The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted document data that could contain malicious instructions designed to subvert the agent's logic once extracted and processed.
Ingestion points: Documents are ingested via scripts/layout_caller.py and scripts/split_pdf.py from local paths or remote URLs.
Boundary markers: The skill instructions do not specify the use of delimiters or 'ignore' instructions when presenting extracted text to the agent context.
Capability inventory: The skill scripts have access to the local file system (via pathlib, PIL, pypdfium2) and the network (via httpx).
Sanitization: The extracted text is returned as raw markdown without filtering for potential prompt injection patterns.
[COMMAND_EXECUTION]: The skill defines several CLI utilities that the agent is expected to execute locally to perform document processing tasks such as optimization, splitting, and layout analysis.
[EXTERNAL_DOWNLOADS]: The skill fetches documents from external URLs and makes API calls to a user-configured endpoint for document parsing using the httpx library. These operations are performed over HTTPS.

paddleocr-doc-parsing