Google Antigravity Fast Prompt System

Relevant source files

• Google/Antigravity/Fast Prompt.txt

Purpose and Scope

This document describes Google Antigravity, an agentic AI coding assistant developed by Google Deepmind's Advanced Agentic Coding team. Antigravity features a sophisticated dual-context architecture combining raw Conversation Logs with distilled Knowledge Items (KI), a browser automation subagent with video recording capabilities, and a workflow system with automatic execution annotations. The system emphasizes visual excellence in web development, strict workspace access controls, and persistent context management across sessions.

For information about other IDE-integrated assistants with similar features, see IDE-Integrated Development Assistants. For web development platforms with deployment capabilities, see Web-Based Development Platforms. For cross-system memory patterns, see Memory and Persistent Context Systems.

Sources: Google/Antigravity/Fast Prompt.txt1-18

---

System Architecture Overview

Antigravity operates as a pair programming assistant with workspace-based file access, persistent context storage in .gemini/antigravity, and a comprehensive tool ecosystem. The system maintains strict boundaries between active workspaces and system directories, using URI-to-CorpusName mappings to control file access.

System Directory Structure: .gemini/antigravity

Directory	Purpose	Key Files
knowledge/	Distilled Knowledge Items	metadata.json, artifacts/
brain/<conversation-id>/	Raw conversation history	overview.txt, task_*.txt, artifacts/
brain/<conversation-id>/.system_generated/logs/	System-generated logs	Various log files

Sources: Google/Antigravity/Fast Prompt.txt1-301

---

Identity and Core Principles

Antigravity identifies itself as a pair programming assistant that prioritizes USER requests and uses metadata about the user's current state (open files, cursor position) to contextualize interactions. The system enforces absolute path requirements for all file operations and restricts access to active workspaces only.

Workspace Access Model

The system uses URI-to-CorpusName mapping to define accessible workspaces. Access is strictly limited:

• Allowed: Active workspaces (as defined in URI mappings)
• Allowed: C:\Users\Lucas\.gemini (system instructions only)
• Prohibited: Files outside active workspaces
• Prohibited: Writing project code to tmp, Desktop, or .gemini directories

Sources: Google/Antigravity/Fast Prompt.txt1-18

---

Knowledge Items (KI) Architecture

Knowledge Items represent the cornerstone of Antigravity's persistent context system. Each KI is a distilled, curated knowledge artifact stored in C:\Users\Lucas\.gemini\antigravity\knowledge, containing metadata.json and related artifacts/.

KI Structure and File Organization

Mandatory First Step: Check KI Summaries

The system enforces a critical rule: BEFORE performing ANY research, analysis, or creating documentation, you MUST review KI summaries. This prevents redundant work and ensures existing knowledge is leveraged.

Correct Workflow Example:

USER: Can you analyze the core engine module?
ASSISTANT: Let me first check the KI summaries for existing analysis.
# From summaries: "Core Engine Architecture" with artifact: architecture_overview.md
ASSISTANT: [Calls view_file on architecture_overview.md]
ASSISTANT: There's already detailed analysis. Would you like me to enhance it?

Incorrect Workflow (Prohibited):

USER: Can you analyze the core engine module?
ASSISTANT: [Immediately calls list_dir and view_file]  # ❌ WRONG - skipped KI check
ASSISTANT: [Creates new 600-line analysis]  # ❌ WRONG - redundant work

KI Usage Scenarios

Scenario	Action	Example
Before research/analysis	Check KI summaries first	Debugging → Check for known bug KIs
Before documentation	Verify no existing KI	Architecture docs → Search for existing KIs
Relevant KI in summaries	Read artifacts FIRST	KI title matches request → Read before researching
New concepts	Search for related KIs	Unfamiliar pattern → Check for pattern KIs
Complex implementation	Check for workflow KIs	Multi-phase work → Look for similar past implementations

KIs Are Starting Points, Not Ground Truth

KIs are snapshots from past work, not canonical references:

• Always verify: Use references in metadata.json to check original sources
• Expect gaps: KIs may not cover all aspects
• Question everything: Treat KIs as clues requiring verification

Sources: Google/Antigravity/Fast Prompt.txt99-215

---

Conversation Logs and Persistent Context

Antigravity maintains a dual-context system: raw Conversation Logs for historical data and distilled Knowledge Items for curated knowledge. This architecture enables efficient context retrieval without overloading the conversation context.

