The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from PDF documents that could contain malicious instructions designed to influence the agent's behavior.
Ingestion points: The ingest_policy_document function reads content from PDF files and stores them in a PostgreSQL database (policy_chunks table).
Boundary markers: The prompt uses a clear 'Policy Documents:' header and explicit instructions to answer 'ONLY based on the provided policy documents'.
Capability inventory: The skill utilizes psycopg2 for database operations and pypdf for file reading; it does not perform arbitrary shell commands or external network requests (other than to the database).
Sanitization: The skill uses parameterized SQL queries (%s placeholders) which effectively prevents SQL injection during both ingestion and retrieval.
[EXTERNAL_DOWNLOADS]: The skill downloads pre-trained machine learning models from well-known services.
Evidence: Downloads sentence-transformers/all-MiniLM-L6-v2 and cross-encoder/ms-marco-MiniLM-L-6-v2 from HuggingFace repositories.
[COMMAND_EXECUTION]: The skill requires Python code execution to process embeddings, manage database connections, and run the RAG logic.
Evidence: The frontmatter explicitly enables code_execution: python and the workflow involves executing several Python functions defined in the markdown.

policy-qa