The Agent Skills Directory

[PROMPT_INJECTION]: The skill is designed to process arbitrary construction documents (PDF, DXF, DWG, etc.), which creates a significant surface for indirect prompt injection. Malicious instructions embedded in uploaded construction specifications or plans could attempt to override agent behavior or exfiltrate data during the extraction process.
Ingestion points: SKILL.md and references/extraction-rules.md describe the automated processing of any user-provided construction documents.
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are present in the extraction logic.
Capability inventory: The skill possesses file system access, local command execution (pdftoppm, tesseract), and optional network capabilities (Gemini API).
Sanitization: The skill lacks mechanisms to sanitize or filter natural language instructions found within processed documents.
[COMMAND_EXECUTION]: The skill includes executable Python scripts such as references/convert_dwg.py and references/parse_dxf.py. The convert_dwg.py script uses subprocess.run to execute external system binaries (ODAFileConverter) and additional Python scripts to perform CAD file conversions.
[EXTERNAL_DOWNLOADS]: The documentation in references/visual-extraction-reference.md and references/mcp-extraction-tools.md outlines optional integrations with external services, including the Google Gemini API for document validation and Cloudflare D1/R2 for state persistence, which involve data transfer to third-party infrastructure.

document-intelligence