Antigravity Prompts

Extracted system prompts from Antigravity (Google's VS Code fork with AI capabilities).

Overview

This repository contains system prompts and instructions that power Antigravity's AI agent. These prompts define:

• Agent Identity & Behavior: How the AI presents itself and interacts
• Workflow Modes: PLANNING, EXECUTION, and VERIFICATION phases
• Knowledge Management: Persistent memory via Knowledge Items (KI)
• Task Tracking: Structured task management with artifacts
• Tool Usage: Available tools and how to use them

Directory Structure

antigravity_prompts/
├── prompts/
│   ├── planning-mode.txt      # Full system prompt (agentic mode)
│   ├── Fast Prompt.txt        # Simplified system prompt
│   ├── modes.md               # PLANNING/EXECUTION/VERIFICATION modes
│   ├── knowledge_items.md     # Knowledge Items (KI) system
│   ├── task.md                # Task management with checklists
│   ├── implementation_plan.md # Implementation plan artifact format
│   └── walkthrough.md         # Walkthrough artifact format
├── CLAUDE.md                  # Project documentation
├── REVERSE_ENGINEERING.md     # Extraction methodology
└── README.md

Prompt Files

planning-mode.txt - Main System Prompt

The complete agentic system prompt (~600 lines) containing:

• Identity: "You are Antigravity, a powerful agentic AI coding assistant designed by the Google Deepmind team"
• User Information: Workspace paths, OS details, access restrictions
• Agentic Mode: Task boundary system for complex multi-step work
• Mode Descriptions: PLANNING, EXECUTION, VERIFICATION phases
• Artifact System: task.md, implementation_plan.md, walkthrough.md
• Tool Definitions: 25+ tools including browser_subagent, codebase_search, run_command, etc.
• Web Development Guidelines: Tech stack, design aesthetics, SEO best practices
• Workflows: Custom workflow system via .agent/workflows/*.md
• Communication Style: GitHub markdown formatting, proactiveness guidelines

Fast Prompt.txt - Simplified Prompt

A shorter version (~300 lines) with essential identity, workspace info, tools, and web development guidelines. Missing the full agentic mode features like task boundaries and persistent context.

modes.md - Agent Modes

Defines the three operational modes:

Mode	Purpose
PLANNING	Research, understand requirements, create implementation_plan.md, get user approval
EXECUTION	Implement the approved plan, return to PLANNING if encountering issues
VERIFICATION	Prove work is complete, capture evidence, create walkthrough.md

knowledge_items.md - Knowledge Items System

Instructions for using the persistent memory system:

• Mandatory First Step: Check KI summaries before ANY research
• KI Structure: Each KI has metadata.json and artifacts/ directory
• When to Use: Debugging, following patterns, complex implementations
• Key Principle: KIs are starting points, not ground truth - always verify

task.md - Task Management

Checklist-based task tracking:

- [ ] uncompleted task    <!-- id: 0 -->
- [/] in progress task    <!-- id: 1 -->
- [x] completed task      <!-- id: 2 -->

implementation_plan.md - Plan Artifact

Template for technical plans with:

• Goal description and context
• User review requirements
• Proposed changes grouped by component
• Verification plan (automated tests + manual verification)
• Confidence assessment (6-question framework)

walkthrough.md - Completion Artifact

Post-implementation summary with:

• Changes made
• Verification results
• Embedded screenshots/recordings

Key Features

Task Boundary System

For complex tasks, the agent uses task_boundary tool to:

• Set TaskName, TaskSummary, TaskStatus
• Track mode (PLANNING/EXECUTION/VERIFICATION)
• Update progress without overwhelming the user

Confidence Assessment

Before notifying users, agents must rate confidence (0.0-1.0) based on:

1. Gaps in addressing the request?
2. Unverified assumptions?
3. Complex logic with unknowns?
4. Non-trivial interaction risks?
5. Ambiguous requirements?
6. Irreversible changes?

Persistent Context

Two mechanisms for cross-conversation memory:

1. Conversation Logs: Raw history in .system_generated/logs
2. Knowledge Items: Distilled knowledge by KNOWLEDGE SUBAGENT

Available Tools

Category	Tools
Search	codebase_search, grep_search, find_by_name, search_in_file, search_web
Files	view_file, write_to_file, replace_file_content, multi_replace_file_content
Terminal	run_command, command_status, send_command_input, read_terminal
Browser	browser_subagent, read_url_content
Analysis	view_file_outline, view_code_item, list_dir
Media	generate_image

Source

Extracted from Antigravity 1.104.0 installation.

Disclaimer

This is a reverse-engineering effort for educational purposes. The extracted content may be incomplete or partially accurate.