The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface because it processes untrusted data from external documents.\n
Ingestion points: The skill ingests user-provided PDF, DOCX, PPTX, and XLSX files which are converted to text using the Docling library, as described in opensearch-launchpad/SKILL.md and opensearch-launchpad/document_processing_guide.md.\n
Boundary markers: Instructions do not specify the use of delimiters or 'ignore' directives to separate the ingested document content from the agent's internal logic.\n
Capability inventory: The skill can execute shell commands (via bash and uv), write to the filesystem (indexing), and perform network operations to OpenSearch and AWS Bedrock APIs.\n
Sanitization: There is no mention of text sanitization or validation for the content extracted from processed documents before it is used in indexing or evaluation.\n- [EXTERNAL_DOWNLOADS]: The skill downloads and installs several external dependencies and tools from public registries.\n
Installs the docling Python library from IBM Research for document conversion.\n
Fetches and runs MCP servers including duckduckgo-mcp-server and the vendor's own opensearch-mcp-server-py using the uvx tool runner.\n- [COMMAND_EXECUTION]: The skill workflow involves executing various scripts and dynamic code snippets.\n
Runs local bash and Python scripts (scripts/start_opensearch.sh and scripts/opensearch_ops.py) to manage the OpenSearch infrastructure.\n
Executes inline Python code snippets via uv run python -c in opensearch-launchpad/evaluation_guide.md to calculate search quality metrics.

search