The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-supplied documents (PDF, images, Office) and remote URLs, converting their content into Markdown. Malicious instructions embedded in these documents could influence subsequent agent actions.
Ingestion point: The main entry point scripts/convert.py accepts file paths and URLs as input.
Boundary markers: No explicit prompt boundary markers or instruction-ignoring delimiters are used during processing.
Capability inventory: The skill can write files to the local system, download remote assets, and execute Python scripts via uv run.
Sanitization: Basic filename sanitization is performed, and output is cleaned using regex patterns.
[EXTERNAL_DOWNLOADS]: The skill downloads document processing results and image assets from remote sources.
The MinerU backend (scripts/mineru_ocr.py) downloads ZIP archives containing conversion results and extracts them locally.
The skill fetches image assets from remote URLs found within processed documents using urllib.request.urlretrieve.
[COMMAND_EXECUTION]: The skill executes Python scripts using the uv run tool and includes a macOS JXA wrapper (scripts/convert.js) that executes shell commands via app.doShellScript.
[DYNAMIC_EXECUTION]: The skill uses dynamic module loading for optional dependencies and diagnostics.
Late imports are used in scripts/rapid_ocr.py for rapidocr, numpy, and pillow to handle optional local OCR capabilities.
The __import__('sys') function is called in scripts/mineru_ocr.py and scripts/paddle_ocr.py for diagnostic logging to standard error.

legal-ocr