The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes video transcripts, which are untrusted external data sources, to drive automated video editing decisions and timeline modifications.
Ingestion points: Transcripts are read via read_script and from files in the library/ directory, as described in SKILL.md and references/other-a-roll-editing-scenarios.md.
Boundary markers: Edits are staged in a timeline.md file using specific strikethrough and Unicode arrow (→) syntax to delimit changes before application.
Capability inventory: The skill utilizes apply_script to commit timeline changes, create_motion_graphic_from_code to generate JSX assets, and manage_timelines to duplicate or create project structures.
Sanitization: The skill enforces structural rules (using [sN] segment IDs) and warns against modifying raw ASR or unstruck words directly, though it relies on the model's interpretation of the transcript content.
[DYNAMIC_EXECUTION]: The skill provides instructions for generating and editing JSX code at runtime for Motion Graphics components using the create_motion_graphic_from_code and edit_asset tools.
Evidence: The reference file references/motion-graphics.md contains detailed guidance on authoring JSX code for the active Motion Graphics workflow, including the assembly of properties and visual structures.
[EXTERNAL_DOWNLOADS]: The skill facilitates the acquisition of stock media and the use of external generation models.
Evidence: references/b-roll.md describes a workflow for using search_stock_media to download results into the sandbox and importing them into the project. It also references the video-gen skill and DeepFilterNet3 for voice isolation.

talking-head-guide