The Agent Skills Directory

[Architecture for Data Ingestion]: The documentation describes the ingestion of external data such as user preferences, shopping carts, and page content into a dynamic context block. While this represents a surface for indirect prompt injection, the skill advocates for the use of architectural boundaries to maintain separation between instructions and data.
[Boundary and Security Controls]: The skill outlines the use of 'data fences' and 'fence notices' within the prompt assembly process. These are intended to serve as delimiters that help the agent distinguish between its core instructions and potentially untrusted data provided during a turn.
[Tool Registry and Access Control]: The reference describes a centralized tool registry where capabilities like web search and data analysis are managed. It specifies that tool availability is determined by a fixed configuration and an executor, rather than being dynamically generated from per-request input, which is a positive security pattern.
[Verification and Testing]: The skill includes instructions for verifying the integrity and efficiency of prompt assembly through logging and automated tests. This promotes transparency and allows for the detection of unexpected changes in prompt structure or content.

commerce-prompt-caching