The Agent Skills Directory

[COMMAND_EXECUTION]: The skill interpolates variables directly into shell commands within SKILL.md and its references. Specifically, the command agent-voice speak "$RESPONSE_AS_TEXT" --label "$LABEL" and the delivery template agent-voice play "$path" use double-quoted variable interpolation. If these variables contain shell-active characters like backticks or semicolons, it could lead to unintended command execution in the user's terminal.- [INDIRECT_PROMPT_INJECTION]: The skill processes assistant responses which may be influenced by external data or user prompts, creating an injection surface.
Ingestion points: The RESPONSE_AS_TEXT variable is populated from current, previous, or thread-based assistant responses.
Boundary markers: No boundary markers or specific delimiters are used to wrap the interpolated text or warn the shell environment to ignore embedded control characters.
Capability inventory: The skill has the ability to execute shell commands (agent-voice speak, uv tool install), perform setup operations, and create/remove temporary files outside the workspace.
Sanitization: There is no evidence of sanitization, escaping, or validation of the text content before it is passed to the shell.- [EXTERNAL_DOWNLOADS]: The skill instructions include commands to install a third-party tool using uv tool install agent-voice and references the source code hosted at https://github.com/yoav0gal/agent-voice. This introduces an external dependency that is downloaded and executed on the host system.

spoken-response