The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes instructions to detect and flag adversarial directives like 'ignore previous instructions' within user-provided role definitions. While these phrases triggered static analysis hints, they are part of a defensive hardening strategy designed to protect the agent rather than a malicious attempt to bypass safety guidelines.\n- [NO_CODE]: The skill consists entirely of instructional text and design patterns. It contains no scripts (Python, JavaScript, etc.) or binary components.\n- [COMMAND_EXECUTION]: The skill is restricted to read-only tool access (Read, Grep, Glob) and does not request tools for shell command execution or write access, minimizing the potential impact of adversarial input.\n- [DATA_EXFILTRATION]: No network tools or external communication capabilities are requested, and the skill does not interact with remote domains.\n- [SAFE]: The skill incorporates an explicit 'Injection Hardening' boundary and a 'Prompt Injection Safety Notice,' emphasizing that it provides architectural recommendations only and treating all processed artifacts as untrusted input.

rbac-design