The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a 'Quality Evaluation (agent-as-judge)' workflow that is susceptible to indirect prompt injection. \n
Ingestion points: Document contents are ingested via the --source argument in scripts/opensearch_ops.py and converted to chunks.\n
Boundary markers: The guide does not specify the use of delimiters or instructions to ignore embedded commands when the agent reviews 'representative chunk samples'.\n
Capability inventory: The agent reads extracted text and writes a verdict file (save-quality) which can influence subsequent re-ingestion decisions.\n
Sanitization: There is no mention of sanitization or filtering of the extracted document content before it is presented to the agent for judging.\n- [EXTERNAL_DOWNLOADS]: The skill performs automatic downloads of external resources.\n
Evidence: document_processing_guide.md notes that the first execution 'downloads AI models (~1.5 GB) automatically'. These are required by the Docling library for processing tasks like OCR and table extraction.\n- [COMMAND_EXECUTION]: The skill relies on the execution of local scripts to perform its primary functions.\n
Evidence: Multiple commands in document_processing_guide.md invoke uv run python scripts/opensearch_ops.py to handle ingestion, evaluation, and quality logging.

document-processing