The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes tools to render and install client adapters into local development environments (e.g., VS Code, Cursor). These actions are performed via an MCP server and involve file system writes. The skill mitigates risk by requiring the agent to display the full file content and the exact destination to the user before any modification occurs.
[INDIRECT_PROMPT_INJECTION]: The skill ingests user-provided data such as workspace directories and model identifiers, which represents a potential injection surface.
Ingestion points: User input for workspace paths and client-native model IDs as defined in SKILL.md.
Boundary markers: The instructions mandate the use of unique 'install tokens' that the user must repeat verbatim, specifically forbidding the acceptance of generic 'yes' or 'confirm' responses.
Capability inventory: The skill can call save_preferences, render_client_adapter, and install_client_adapter to modify project or user settings.
Sanitization: The underlying MCP server is instructed to compute allowlisted destinations based on the provided workspace, preventing the skill from being used to write to arbitrary system paths.

setup