The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents patterns for Retrieval-Augmented Generation (RAG) and autonomous agents that ingest data from external sources (databases, vector stores, and user-provided text).
Ingestion points: Data enters the context via retriever.ainvoke, search_database tool calls, and state variables like state['text'] in workflows.
Boundary markers: The RAG prompt template includes basic delimiters and instructions to avoid hallucination ("Answer based on the context below. If you cannot answer, say so.").
Capability inventory: Agents are configured with tools for database searching and calculation.
Sanitization: The skill provides a hardened calculate tool using Python's ast module to prevent arbitrary code execution.
[DYNAMIC_EXECUTION]: The calculate tool processes string-based mathematical expressions. It uses a recursive Abstract Syntax Tree (AST) evaluation method with a strict whitelist of allowed operators (ast.Add, ast.Sub, etc.), which is the industry-standard approach for safe runtime evaluation of arithmetic expressions.
[CREDENTIALS_UNSAFE]: The skill mentions sensitive configuration parameters such as PINECONE_API_KEY, LANGCHAIN_API_KEY, and database connection strings. However, these are implemented using secure practices (environment variable access via os.environ) or use obvious placeholders like user:pass and your-api-key. No actual secrets are exposed.

langchain-architecture