The Agent Skills Directory

User-Provided Input Handling: The skill incorporates a $ARGUMENTS placeholder to allow users to specify a focus for the review. While this is a standard feature for user interaction, it represents a potential surface for prompt injection if a user provides malicious instructions instead of a focus hint.- Context Processing and Ingestion: The skill analyzes the current conversation to generate a 'Review Brief.' This involves ingesting untrusted data from the chat history. To mitigate risks, the skill includes explicit instructions to distill this information neutrally and to separate it from the original chat context before further processing.- Tool Orchestration and Delegation: The skill uses a delegate tool to spawn sub-sessions for different review personas. Notably, it specifies context_depth="none" for these sub-sessions, ensuring that each persona agent only sees the distilled brief and not the full conversation history. This isolation is a positive security practice that limits information exposure and prevents cross-context prompt injection.

council-here