The Agent Skills Directory

[COMMAND_EXECUTION]: The tool executes local adapter scripts via subprocess.run. This is safely implemented by requiring an adapter configuration file that must reside outside the project root, preventing malicious projects from injecting execution payloads. It uses an argument list rather than a shell string to prevent command injection.\n- [DATA_EXPOSURE]: The skill implements robust file system protections. It uses O_NOFOLLOW and directory file descriptors (dir_fd) where available to prevent symbolic link attacks and directory traversal. It explicitly prevents reading or writing sensitive project files like short-drama.json.\n- [EXTERNAL_DOWNLOADS]: The skill interacts with well-known AI services (OpenAI, Volcengine/Seedance, and MiniMax) to generate and download media. These network operations are limited to the specific tasks defined in the adapters and are used solely for downloading the results of production jobs.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests prompt text and references from the project files. While this creates a surface for indirect injection, the skill enforces a "Confirmation Gate" requiring the user to explicitly review a JSON preview of the final payload and issue a specific confirmation code before any production task is executed.

short-drama-produce