The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill outlines procedures for ingesting unstructured data from local documents, which establishes a potential surface for indirect prompt injection attacks. 1. Ingestion points: SKILL.md and references/R2-sop-workflow.md utilize SimpleDirectoryReader to load content into the agent context. 2. Boundary markers: The skill focuses on architectural strategies like tree_summarize to mitigate context confusion rather than specifying strict input delimiters in baseline examples. 3. Capability inventory: The provided snippets involve file system reads and network calls to LLM providers. 4. Sanitization: The skill mitigates risks by instructing the agent to implement evaluation loops using FaithfulnessEvaluator and RelevancyEvaluator.
[COMMAND_EXECUTION]: The skill provides Python code snippets for the agent to execute, performing operations like building vector indices and managing ingestion pipelines. Evidence: Files such as SKILL.md and references/R2-sop-workflow.md contain code for initializing VectorStoreIndex and global Settings, which are standard operations for the described RAG workflow using well-known services.

agentsop-llamaindex