The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests untrusted data from Knowledge Planet topics and comments across multiple scenarios (e.g., daily-patrol.md, monitor-risky-content.md). This creates a surface for indirect prompt injection, where malicious instructions embedded in posts could influence the agent's behavior during summarization or script generation. \n
Ingestion points: Topic content and comments fetched via group +topics and get_topic_comments. \n
Boundary markers: Not explicitly defined in instructions to isolate external data. \n
Capability inventory: Includes posting content, deleting topics, and executing shell commands. \n
Sanitization: Lacks explicit sanitization of ingested content before processing. \n- [COMMAND_EXECUTION]: The skill makes extensive use of local shell commands for management and rendering. The record.js script invokes ffmpeg via execSync to encode video files, and the migration scenario involves file system operations like mv and ls. These capabilities are functional requirements but represent a sensitive command execution surface. \n- [EXTERNAL_DOWNLOADS]: The skill provides instructions to install the zsxq-cli tool globally and puppeteer locally. These downloads target official registries and the trusted author's own repositories, which are safe sources for the skill's intended purpose. \n- [SAFE_PRACTICE]: The skill incorporates robust safety guidelines, including strict prohibitions on exposing authentication tokens and mandatory user confirmation checkpoints for all write and delete operations.

zsxq