The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to execute external media adapters. This is managed by an adapter configuration file located outside the project root to prevent project-controlled command definition. Commands are executed as argument arrays without a shell, and are gated by a mandatory human-in-the-loop confirmation process where a cryptographic-like confirmation phrase must be provided.
[EXTERNAL_DOWNLOADS]: The built-in adapters fetch media production results from well-known service providers including OpenAI, MiniMax, and Volcengine (ByteDance). These downloads are restricted to successful job completions and are part of the intended primary functionality. The implementation verifies file content against expected media signatures (PNG, MP4, etc.) after download.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from project Markdown files to extract prompts and reference bindings. It mitigates injection risks by using strict Markdown header matching (e.g., IMG-*), prompt boundary enforcement (Markdown quotes), and an explicit confirmation step where the user must review the final job specification before execution. Any change in the input file hashes immediately invalidates the confirmation.
[SAFE]: The implementation includes comprehensive security measures, such as SHA-256 integrity checks for all inputs and outputs, strict path validation to prevent traversal, and O_NOFOLLOW file operations to prevent symlink attacks. It also includes a recursive scanner to detect and block secret keys in parameters before they are passed to adapters.

short-drama-produce