The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes subprocess.run to execute local media provider adapters. This execution is highly controlled: it uses argument arrays instead of shell strings to prevent command injection, enforces strict timeouts, and requires the adapter configuration to be located in a directory outside the project root to prevent the agent from modifying its own execution environment.
[EXTERNAL_DOWNLOADS]: The skill facilitates communication with well-known and reputable AI services, specifically OpenAI, MiniMax, and Volcengine/Seedance, to perform media generation. These network operations are conducted via HTTPS using the standard library and target established domains associated with these services.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from project documents, which serves as a potential surface for indirect prompt injection. To mitigate this, the tool implements a "Reference Contract" mechanism that appends structured metadata and boundary markers to the prompts sent to AI models, helping them distinguish between instructions and descriptions of reference assets. Furthermore, the mandatory manual confirmation step requires the user to review a full preview of the task before execution.
[DATA_EXFILTRATION]: The skill transmits project-specific prompts and reference files to AI providers for media generation, which is the primary function of the skill. Data integrity and confidentiality are protected through strict project-root path validation, input hashing, and the use of system environment variables for sensitive credentials rather than storing them in project files.

short-drama-produce