The Agent Skills Directory

[COMMAND_EXECUTION]: The skill extensively uses 'subprocess.run' to invoke system binaries including 'ffmpeg', 'ffprobe', and 'rsvg-convert' across multiple scripts ('composite_variants.py', 'music_and_mux.py', and 'render_overlays.py'). While command arguments are passed as lists to mitigate shell injection, the complexity of the generated 'filter_complex' strings for FFmpeg poses a risk if input variables are not strictly validated.
[COMMAND_EXECUTION]: Scripts 'music_and_mux.py' and 'strip_product_backgrounds.py' perform dynamic library loading by modifying 'sys.path' to include a computed path ('../../../../skills/atoms/_shared'). This allows the skill to import and execute code from a location outside the controlled skill directory, which could be exploited if the surrounding environment is compromised.
[EXTERNAL_DOWNLOADS]: The skill interacts with Fal.ai services to generate music and remove product backgrounds. It uses a 'download' function from the dynamically loaded 'fal_helpers' to fetch media assets from remote URLs ('fal.media') into the local 'assets' folder.
[PROMPT_INJECTION]: The skill is designed to process untrusted external data (scraped product images and background videos) and uses configuration from an external database to drive the assembly logic. This creates an attack surface for indirect prompt injection where malicious instructions could potentially influence the behavior of the agent during the multi-step rendering process.

render-vignette