The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from several sources (including user prompts, image content via URLs, and text from PDF documents) which are subsequently integrated into the context of various LLM models.
Ingestion points: User instructions handled via llm.invoke, multimodal analysis in reference/multimodal.md, and PDF-to-image extraction logic.
Boundary markers: The skill demonstrates the use of ChatPromptTemplate and SystemMessage to organize model instructions, although it does not implement specific escape delimiters or sanitization for all external data types within the provided examples.
Capability inventory: The skill enables significant model capabilities including network communication with the OpenRouter API and tool calling for web searches, database queries, and mathematical evaluations.
Sanitization: The calculate tool in reference/langchain-integration.md follows security best practices by using the ast module to safely parse and evaluate mathematical expressions, which prevents potential code injection vulnerabilities typically associated with the use of the eval() function.

openrouter-skill