The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides shell command templates for interacting with the grok CLI across different modes (Peer, Builder, Reviewer). These commands include flags for managing session state, reasoning effort, and tool permissions.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates an agent processing repository content through an external CLI tool, creating a surface for indirect prompt injection where instructions embedded in project files could influence the agent's behavior.
Ingestion points: The agent is instructed to read repository files and directory structures using tools such as read_file, grep, and list_dir as specified in the Peer and Reviewer mode configurations.
Boundary markers: The instructions recommend using isolated prompt files (--prompt-file) and specific environment variables (e.g., GROK_MEMORY=0) to disable session memory and external tool integrations during sensitive tasks.
Capability inventory: The skill documents different capability tiers, ranging from read-only access for reviews to full shell access in 'Builder' mode (utilizing --always-approve).
Sanitization: The skill emphasizes capability restriction through tool allowlists (--tools) and explicit deny rules (--deny) to prevent the agent from executing unauthorized commands based on the data it processes.

grok-agents