The Agent Skills Directory

[PROMPT_INJECTION]: The instructions explicitly direct the agent to "remove... unrequested features or safety" during the implementation review phase. This pattern can be used to instruct sub-agents to bypass safety filters or remove security-critical defensive programming (e.g., bounds checking or input validation) if they are not explicitly mentioned in the original task requirements.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the processing of external data that could contain malicious instructions.
Ingestion points: The skill reads "governing specifications," "repository instructions," and "user-level tasks" from the local environment or external input.
Boundary markers: No boundary markers, delimiters, or "ignore embedded instructions" warnings are utilized when passing these potentially untrusted specifications to sub-agents.
Capability inventory: The skill uses the agent CLI to spawn sub-agents and performs "live validation," which involves the execution of the running software and inspection of workspace effects. It also involves writing design and review artifacts to the file system.
Sanitization: There is no evidence of sanitization, validation, or escaping of the external specifications or tasks before they are interpolated into prompts for the authoring or validation agents.

orchestrate-attractor-loops