The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local Python and Bash scripts (rules_stamp.py, last-article.sh) to locate the most recent article in a Hugo repository, calculate cryptographic hashes (SHA-1) of rubric files, and record review results. It also interacts with sibling vendor skills such as write-status and write-article for integrated workflow management.
[INDIRECT_PROMPT_INJECTION]: The skill processes external content—including files, pasted text, and data from URLs—to perform its primary review function. This introduces an attack surface where instructions embedded in the target article could attempt to manipulate the review output or scoring.
Ingestion points: Target articles are provided via the [article] parameter, which supports paths, text, or URLs.
Boundary markers: The skill contains explicit instructions to maintain a read-only posture ("A review never edits") and prohibits the invention of facts, which provides a level of protection against adversarial instruction following.
Capability inventory: File system access for status tracking, execution of local scripts, and network access for URL fetching.
Sanitization: Invokes a specialized ai-sanitize skill to detect AI patterns, although it does not provide explicit filtering for prompt injection payloads.
[EXTERNAL_DOWNLOADS]: The skill can fetch content from user-provided URLs to evaluate web-based articles.
[OBFUSCATION]: Documentation examples in references/writing-style.md contain zero-width space characters (U+200B) within Hugo template tags (e.g., {{< ref ... }}). Analysis confirms this is a benign escaping technique used to prevent the template from being rendered or executed during prompt processing, rather than a malicious attempt to hide executable code.

write-review