The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data, including image URLs and script text, which are passed into shell commands for processing by AI models.
Ingestion points: The workflow in SKILL.md ingests user-supplied image URLs for character training (phota/train) and generated script text for speech synthesis (inworld/text-to-speech-2).
Capability inventory: The skill utilizes the Bash tool to invoke the infsh CLI, allowing it to execute commands that process these inputs.
Boundary markers: There are no defined delimiters or instructions to the agent to treat user-provided data as strictly data, which could lead to instructions in the script being interpreted as commands by the agent or downstream models.
Sanitization: No sanitization or escaping mechanisms are described for the text and URL parameters before they are interpolated into the shell command JSON structures.
[EXTERNAL_DOWNLOADS]: The skill requires the processing and transmission of remote URLs to external AI services.
The character creation phase requires passing a list of remote image URLs to the training service.
Video generation steps depend on supplying remote URLs for both base images and audio files to the avatar animation tools.
[COMMAND_EXECUTION]: The skill defines a complex sequence of shell commands using the infsh utility to orchestrate the video production pipeline. This involves running character generation, audio synthesis, and video merging tools directly from the agent's command line environment.

ai-podcast