The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to install a third-party tool premiere-cli from the author's GitHub repository (https://github.com/stefanwebb/premiere-cli) using pipx or uv. It also executes a command to install a CEP panel and enable PlayerDebugMode within Adobe Premiere Pro.
[COMMAND_EXECUTION]: The skill relies extensively on executing shell commands via the premiere-cli binary to interact with Adobe Premiere Pro, manage sequences, and automate desktop actions.
[DYNAMIC_EXECUTION]: The inspect-dom-object command provides a mechanism for API discovery by evaluating property-path expressions (e.g., app.project.activeSequence) against the live Premiere Pro DOM. While it attempts to restrict method calls to maintain read-only status, it remains a dynamic evaluation vector.
[PRIVILEGE_ESCALATION]: The desktop-* suite of commands requires the user to grant macOS Accessibility permissions. This allows the skill to generate synthetic keyboard and mouse events (clicks, typing, screenshots), which are powerful capabilities that could be misused if the agent is directed to other applications.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a significant attack surface for indirect injection by ingesting and processing untrusted data from the Premiere Pro environment:
Ingestion points: Processes project metadata, sequence markers, XMP metadata, and external cuts files (--intervals-file).
Capability inventory: Extensive capabilities including file writing, subprocess execution, and UI automation (macOS Accessibility).
Sanitization: Large responses like XMP metadata and DOM inspections are truncated, but there is no explicit sanitization for instructions embedded within metadata fields.
Boundary markers: The skill lacks explicit instructions to ignore potentially malicious content found within project metadata or external files.

premiere-cli