The Agent Skills Directory

[PROMPT_INJECTION]: The skill was evaluated for indirect prompt injection risks as it ingests user input to modify agent guidelines. Ingestion points: User conversation history matching rule-detection regexes in SKILL.md. Boundary markers: The skill uses markdown blockquotes and headers to encapsulate user input. Capability inventory: Persistence is achieved via file-write operations to the .agents/SYSTEM/ directory. Sanitization: A mandatory human-in-the-loop confirmation step ensures users must explicitly approve captured rules.
[DATA_EXFILTRATION]: No network operations or external data transfer mechanisms were detected. The skill only writes to local configuration files.
[REMOTE_CODE_EXECUTION]: No remote scripts, package installations, or dynamic code execution patterns were found.

rules-capture