The Agent Skills Directory

[DYNAMIC_CONTEXT_INJECTION]: The skill uses the !command`` syntax to execute echo ${CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS:-disabled} at load time. This shell command is used solely to verify the status of a specific environment variable for feature detection and does not exfiltrate data or execute unsafe user input.
[INDIRECT_PROMPT_INJECTION]: The skill processes discovery briefs and reads arbitrary codebase files, creating a surface for indirect prompt injection. The skill implements substantial mitigations including an 'information barrier' where sub-agents are isolated from intent-heavy context, and a mandatory 'Bias Validation Gate' for all dispatched tasks. (Ingestion points: Discovery briefs and codebase files accessed during research waves. Boundary markers: Neutral task descriptions and intent-free spawn prompts. Capability inventory: Spawning sub-agents with the Agent tool, managing tasks, and compiling reports. Sanitization: Formal 'Question Generation Rules' and a 'Bias Validation Gate' to neutralize instructions before sub-agent ingestion.)

research