The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/production_tool.py script executes external adapter tools using subprocess.run with shell=False. The commands are defined in an external adapter_config JSON file which must reside outside the project directory.
[EXTERNAL_DOWNLOADS]: The scripts/provider_adapters.py script downloads generated media files from external URLs (e.g., OpenAI, Volcengine) to the project's production directory. It validates that URLs use the HTTPS protocol.
[DATA_EXFILTRATION]: The skill transmits project data, including prompts and reference images, to external generative AI service providers (OpenAI, Seedance/Volcengine, and MiniMax) to perform production tasks.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes untrusted content from project Markdown files (e.g., 图片提示词.md, 视频提示词.md) to extract prompts for generative AI tools.
Ingestion points: The _markdown_section and _copyable_prompt functions in scripts/production_tool.py read data from user-editable project files.
Boundary markers: Prompts are extracted from specific Markdown sections and are expected to be prefixed with blockquote markers (>).
Capability inventory: The tool can execute local adapter processes and make network requests to AI providers.
Sanitization: The skill implements a robust 'Hard Gate' process requiring the agent to generate a job fingerprint and the user to provide a specific confirmation token (CONFIRM <job_id> <fingerprint>) before any production command is executed. It also filters parameters for common secret keys and enforces strict path sanitization to prevent directory traversal.

short-drama-produce