# Antigravity CLI Implementation

> Antigravity CLI (Gemini) optimization strategies including GEMINI.md separation, Implementation Plan artifacts, and hook-based guardrails

- **Category**: optimizing-token-usage-practices

- **Canonical URL**: https://designpattern.fyi/optimizing-token-usage-practices/part-iii-antigravity-cli/

---

## Description
Antigravity CLI (Gemini) optimization strategies including GEMINI.md separation, Implementation Plan artifacts, and hook-based guardrails








## Additional Notes

## Antigravity CLI Token Optimization

### Layer 1 — Platform Mechanics

Keep `GEMINI.md` to Antigravity-only overrides; put shared, cross-tool rules in `~/.gemini/AGENTS.md` to sidestep the Gemini CLI path conflict. Skills in `~/.gemini/config/skills/` follow the same load-on-match principle as Claude/Codex. Route lightweight or parallel subagents (e.g. a browser-testing subagent) to a faster model tier via their YAML `model` field, reserving your primary tier for the main coding loop.

**Key Strategies:**
- **File Separation**: Separate Antigravity-specific from shared rules
  - `GEMINI.md`: Antigravity-only overrides (keep thin)
  - `~/.gemini/AGENTS.md`: Shared, cross-tool rules
  - Avoids Gemini CLI path conflict
- **Skills Progressive Disclosure**: Skills in `~/.gemini/config/skills/`
  - Load-on-match principle like Claude/Codex
  - Reduce always-loaded token overhead
- **Tiered Subagent Routing**: Route subagents by workload type
  - Lightweight/parallel subagents: Faster model tier
  - Main coding loop: Primary model tier
  - Configure via YAML `model` field in agent definitions

### Layer 2 — Agent Operation

Antigravity's Implementation Plan artifact is a built-in plan gate — explicitly ask for a plan first on non-trivial tasks so exploration and editing don't happen uncontrolled in the same pass. Use `PreToolUse`-style hooks to hard-block expensive or risky operations (e.g. a forced push) rather than relying on the model to self-police.

**Operational Best Practices:**
- **Implementation Plan Artifact**: Built-in plan gate
  - Explicitly request plan first on non-trivial tasks
  - Separate exploration from editing
  - Prevent uncontrolled mixed-mode operations
- **Hook-Based Guardrails**: Use `PreToolUse`-style hooks
  - Hard-block expensive operations (e.g., force push)
  - Don't rely on model self-policing
  - Enforce operational constraints at tool level
- **Session Management**: Clear separation between planning and execution phases

### Layer 3 — Codebase Architecture

Duplicate the Conventions section into `AGENTS.md` (shared) or `GEMINI.md` only if Antigravity-specific phrasing is genuinely needed.

**Architecture Rules:**
- **Shared Conventions**: Use `AGENTS.md` for cross-tool compatibility
- **Tool-Specific Overrides**: Use `GEMINI.md` only when Antigravity-specific phrasing is needed
- **Avoid Duplication**: Don't duplicate content between files unnecessarily

## Antigravity CLI-Specific Features

### Implementation Plan Artifact
Antigravity's native Implementation Plan provides structured planning:
- Built-in plan gate separate from execution
- Structured approach to complex tasks
- Natural separation of exploration and implementation

### YAML Agent Definitions
Subagent definitions in `.antigravity/agents/` or `~/.config/antigravity/agents/`:
- YAML-based agent configuration
- Per-agent model selection
- Support for complex agent orchestration

### Hook System
Antigravity's hook system allows operational control:
- `PreToolUse` hooks for blocking operations
- `PostToolUse` hooks for logging and auditing
- Hard enforcement of operational constraints

### Skills Integration
Skills in `~/.gemini/config/skills/`:
- Follow same progressive disclosure pattern as other tools
- Load-on-demand based on task matching
- Reduce always-loaded context overhead

### Path Conflict Resolution
Critical note for users with both Gemini CLI and Antigravity CLI:
- Both write to `~/.gemini/GEMINI.md`
- Can silently pollute context with wrong tool's rules
- Deliberate separation required for clean operation