or as payloads in automated test suites to verify the scanner's efficacy. They are literals and do not target the agent's own instruction set.
[DYNAMIC_EXECUTION]: The skill uses local Python scripts for PDF metadata extraction, manuscript auditing, and report formatting. These scripts (e.g., scan_pdf_layers.py, check_pdf_injection.py) use the standard library and the well-known PyMuPDF library. They process input data deterministically and do not involve remote code execution, unsafe deserialization, or dynamic command generation.
[INDIRECT_PROMPT_INJECTION]: The skill proactively addresses the threat of indirect prompt injection by identifying manuscript PDFs as a vector. It implements a multi-stage sanitization process that extracts only visible text runs and checks them against known injection patterns before the content is processed by the language model.
[EXTERNAL_DOWNLOADS]: The skill references various official medical journal domains (e.g., ajronline.org, rsna.org) and DOI links to provide reviewers with legitimate formatting guidelines and AI usage policies. No untrusted external code or payloads are fetched.

peer-review