The Agent Skills Directory

[SAFE]: The skill performs its stated function of reviewing literary content. It handles project files such as story drafts, outlines, and character settings, which is appropriate for its use case.
[COMMAND_EXECUTION]: The skill invokes local Node.js scripts (scripts/normalize-punctuation.js and scripts/check-ai-patterns.js) using the node command. These scripts are bundled with the skill and are used to detect formatting issues and specific linguistic patterns in the story text. They do not access network resources or sensitive system directories (like .ssh or .aws).
[REMOTE_CODE_EXECUTION]: No remote code execution or piped shell commands (e.g., curl | bash) were detected. All executable content is local to the skill package.
[PROMPT_INJECTION]: The instructions provide a structured framework for the AI to act as a 'review coordinator.' While it includes 'iron rules' (铁律), these are task-specific constraints to ensure the agent focuses on identifying problems rather than validating correctness, and do not constitute a safety bypass.
[DATA_EXFILTRATION]: There is no evidence of data exfiltration. The skill reads local project files to perform its analysis, but does not include any network-enabled tools or commands to send this data to external servers.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted story content provided by the user. While this content is passed to sub-agents, the primary instructions focus on critical analysis and error detection using defined rubrics, which mitigates the risk of the agent following instructions embedded within the processed text. This is classified as an expected operational surface for a review tool.

story-review