The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data from platform tool outputs and user input, which creates a surface for potential indirect prompt injection.
Ingestion points: Data enters the agent context through tools like list_projects, find_mcp, list_access_roles, and list_access_members, as well as user-provided search queries (SKILL.md).
Boundary markers: The instructions explicitly require the agent to stop at "confirmation points," "ask the user before any write," and present eligible choices for explicit user selection rather than inferring them.
Capability inventory: The skill utilizes mutation tools such as create_mcp_access_role, update_mcp_access_role, and assign_mcp_access_role to modify access control state.
Sanitization: The skill mandates using only "masked identities" and "privacy-safe member counts" provided by the platform tools to avoid exposing raw identifiers or sensitive user data.
[COMMAND_EXECUTION]: The skill references a platform execute tool for running multi-step read operations.
Evidence: The workflow describes how the execute tool may run tools.search, tools.describe, and tools.call on specific platform--<tool name> paths.
Safeguards: These operations are scoped to platform paths, and the skill includes a critical safety rule to "ask the user before any write" and end each execute run at a confirmation point.

manage-mcp-access