The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill implements strict policies against the exposure of authentication tokens, stating they must never be outputted to the terminal or shared. Sensitive operations like WeChat payments are managed through a secure 'Skill Pay' flow that requires a specific host environment (WorkBuddy) and explicit user authorization for each transaction. File access is limited to the skill's own operational directories (e.g., temporary folders for video rendering).
[COMMAND_EXECUTION]: The video generation scenario uses a local script (record.js) that invokes ffmpeg via execSync. The execution is safely managed: the script searches for ffmpeg in a pre-defined list of trusted system paths, and the command arguments are structured based on internal logic. There is no evidence of arbitrary or untrusted command execution.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external content from Knowledge Planet, such as posts and comments. It includes detailed instructions and 'Refinement Rules' to ensure the agent summarizes data accurately while ignoring any potential instructions embedded in the source text. Safety is further enforced by mandatory human-in-the-loop checkpoints for every public action (posting, replying, or modifying settings).
[EXTERNAL_DOWNLOADS]: The documentation provides instructions for installing legitimate dependencies, such as zsxq-cli and puppeteer, which are standard for the skill's functionality. It does not perform unauthorized or silent downloads from unknown sources.
[DYNAMIC_EXECUTION]: The skill uses local scripts to generate HTML templates and then render them into videos or images using browser-based tools (Puppeteer). This dynamic content generation is restricted to the intended business logic (creating reports and media) and is fully transparent to the user.

zsxq