The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection due to its document conversion capabilities.
Ingestion points: The skill ingests untrusted data from PDFs, DOCX, XLSX, and YouTube transcripts via the markitdown library in scripts such as scripts/batch_convert.py and scripts/convert_literature.py.
Boundary markers: No explicit boundary markers or instructions to ignore embedded commands within converted text are present.
Capability inventory: The skill has permissions for file writing, network access via the requests library in scripts/generate_schematic_ai.py, and subprocess execution in scripts/generate_schematic.py.
Sanitization: There is no evidence of sanitization or filtering of extracted content before it is processed by the AI or saved to the filesystem.
[COMMAND_EXECUTION]: The script scripts/generate_schematic.py uses subprocess.run to execute a local helper script (generate_schematic_ai.py). The command construction uses an argument list which prevents standard shell injection vulnerabilities, and the execution is restricted to internal scripts.
[EXTERNAL_DOWNLOADS]: The skill facilitates the installation of the markitdown package and references Microsoft's official GitHub repository for source code. These are verified as trusted external sources.
[DATA_EXPOSURE]: The skill transmits document metadata and image content to OpenRouter and Google Gemini APIs for AI-powered processing. These are well-known services, and sensitive API keys are correctly managed via environment variables.

markitdown