The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill utilizes scripts/bootstrap.py and scripts/desktop.py to download the core editing engine and the desktop application from GitHub releases (targeting the full-aigc-skills and storytold repositories). These downloads are secured via hardcoded SHA256 checksums, and the application is further verified using the system's codesign utility before use.
[COMMAND_EXECUTION]: The skill frequently invokes subprocesses to run the filmcraft-cli binary and various macOS system tools like hdiutil and lipo. These executions are performed through a security wrapper that uses sandbox-exec to enforce a profile that limits file system and network access, providing a high degree of process isolation.
[DYNAMIC_EXECUTION]: Local Python modules within the skill directory are loaded dynamically using importlib.util. This is used for modularizing the permission system, command registry, and session management logic without resorting to unsafe execution of external code.
[INDIRECT_PROMPT_INJECTION]: While the skill processes user-supplied JSON plans and external media metadata (a potential attack surface), it mitigates risk through comprehensive schema validation, strict parameter filtering, and by running the editing engine in a restricted sandbox environment that drops network capabilities and limits disk writes.
[DATA_EXFILTRATION]: The skill implements a whitelist for environment variables (PATH, HOME, TMPDIR, etc.), ensuring that sensitive credentials like API keys present in the host's environment are not passed to the child processes.

filmcraft-cli-subtitles