The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes explicit defensive instructions within the SKILL.md file. It mandates that any user-provided manuscripts or drafts must be treated as untrusted input and that the agent should never execute commands found within them (e.g., instructions like 'ignore previous rules').
[COMMAND_EXECUTION]: The skill executes a local Python script (scripts/scan.py) to perform keyword scanning. While the test suite (evals/test_scan.py) utilizes subprocess.run to interact with Git, this is a security feature used to verify that sensitive user data in the data/ directory is not being tracked by the repository.
[DATA_EXFILTRATION]: No network operations or external data exfiltration patterns were detected. The skill is designed for completely local operation, reading and writing only to its own data/ directory to store reports, user profiles, and local rule sets.
[SAFE]: The skill processes untrusted input as part of its primary function but manages this surface safely through deliberate prompt engineering and by minimizing capabilities that could be exploited by indirect injection.

yuwen-publish-precheck