Security audit
Security checks for vulnerabilities and agentic risk
This is a disclosed writing-revision skill with a local read-only checking script and an optional blind-review subagent, with no evidence of hidden persistence, exfiltration, or destructive behavior.
Before installing, expect this skill to work best on Chinese or English argumentative/expository prose and to run a local checking script plus possibly a separate blind-review subagent. Avoid using it for sensitive text unless you are comfortable with that review flow; use a simpler rewrite path if you do not want delegation.
The README says the skill rewrites already-written Chinese or English argumentative prose, which constrains supported language/locale behavior in natural language. There is no opt-in choice or explicit justification for excluding other languages, so this appears to impose a language policy without user selection.
The document contains prescriptive instructions centered on Chinese-language wording and forbids specific Chinese phrases, effectively constraining output language/register behavior. There is no visible opt-in, user choice, or stated region-specific justification in this file, so this appears to impose a locale/language policy unilaterally.
The text uses Chinese examples and prescriptive statements such as banning certain Chinese phrases and later requiring 干练简洁, which amounts to a natural-language style/locale constraint. The file does not indicate that this language expectation is optional, user-selected, or tied to a justified region-specific context.
The rubric hard-codes Chinese connective examples ("所以/也就是/这意味着/于是") as a required tell, which imposes a language-specific evaluation frame. The file also repeatedly uses Chinese-only examples elsewhere, but does not offer any language choice or explain that the skill is limited to Chinese texts, creating a natural-language locale policy concern.
The README states that the skill may execute a local script and invoke an independent blind-review sub-agent, but it does not clearly surface those operational behaviors as user-visible side effects requiring consent at time of use. In an agent environment, undisclosed tool execution or delegation can violate user expectations, create unnecessary data exposure to another model instance, and trigger actions the user did not explicitly authorize.
No suspicious patterns detected.