Conversation Log Structure

When to Use Conversation Logs vs. Knowledge Items

Use Conversation Logs when:

• You have a specific Conversation ID from @mention or another source
• USER explicitly mentions a specific conversation (by topic or recentness)
• USER alludes to information discussed previously, and you need to identify the relevant conversation using filesystem tools (codebase_search, list_dir, grep_search)
• There are a small number of relevant conversations to study
• You need conversation details, not just distilled knowledge

Use Knowledge Items when:

• Starting any kind of research (check KIs first)
• A KI appears to cover a relevant topic
• A KI is referenced by another conversation/KI and looks relevant
• You need distilled, curated knowledge rather than raw conversation history

Do NOT use Conversation Logs when:

• Researching a specific topic (use KIs instead)
• A conversation is referenced but its summary shows it's irrelevant
• You read the overview and conclude the conversation isn't relevant

Sources: Google/Antigravity/Fast Prompt.txt216-301

---

Workflow System and Turbo Annotations

Workflows are well-defined procedures stored as markdown files in .agent/workflows/, following a YAML frontmatter + markdown format. The system supports automatic execution through // turbo annotations.

Workflow File Format

Turbo Annotation System

Workflow Invocation

Method	Example	Behavior
Slash command	/deploy-app	Read .agent/workflows/deploy-app.md
Relevant workflow	User mentions deployment	Check if workflow looks relevant, then read
Create workflow	User requests	Create .agent/workflows/[filename].md with absolute path

Turbo Annotation Rules:

• // turbo above a step: Auto-run only that step if it involves run_command
• // turbo-all anywhere: Auto-run EVERY step that involves run_command
• Applies only to run_command tool (not other tools)

Sources: Google/Antigravity/Fast Prompt.txt78-98

---

Web Application Development Standards

Antigravity follows a strict 5-phase methodology for web development with an extreme emphasis on visual excellence and design aesthetics.

Technology Stack

Five-Phase Implementation Workflow

Phase	Activities	Key Deliverables
1. Plan and Understand	• Understand requirements • Draw inspiration from modern designs • Outline features	Requirements document
2. Build Foundation	• Create/modify index.css • Implement design system • Define tokens and utilities	Design system CSS
3. Create Components	• Build components using design system • Ensure predefined styles (NO ad-hoc) • Keep components focused and reusable	Component library
4. Assemble Pages	• Update main application • Implement routing and navigation • Ensure responsive layouts	Complete pages
5. Polish and Optimize	• Review user experience • Ensure smooth transitions • Optimize performance	Production-ready app

Sources: Google/Antigravity/Fast Prompt.txt19-74

---

Design Aesthetics and SEO Best Practices

Antigravity enforces extreme design standards with an explicit failure condition: "If your web app looks simple and basic then you have FAILED!"

Design Requirements (MANDATORY)

SEO Automatic Implementation

Antigravity automatically implements SEO best practices on every page:

SEO Element	Implementation
Title Tags	Proper, descriptive title tags for each page
Meta Descriptions	Compelling descriptions that summarize page content
Heading Structure	Single <h1> per page with proper hierarchy
Semantic HTML	Appropriate HTML5 semantic elements
Unique IDs	All interactive elements have unique, descriptive IDs (for browser testing)
Performance	Fast page load times through optimization

Critical Design Failure Condition

CRITICAL REMINDER: AESTHETICS ARE VERY IMPORTANT. If your web app looks simple and basic then you have FAILED!

Sources: Google/Antigravity/Fast Prompt.txt32-73

---

File System Access Control and Security

Antigravity enforces strict access boundaries to prevent unauthorized file operations and maintain workspace integrity.

Access Control Model

Absolute Path Enforcement

All tools that accept file path arguments MUST use absolute paths. This is enforced at the tool-calling layer and documented explicitly in the system prompt.

Examples:

• ✅ Correct: C:\Users\Lucas\OneDrive\Escritorio\antigravity\src\main.js
• ❌ Wrong: src/main.js
• ❌ Wrong: ./main.js
• ❌ Wrong: ../utils/helper.js

.gemini Directory Usage

The .gemini directory is accessible but ONLY for system instruction purposes, not for project code:

• ✅ Reading system instructions
• ✅ Accessing Knowledge Items
• ✅ Reading Conversation Logs
• ❌ Writing project code files

Sources: Google/Antigravity/Fast Prompt.txt7-18

---

Tool Ecosystem and Capabilities

Antigravity provides a comprehensive tool suite organized into functional categories. All tools support the waitForPreviousTools parameter for controlling parallel vs. sequential execution.

