critique
Pass
Audited by Gen Agent Trust Hub on Aug 14, 2026
Risk Level: SAFEPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The instructions utilize strong directive language such as 'CRITICAL' and 'most important check' to emphasize specific evaluation criteria (e.g., AI slop detection). These are natural instructional markers for a critique persona and do not constitute an attempt to bypass agent safety filters or extract system prompts.
- [DATA_EXFILTRATION]: The skill does not contain any code or instructions for network operations, external data transmission, or access to sensitive local files (e.g., credentials, SSH keys).
- [COMMAND_EXECUTION]: While the skill suggests follow-up commands like /animate or /polish, these are references to other internal agent skills meant for design refinement. There is no evidence of arbitrary shell command execution or unauthorized system access.
- [INDIRECT_PROMPT_INJECTION]: The skill acts as a processor for external data (user designs and the 'area' argument).
- Ingestion points: The 'area' argument and the UI design context provided by the user.
- Boundary markers: Explicit boundary markers or 'ignore' instructions for embedded data are absent.
- Capability inventory: The skill can trigger other internal design-focused skills (/animate, /quieter, /optimize, /adapt, /clarify, /distill, /delight, /onboard, /normalize, /audit, /harden, /polish, /extract, /bolder, /critique, /colorize).
- Sanitization: No explicit sanitization or filtering of input data is defined.
- Note: This represents a standard surface for a tool-use skill and does not indicate malicious intent.
Audit Metadata