The Agent Skills Directory

[SAFE]: The skill consists of documentation and illustrative code snippets for LLM application architectures. It provides templates for common industry patterns including RAG pipelines, multi-agent collaboration, and observability frameworks using standard tools.
[PROMPT_INJECTION]: The skill documents patterns for processing external data that are vulnerable to indirect prompt injection. \n
Ingestion points: The RAG_PROMPT_TEMPLATE and REACT_PROMPT templates in SKILL.md are designed to interpolate untrusted external data, such as retrieved document chunks and tool output observations, directly into the model context. \n
Boundary markers: The templates rely on weak instructional boundary markers (e.g., "Answer... based ONLY on the following context") which are known to be bypassable by adversarial content embedded in the ingested data. \n
Capability inventory: The documented patterns include capabilities for generating text via LLM APIs and executing tools based on model output, which could be leveraged if an injection succeeds. \n
Sanitization: The provided code snippets lack examples of sanitization, validation, or escaping of external content before it is processed by the LLM.

llm-app-patterns