The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input from the user_message to generate shooting scripts and prompts for external media generation models.
Ingestion points: The initial user request is ingested in the intake_extract step and subsequently used in the script_draft and script_revised steps (via the user_message input).
Boundary markers: The skill employs xml_escape and truncate filters when interpolating the user message into step tasks.
Capability inventory: The workflow has the capability to write files (filesystem-write), execute system processes (process-control for FFmpeg/FFprobe), and trigger external API calls for media generation (external_paid_submit).
Sanitization: Input is sanitized using xml_escape. Crucially, the skill implements two mandatory human-in-the-loop review gates (review_gate and revision_confirm_gate) which require explicit user approval before any external media submission or billing occurs.
[COMMAND_EXECUTION]: The skill requires the presence of system binaries and executes them for video editing tasks.
Evidence: The metadata requires ffmpeg and ffprobe. Steps such as merge, subtitled_final, and delivery_audit (via composition skills like video-merger and subtitle-burner) utilize these binaries to process video assets locally.
[EXTERNAL_DOWNLOADS]: The skill interacts with external AI service providers to generate images and videos.
Evidence: The skill calls external providers through nano-banana-pro (Gemini models) and seedance-2-prompt (Seedance models) for image and video creation. These steps are marked as external_paid_submit and are gated by user consent.

meta-short-drama