The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute several system and local utilities to perform its functions:
Executes ffmpeg and ffprobe for video processing, conversion, and metadata verification.
Uses the macOS system utility screencapture for native window recording.
Runs a bundled Swift script (scripts/window-id.swift) to resolve window IDs for targeted capture.
Uses playwright (Node.js) to automate browser interactions and screenshots.
[EXTERNAL_DOWNLOADS]: The skill requires external dependencies to function:
Instructs the agent to install Playwright browser binaries via playwright install.
Recommends the use of static ffmpeg builds on CI runners or hosted environments to avoid package manager locks.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data by design:
It navigates to and captures arbitrary web pages, which could theoretically contain malicious content designed to influence the agent's behavior.
It extracts text captions directly from the DOM of captured pages to create subtitles.
However, the risk is mitigated by the fact that the ingested data is primarily rendered into images or video files, rather than being used to dynamically construct further executable commands.
[DATA_EXFILTRATION]: While the skill is designed to capture potentially sensitive interface data (including logged-in states), it includes robust security protocols:
Explicitly mandates the use of throwaway browser profiles (launchPersistentContext) for authenticated sessions.
Provides a double-nested try...finally logic to ensure profile directories are deleted from disk even if the process crashes or an error occurs.
Emphasizes verifying that the profile directory is actually gone before completing the task.

vp-recording