The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted customer data from inputs/reviews/ and inputs/comments/. The instructions require the agent to use this language verbatim, which creates a potential surface for indirect prompt injection. 1. Ingestion points: External text files in the inputs/ directory and marketing context files. 2. Boundary markers: None identified for the ingested text. 3. Capability inventory: Includes curl for network requests, npx for package execution, and ffmpeg for video processing. 4. Sanitization: No prompt injection sanitization is defined for the ingested text, although technical escaping is used for the HTML output template.
[COMMAND_EXECUTION]: The skill provides instructions for the agent to execute shell commands using tools like ffmpeg for video assembly and npx for rendering video via Remotion. Evidence is found in references/imessage-video-ads.md and references/generative-tools.md.
[DYNAMIC_EXECUTION]: The skill uses a data-driven HTML template (assets/creative-review-template.html) and leverages React-based video generation through Remotion. Evidence is in references/creative-review-page.md and references/generative-tools.md.
[EXTERNAL_DOWNLOADS]: The skill includes references to external assets, such as audio files from bigsoundbank.com, for use in video production. Evidence is in references/imessage-video-ads.md.

suede-ad-creative