The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input (product names, descriptions, and shopping requirements) to generate video scripts and visual content.
Ingestion points: User requirements and product information are captured in Step 1 (SKILL.md) and propagated to the Scriptwriter and Visual Designer agents.
Boundary markers: The skill relies on rigid JSON schemas for agent inter-communication, which provides implicit data boundaries, though it lacks explicit instructions to ignore embedded commands in the raw product data.
Capability inventory: The skill documented in 'references/tech-integration.md' has capabilities for network communication (AI APIs) and local shell execution (FFmpeg).
Sanitization: There is no explicit documentation of sanitization or escaping mechanisms for the user-supplied product strings before they are incorporated into agent prompts.
[COMMAND_EXECUTION]: The skill instructions include system-level commands for setting up dependencies and processing media.
Evidence: 'references/tech-integration.md' specifies the installation of media tools using 'apt-get install -y ffmpeg'.
Evidence: 'references/tech-integration.md' provides Python code to execute the 'ffmpeg' binary using the 'subprocess.run' function for video composition.
[DYNAMIC_EXECUTION]: The skill utilizes dynamic execution to handle media processing at runtime.
Evidence: Media synthesis is handled by invoking the 'ffmpeg' binary with dynamically constructed arguments for scaling, overlays, and encoding.
[EXTERNAL_DOWNLOADS]: The skill fetches resources and interacts with external AI service providers.
Evidence: Integration code in 'references/tech-integration.md' makes network requests to well-known services including Azure TTS (microsoft.com), Baidu TTS (baidu.com), HeyGen (heygen.com), D-ID (d-id.com), and Jianying/CapCut (jianying.com). These interactions are documented for legitimate functionality.

digital-avatar-shopping-video