The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions to skip the creation of an implementation plan (implementation_plan.md) and immediately execute file operations. This overrides standard safety protocols intended to allow for human review before the agent performs significant environment modifications.\n- [PROMPT_INJECTION]: The skill defines a surface for indirect prompt injection by ingesting unstructured user requirements to generate system prompts and operational logic for sub-agents.\n
Ingestion points: User-provided domain requests and project-specific descriptions.\n
Boundary markers: None specified within the templates provided for generated SKILL.md or agent.json files to separate data from instructions.\n
Capability inventory: Includes write_to_file, run_command, and define_subagent to deploy and register generated agent personas.\n
Sanitization: No procedures for sanitizing or escaping user input before it is interpolated into agent instructions are provided.\n- [COMMAND_EXECUTION]: The skill utilizes run_command and define_subagent to manage the agent lifecycle and register new personas within the Antigravity CLI environment. These capabilities are used for the skill's primary purpose but involve modifying the runtime configuration.\n- [SAFE]: The skill includes a requirement to store all intermediate artifacts, code, and analysis data in a dedicated _workspace/ folder. This follows isolation best practices to keep the main project directory clean and prevent accidental modification of production code.

harness