AntiGravity Token & Budget Optimizations

This repository contains optimized templates, configuration files, and automation scripts designed for the AntiGravity AI coding assistant (and similar LLM-based agent ecosystems).

The primary goal of these optimizations is to minimize prompt/context size (input tokens) and limit verbosity in responses (output tokens) to conserve your plan usage cap and budget without losing functional capability.

Project Structure

.
├── scripts/
│   ├── compress_all_skills.py   # Script to compress science/tool SKILL.md files
│   └── compress_mcp_schemas.py  # Script to recursively compress MCP JSON schemas
├── templates/
│   ├── rules/
│   │   └── AGENTS.md            # Highly optimized global system rules
│   └── skills/                  # Hand-crafted global pipeline skills
│       ├── autonomous-dev-planning/SKILL.md
│       ├── autonomous-dev-implementation/SKILL.md
│       ├── autonomous-dev-review/SKILL.md
│       ├── document-driven-design/SKILL.md
│       ├── document-driven-design-only/SKILL.md
│       ├── git-check-commit-push/SKILL.md
│       ├── workspace-onboarding-planning/SKILL.md
│       ├── workspace-onboarding-development/SKILL.md
│       └── workspace-onboarding-review/SKILL.md

Optimization Concepts & Intentions

1. Context Size Minimization

By default, AntiGravity loads the contents of active system rules (AGENTS.md), triggered skill instructions (SKILL.md), and registered MCP tool schemas (*.json) into the model's prompt on every turn. In large projects, this startup overhead can consume 50,000+ tokens per message.

• Pruning Verbal Fluff: Removing lengthy background explanations, common mistakes lists, and verbose workflows from skills.
• Structural Filtering: Eliminating empty headings and sections that do not contain actionable commands or parameters.
• Rules Condensation: Shortening system rules (Node.js worker thread management, PDF.js ingestion invariants, PowerShell constraints) to direct imperatives.

2. Output Verbosity Controls

Generating verbose text is the most expensive part of LLM usage (in terms of time and API budget).

• Strict Constraints: Adding explicit limits (e.g., maximum line limits for plans, bullet point constraints, code-only outputs, removing conversational padding) directly in the skills' instructions.
• Narration-Free Execution: Enforcing that code implementation phases output pure code/tests with zero explanations.

3. Minification of JSON Schemas

Tool definitions are loaded as JSON structures.

• Minifying schema JSON files (removing pretty-print spacing and indentation).
• Compressing description and parameter description texts while preserving schemas (like GraphQL schema syntax) to keep tools fully functional.

Metrics & Combined Savings

• Skill Files: ~164,000 characters saved (49.3% size reduction).
• MCP Schemas & Rules: ~18,500 characters saved (10.5% size reduction).
• Total Savings: ~183,000 characters (equivalent to ~45,000 tokens saved on every single message turn).

Installation & Usage

Prerequisites

• Python 3.8+

Step 1: Apply Hand-Crafted Global Skills & Rules

Copy the optimized global templates directly into your global configuration directory:

• Copy templates/rules/AGENTS.md to ~/.gemini/config/AGENTS.md.
• Copy templates/skills/* folders to ~/.gemini/config/skills/.

Step 2: Compress Dynamic Tool Skills

Run the script to automatically parse and compress the science and Android-specific tool skills:

python scripts/compress_all_skills.py --execute

Step 3: Compress MCP Tool JSON Schemas

Run the schema minification script to compress descriptions and remove spacing:

python scripts/compress_mcp_schemas.py --execute