The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides instructions to install and execute third-party packages from the NPM registry using npx. Specifically, it references gooseworks (in references/imessage-video-ads.md) and create-video (in references/generative-tools.md).- [COMMAND_EXECUTION]: The documentation includes various examples of shell commands for the agent to run, such as curl for interacting with external APIs (Gemini, ElevenLabs), ffmpeg for video processing, and git for cloning repositories.- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data. Ingestion points: Untrusted content is ingested from inputs/reviews/ and inputs/comments/. Boundary markers: The instructions lack explicit prompt delimiters to separate this untrusted content from the agent's instructions. Capability inventory: The agent has the ability to execute shell commands, interact with advertising platform APIs, and generate code files. Sanitization: The documentation recommends basic character escaping (< to \u003c) when generating HTML artifacts, but no thorough sanitization is defined for the raw data ingested.- [DYNAMIC_EXECUTION]: The skill involves generating React components for the Remotion framework and creating self-contained HTML review pages containing embedded JSON data blocks populated at runtime.- [EXTERNAL_DOWNLOADS]: The skill fetches audio assets from bigsoundbank.com and clones the voicebox repository from GitHub. While these are documented neutrally, they involve external network requests and resource retrieval.

ad-creative