The Agent Skills Directory

[SAFE]: The skill defines a logical framework for managing parallel AI agent tasks. It includes phases for framing tasks, fanning out to multiple candidates, cross-judging results, and grafting the best components into a final artifact. No malicious patterns such as credential theft, remote code execution, or data exfiltration were found.- [INDIRECT_PROMPT_INJECTION]: The skill functions as an orchestrator for untrusted inputs (tasks) and intermediate outputs (candidate results). While it possesses the surface for indirect injection via these ingestion points (Phase A and Phase C), its capabilities are restricted to spawning subagents and writing to local file paths like .scratchpad or git worktrees. No specific sanitization or boundary markers are defined in the instructions, but the risk is consistent with standard agent operations.

arena