The Agent Skills Directory

[DYNAMIC_EXECUTION]: The calculatorTool example uses the JavaScript eval() function to process user-provided mathematical expressions. This allows for arbitrary code execution within the application environment if the input is not strictly sanitized or restricted to safe mathematical operations.
Evidence: const result = eval(expression); // Use safer math parser in production in SKILL.md.
[INDIRECT_PROMPT_INJECTION]: The skill provides templates for Retrieval-Augmented Generation (RAG) and Tool use that ingest external data (files, search results, database records) and interpolate them directly into LLM prompts without explicit sanitization or robust boundary markers to separate untrusted data from system instructions.
Ingestion points: DirectoryLoader (local file system), searchTool (external search API), and dbQueryTool (database records) in SKILL.md.
Boundary markers: Standard prompt templates are used (e.g., Context: {context}), but no explicit instructions are provided to the model to ignore potential directives embedded within that context.
Capability inventory: The skill templates include file reading, network searching, and database querying capabilities.
Sanitization: No sanitization, escaping, or validation logic is included in the retrieval or tool output processing examples.
[COMMAND_EXECUTION]: The skill provides instructions for installing multiple Node.js packages. While the packages listed are well-known and standard for LangChain development, this illustrates a pattern of modifying the local environment via package managers.
Evidence: npm install langchain @langchain/openai @langchain/community in SKILL.md.

langchain-workflow-builder