The Agent Skills Directory

[UNVERIFIABLE_DEPENDENCIES]: The skill executes npm install and python3 -m pip install pymupdf during its initial setup. Automated installation of packages from public registries or the local plugin directory can introduce supply chain risks if dependencies are not pinned to specific versions or vetted for security.\n- [COMMAND_EXECUTION]: The workflow relies on executing various Node.js and Python scripts (e.g., download-pdf.cjs, parse-pdf.js, extract-images.py) located within the resolved plugin directory. These scripts handle file system operations and external downloads.\n- [EXTERNAL_DOWNLOADS]: The skill is designed to download PDFs from user-supplied URLs, including arXiv links. While this is the intended purpose, it involves fetching and processing content from external, untrusted sources.\n- [DYNAMIC_EXECUTION]: The skill generates runnable Python scripts, Jupyter notebooks, and interactive HTML explorers based on the content of the analyzed PDFs. This creates a surface where logic derived from untrusted input is converted into executable or active content.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted PDF data which is then used by the agent to generate code and documentation. \n
Ingestion points: The paper.txt file and metadata generated by the PDF parser serve as the primary input for all generated materials.\n
Boundary markers: The instructions lack explicit delimiters or warnings to the agent to ignore potential instructions embedded within the research papers.\n
Capability inventory: The skill has access to Bash, Write, Edit, and Read tools, providing a significant range of actions that could be abused if the agent follows malicious instructions hidden in a document.\n
Sanitization: There is no specified sanitization or filtering process for the extracted text before it is processed by the agent.

claude-paper-study