The Agent Skills Directory

[SAFE]: The skill performs self-critique based on predefined design dimensions (clarity, hierarchy, typography, motion, brand consistency). It writes results to a local critique.json file and emits signals for iteration control. No sensitive files are accessed, and no network operations are performed.\n- [METADATA_POISONING]: The open-design.json manifest lists a capability prompt:inject. While the term 'inject' can be a red flag, in this context it refers to the platform-specific mechanism for how the critique feedback is reintroduced into the agent's context for the next iteration of the development loop. It is not an adversarial prompt injection attack.\n- [COMMAND_EXECUTION]: The skill mentions the command od ui respond, which is part of the legitimate user interface interaction for the Open Design platform to allow users to manually break out of loops. This is a standard functional feature and not an unauthorized command execution risk.\n- [DATA_EXPOSURE]: The skill writes critique data to the project current working directory in a file named critique.json. This is intended behavior for providing feedback to the user and the system, and it does not expose sensitive user credentials or system information.

critique-theater