The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external manifest files and media content to drive the assembly pipeline, creating a potential surface for indirect injection.
Ingestion points: Content manifest (spots.manifest.json), graphics configuration (spots.graphics.json), and generated video/audio plates.
Capability inventory: The pipeline utilizes ffmpeg, python3, and npx for media processing, assembly, and rendering.
Sanitization: The skill includes explicit instructions to validate slugs, paths, and copy text before they are used in shell commands or filter_complex strings.
Boundary markers: The design recommends using semantic gates with falsifiable assertions and human-in-the-loop review to verify depicted content.
[COMMAND_EXECUTION]: The workflow involves dynamic assembly of ffmpeg command strings and execution of external tools like npx remotion and python3 scripts.
Evidence: Sections 4 and 5 of SKILL.md describe using ffmpeg filters and python3 for timing derivation and assembly.
Mitigation: The skill emphasizes that inputs must be validated before reaching shell commands to prevent injection.
[DYNAMIC_EXECUTION]: The repository provides a reference design for an agent to implement and execute custom Python and Node.js modules for video processing.
Evidence: references/pipeline.md defines interfaces for modules like spotkit.py and export_timing.py to be built and run by the agent.
[EXTERNAL_DOWNLOADS]: The skill references documentation and licensing from the vendor's official domain and well-known services.
Evidence: Mentions docs.bfl.ai and api.bfl.ai (vendor resources) and remotion.dev (a well-known video framework).

flux-3-product-ads