The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill analyzes external performance data (CSV, manual pastes, or API outputs) to iterate on ad copy. Evidence: Ingestion points are defined in SKILL.md for metrics and existing copy; boundary markers to isolate data from instructions are absent; the skill possesses capabilities like network and command execution; no sanitization of the input data is mentioned. This surface could be exploited if an attacker provides poisoned performance metrics containing hidden instructions.\n- [EXTERNAL_DOWNLOADS]: The skill references the installation of third-party tools from external sources, specifically cloning the 'Voicebox' repository from GitHub and using 'npx' to run remote video generation packages. These actions execute code from repositories outside of the immediate platform control.\n- [COMMAND_EXECUTION]: The skill provides examples of shell commands for retrieving ad performance data via local Node.js scripts, interacting with generative AI APIs through curl, and processing media using ffmpeg.\n- [DYNAMIC_EXECUTION]: The skill integrates with the Remotion framework, which renders video assets using React code at runtime. This involves programmatic assembly and execution of code to generate final production assets based on dynamic data feeds.

ad-creative