The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted customer data (reviews and comments) from the inputs/ directory to 'ground' its generation. This constitutes an indirect prompt injection surface as these external files could contain malicious instructions.
Ingestion points: inputs/reviews/ and inputs/comments/ files.
Boundary markers: Absent; the skill encourages verbatim use of customer language to ensure authenticity.
Capability inventory: The skill produces file outputs to the outputs/ directory and provides examples for making network API calls and executing shell commands.
Sanitization: Not explicitly mentioned in the workflow.
[COMMAND_EXECUTION]: The documentation includes various shell command examples for interacting with advertising platform APIs and processing media files using ffmpeg and local scripts.
Evidence: node tools/clis/google-ads.js and ffmpeg -i video.mp4 -i voiceover.mp3 commands provided in SKILL.md and references/generative-tools.md.
[EXTERNAL_DOWNLOADS]: The skill references several external repositories and packages for specialized functionality such as voice cloning and video rendering.
Evidence: Instructions to git clone https://github.com/jamiepine/voicebox.git and use npx create-video for Remotion projects.
[DYNAMIC_EXECUTION]: The skill utilizes code-based video frameworks like Remotion and Hyperframes, which involve rendering video content from programmatically generated React or HTML/JavaScript code.
Evidence: Detailed guidance on using code-based video for templated, brand-consistent variations in references/generative-tools.md.

ad-creative