The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill automatically reads and processes tasks from a local Markdown file which serves as a task queue. This creates a surface where external data (tasks written to the file by other processes, plugins, or users) can influence the agent's routing and execution logic.
Ingestion points: The file 50-系统/40-自动化/AI任务队列.md is read by the /ray command to identify pending collaborative tasks.
Boundary markers: The instructions do not specify any delimiters or safety headers within the task queue file to prevent the agent from interpreting task descriptions as system-level instructions.
Capability inventory: The skill has the ability to trigger a chain of other skills (ray-writer, ray-wechat, ray-x-article, etc.) which possess capabilities for file modification, network communication, and platform-specific API interactions.
Sanitization: There is no evidence of input validation or sanitization for the task strings read from the queue before they are passed to the routing logic.
[DYNAMIC_EXECUTION]: The skill documentation instructs the agent to execute specific local Python scripts to handle the archival of published content and to validate oral scripts.
Evidence: references/content-pipeline.md specifies the command python3 record_published.py --article <正式稿路径> --platform x|wechat --url <公开地址> --published-at <带时区时间>.
Evidence: references/content-pipeline.md references the execution of oral_script_check.py to verify the consistency and length of generated scripts.

ray