5 tips for using Antigravity 2.0 on enterprise codebases, planning phase

Key Concepts

• Antigravity 2.0: An AI-powered coding agent designed for complex enterprise environments.
• "Ducky": The persona adopted for the AI agent to frame the user-AI relationship as a mentor-intern dynamic.
• Repo Walls: The boundaries between different code repositories (e.g., front-end, back-end, shared packages).
• Hierarchy of Rules: A structured approach to defining coding standards and project context.
• Blast Radius: The potential scope of damage caused by AI-driven errors or unauthorized actions.
• Sandbox Mode: An OS-level containment feature to restrict destructive shell operations and network calls.
• "/grill me": A command that prompts the AI to interrogate the user to clarify requirements and architectural decisions.

---

1. Breaking Repo Walls

Enterprise software rarely exists in a single repository. Antigravity 2.0 allows users to aggregate multiple repositories (front-end, back-end, shared packages) into a single project. This enables the AI to perform cross-repo refactoring, such as updating a back-end data model and simultaneously synchronizing the corresponding TypeScript interfaces in the front-end.

2. The Hierarchy of Rules

To ensure code consistency, the agent follows a three-tier rule structure:

• Global Rules: Located in ~/.gemini/gemini.md. Used for universal standards (e.g., 80-character line limits).
• Project Rules: Located in .agents/rules within the project root. Used for application-specific coding standards.
• Local Context: readme.md files placed in specific directories. These provide "hyper-local" instructions, such as how to execute integration tests for a specific module.

3. Containing the Blast Radius

Security is paramount when using AI agents on enterprise code. The speaker recommends:

• Custom Security Presets: Always require manual review before the agent executes terminal commands. Users can build an "allow list" of common commands over time.
• Sandbox Mode: A preview feature that acts as an OS-level containment ring to prevent the agent from running wild with destructive shell operations or unauthorized network calls.
• Cloud Infrastructure Isolation: Ensure the development machine’s credentials do not have access to production environments, preventing accidental exposure or modification of live customer data.

4. Voice-First Planning

Instead of typing prompts, which can be restrictive and lead to missing context, the speaker uses native voice input for initial planning.

• Methodology: Treat the AI like a principal engineer. Talk through legacy quirks and messy architectural problems out loud.
• Benefit: This allows for a back-and-forth dialogue that captures the nuance of the problem, which is significantly more effective than writing a long, static prompt.

5. The "/grill me" Command

This is a critical step for requirement gathering. After providing a voice brief, the user executes the /grill me command.

• Function: The AI analyzes the codebase and the user's prompt, then generates 5–6 sharp, probing questions.
• Purpose: This forces the user to clarify fuzzy requirements or justify architectural decisions before the agent begins writing code, preventing the need for costly refactoring later.

---

Synthesis and Conclusion

The core philosophy presented is that AI agents in an enterprise setting should be treated as digital interns rather than "magic" text boxes. By establishing a clear hierarchy of rules, enforcing strict security guardrails (sandbox mode and cloud isolation), and utilizing iterative communication methods like voice planning and the /grill me interrogation, developers can effectively integrate AI into complex, living systems. The ultimate goal is to provide the agent with enough context and constraints so that it can operate safely and accurately within a professional codebase.