The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes ffmpeg and playwright to perform media processing and frame-stepping for video rendering. These commands are executed locally and are necessary for the skill's primary function of generating high-quality video ads.
[PROMPT_INJECTION]: The skill ingests data from external sources such as ad templates and brand research, creating a surface for potential indirect prompt injection.
Ingestion points: Untrusted data enters the context through get_ad_template (library templates) and get_ad_brand (researched brand data).
Boundary markers: The data is handled using structured JSON schemas (recipe.card), reducing the risk of instructions being misinterpreted as commands.
Capability inventory: The skill has the ability to write local files (write_file) and execute subprocesses for video rendering.
Sanitization: The risk is mitigated by a mandatory human-in-the-loop approval gate in Phase 1.5, where the agent must present the drafted ad specification to the user for confirmation before any rendering or final processing occurs.
[SAFE]: The skill includes a pre-flight phase that checks for required system dependencies on the user's machine. If tools are missing, it provides instructions for the user to install them manually rather than attempting to perform unauthorized or silent installations.

remix-photo-grid-ad-from-sample