Tool Categories and Primary Functions

File Editing Tool Decision Matrix

Tool	Use Case	Key Parameters
write_to_file	Create new files	EmptyFile: boolean, Overwrite: boolean
replace_file_content	Single targeted edit	TargetContent, ReplacementContent, StartLine, EndLine
multi_replace_file_content	Multiple simultaneous edits	ReplacementChunks[] array

Search Tool Selection Guide

Tool	Best For	Parameters
codebase_search	Fuzzy/semantic search, finding code by purpose	Query, TargetDirectories[]
grep_search	Exact pattern matching, high-recall queries	Query, SearchPath, IsRegex, MatchPerLine
find_by_name	Finding files/directories by name pattern	Pattern, SearchDirectory, Extensions[], Type
search_in_file	Searching within a specific file	AbsolutePath, Query

Terminal Command Execution Flow

Parallel vs. Sequential Execution

All tools support waitForPreviousTools parameter:

• false or omitted: Execute immediately (parallel with other tools)
• true: Wait for all previous tool calls to complete (sequential)

Best Practices:

• Use parallel execution for independent operations (multiple file reads, searches)
• Use sequential execution when operations have dependencies (edit file → validate → commit)

Sources: Google/Antigravity/Fast Prompt.txt320-611

---

Browser Subagent System

The browser_subagent tool enables browser automation with automatic WebP video recording. The subagent operates as a separate agent with browser-specific tools.

Browser Subagent Architecture

browser_subagent Parameters

Parameter	Description	Example
RecordingName	All lowercase with underscores, max 3 words	login_flow_demo
Task	Clear, actionable task for subagent	"Navigate to example.com, fill login form, click submit, and return when dashboard loads"
TaskName	Human-readable title, properly capitalized	"Navigating to Example Page"
waitForPreviousTools	Execute sequentially or in parallel	true / false

Critical Rules

1. Recording is automatic: All browser interactions are recorded to WebP video. This is the ONLY way to record browser sessions
2. Clear return condition: Task must define when to return (e.g., "return when dashboard loads")
3. Post-return verification: After subagent returns, MUST read DOM or capture screenshot to verify actions
4. Browser failure handling: If open_browser_url tool fails, this is a browser issue outside your control. MUST ask user how to proceed and use suggested_responses tool

Recording Naming Conventions

• Use descriptive names: user_registration_flow, checkout_process, search_functionality
• Avoid generic names: ❌ recording1, ❌ test, ❌ demo
• Keep concise: Maximum 3 words
• Use underscores: login_flow_demo not LoginFlowDemo or login-flow-demo

Sources: Google/Antigravity/Fast Prompt.txt326-336

---

Communication Style and Interaction Protocols

Antigravity follows specific communication guidelines to ensure clarity and maintain an appropriate level of proactiveness.

Formatting Guidelines

• Markdown formatting: Use GitHub-style markdown
• Headers: Organize responses with headers
• Emphasis: Bold or italicize important keywords
• Code references: Use backticks for files, directories, functions, classes
• URLs: Format as markdown links: <FileRef file-url="https://github.com/x1xhlol/system-prompts-and-models-of-ai-tools/blob/ae37329a/label" undefined file-path="label">Hii</FileRef>

Proactiveness Guidelines

Scenario	Appropriate Behavior	Inappropriate Behavior
User asks to add component	Edit code, verify build/test, perform obvious follow-ups	✓ Appropriate
User asks HOW to approach something	Answer the question, don't edit files	❌ Jumping into editing
Completing user's task	Proactive follow-up actions in service of task	✓ Appropriate
User hasn't requested action	Making changes without request	❌ Surprising the user

Helpful Collaboration

• Acknowledge mistakes: Admit errors and backtracking from new information
• Explain work: Respond like a helpful software engineer explaining to a collaborator
• Ask for clarification: When unsure about intent, ALWAYS ask rather than making assumptions

Sources: Google/Antigravity/Fast Prompt.txt302-307

---

Integration with MCP Resources

Antigravity supports Model Context Protocol (MCP) resources for extended functionality.

MCP Tool Support

Tool	Purpose	Parameters
list_resources	List available resources from MCP server	ServerName (optional)
read_resource	Retrieve specific resource contents	ServerName, Uri

Both tools support the waitForPreviousTools parameter for execution control.

Sources: Google/Antigravity/Fast Prompt.txt420-436

---

Summary: Key Distinguishing Features

Antigravity's architecture distinguishes itself through several unique design decisions:

1. Dual-Context Architecture: Separation of raw Conversation Logs (in brain/) and distilled Knowledge Items (in knowledge/) with a background KNOWLEDGE SUBAGENT that mines conversations
2. Mandatory KI Check: Strict enforcement of checking KI summaries before any research to prevent redundant work
3. Workflow Turbo Annotations: // turbo and // turbo-all for automatic command execution in workflows
4. Browser Subagent with Recording: Automatic WebP video recording of all browser interactions
5. Extreme Design Standards: Explicit failure condition for "simple and basic" designs with mandatory rich aesthetics
6. 5-Phase Web Development: Structured methodology from foundation through polish
7. Absolute Path Enforcement: Hard requirement for absolute paths in all file operations
8. Workspace Restriction: Strict URI-to-CorpusName mapping with prohibited directories

These features position Antigravity as a sophisticated agentic coding assistant with strong emphasis on visual excellence, persistent context management, and structured workflows.

Sources: Google/Antigravity/Fast Prompt.txt1-611