The Agent Skills Directory

[SAFE]: Detailed analysis of the skill's instructions and 34 associated scripts revealed no malicious patterns. The skill adheres to a principle of least privilege, specifically forbidding the use of vision models to watch video files directly, and instead relies on metadata generated by separate tools.
[DATA_EXFILTRATION]: The skill includes a dedicated script (private_text.py) and a "privacy pass" stage designed to prevent data exfiltration. This utility uses comprehensive regular expressions to identify and mask sensitive information, including PII, credit card numbers, and credentials (e.g., OpenAI keys, GitHub tokens, AWS IDs) found in screen recordings or transcripts before any drafts are presented to the user.
[COMMAND_EXECUTION]: The skill uses local Python 3 scripts to perform arithmetic and scene building. These scripts are bundled with the skill and are executed with agent-generated configuration files. There is no evidence of arbitrary command execution or the use of unsanitized external input in shell commands.
[INDIRECT_PROMPT_INJECTION]: While the skill processes untrusted data (screen recordings and transcripts), it implements significant mitigations. The instructions explicitly warn against inventing content not present in the recording and provide specific rules for redrawing UI components as motion design rather than interpreting them as instructions. The surface is well-isolated through the use of intermediate tools and local scripts.

make-screen-videos