The Agent Skills Directory

[SAFE]: The skill serves as a development guide and template for building MCP servers with a strong focus on security and reliability in customer care environments.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill implements a robust three-layer context system (Context Variables, Global Store, Local Store) to manage sensitive data like JWT tokens and customer IDs. It explicitly instructs developers to never trust the model to provide authentication data, instead injecting it from secure stores. It uses standard environment variable practices for service credentials (e.g., OpenSearch).
[PROMPT_INJECTION]: Instructions for the agent are workflow-oriented and do not contain attempts to bypass safety filters or override system prompts. The patterns (Direct Response, Hybrid Response) are designed to reduce the risk of the model hallucinating or being manipulated by external data.
[COMMAND_EXECUTION]: No suspicious command execution patterns were found. The provided examples use standard web frameworks (Express, Starlette) and the official Model Context Protocol SDKs.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill references well-known and version-pinned packages for MCP development. There are no patterns involving the download or execution of remote scripts.
[INDIRECT_PROMPT_INJECTION]: The architecture includes security controls such as 'visibility: app' for tools that should not be callable by the model and audience annotations to control whether content is visible to the assistant or sent directly to the user. Input validation via Zod (TypeScript) and Pydantic (Python) is encouraged in the reference implementations.

customercare-mcp-builder