The Agent Skills Directory

[SAFE]: The skill implements a robust framework for multi-agent reviews, utilizing parallel execution via the run_in_background parameter. The workflow is logically structured and requires user confirmation before dispatching agents.
[PROMPT_INJECTION]: The skill includes protection against sub-agent drift by explicitly instructing reviewer personas to "Do NOT write any code — only research and report findings." While the skill ingests untrusted data from project files and browser content, the risk is mitigated by its specific focus and the non-executable nature of the reviewer prompts.
Ingestion points: Project source code at [PATH] and live site content via browser MCP tools.
Boundary markers: The prompt template provides structural headers but does not use explicit delimiters to separate instructions from analyzed content.
Capability inventory: Sub-agents are restricted to reporting and research, preventing them from making unauthorized modifications to the codebase. The parent agent handles consolidation and planning.
Sanitization: No automated sanitization is described, as the skill relies on the LLM's inherent ability to distinguish between code content and operational instructions.
[DATA_EXFILTRATION]: No unauthorized data exfiltration patterns were detected. The skill accesses local files and potentially local dev servers for the purpose of auditing, which is consistent with its stated purpose.

experts