Press "Enter" to skip to content

A Power User Guide for Google Antigravity

After a near-disaster, the agent powering Google Antigravity said something that surprised me. I asked it for more details, and eventually the agent, Claude Opus 4.5 (Thinking) offered to write this Power User Guide. I present it here with no validation or editing, just in case it has any merit. (Does it? I don’t know!)

Expand the first block to see the exact markdown it wrote to disk:

markdown version

This Power User Guide was created by Claude Opus 4.5 (Thinking) via Antigravity 1.13.3. declare-ai.org

# Antigravity Power User Guide

A comprehensive reference for getting the most out of Antigravity.

---

## Glossary

| Term | Meaning |
|------|---------|
| **KI** | **Knowledge Item** — Persistent, curated knowledge distilled from conversations. Stored in `~/.gemini/antigravity/knowledge/` |
| **Artifact** | Files created during tasks: `task.md`, `implementation_plan.md`, `walkthrough.md`. Stored in `brain/<conversation-id>/` |
| **Brain** | The `~/.gemini/antigravity/brain/` directory — one subfolder per conversation containing artifacts |
| **Task Boundary** | The mechanism that creates the structured task UI during multi-step work |
| **Mode** | PLANNING → EXECUTION → VERIFICATION — the phases the agent moves through during a task |
| **Workflow** | Reusable procedure defined in `.agent/workflows/*.md` in your project |
| **Knowledge Subagent** | Separate async process that distills conversations into KIs after they occur |
| **Browser Recording** | WebP videos auto-saved when browser tools are used — stored in `browser_recordings/` |
| **MCP** | Model Context Protocol — for connecting external tools (configured in `mcp_config.json`) |

---

## Directory Structure

```
~/.gemini/antigravity/
├── brain/                      # Conversation artifacts
│   └── <conversation-id>/
│       ├── task.md             # Current task checklist
│       ├── task.md.resolved.N  # Version history
│       ├── implementation_plan.md
│       └── walkthrough.md
├── knowledge/                  # Knowledge Items
│   └── <ki-name>/
│       ├── metadata.json       # Title, summary, source refs
│       ├── timestamps.json     # Access/modification times
│       └── artifacts/          # The actual markdown docs
├── conversations/              # Encrypted conversation state (.pb)
├── implicit/                   # Encrypted implicit context (.pb)
├── browser_recordings/         # WebP videos from browser sessions
├── mcp_config.json            # MCP server configuration
└── user_settings.pb           # User preferences (encrypted)
```

---

## Power User Tips

### 1. Workflows with Turbo Mode

Create project-specific workflows in `.agent/workflows/` that the agent will follow:

```markdown
---
description: Deploy to production
---
// turbo-all
1. Run tests: `bin/agent_rake`
2. Build: `bundle exec rake build`
3. Push: `gem push pkg/*.gem`
```

**Annotations:**
- `// turbo` — Auto-run the next command without approval
- `// turbo-all` — Auto-run ALL commands in the workflow

**Usage:** Type `/deploy` or say "run the deploy workflow"

### 2. AGENTS.md — Your Control File

Create `AGENTS.md` in your project root with rules the agent will follow:

```markdown
## Commit Style
- Use Conventional Commits format
- Include AI attribution: `Co-authored-by: Antigravity <[email protected]>`

## Forbidden Actions
- Never modify files in `vendor/`
- Never push directly to `main`

## Definition of Done
- All tests pass (`bin/agent_rake`)
- CHANGELOG.md updated
- Documentation updated
```

### 3. Reference Past Conversations

The agent can search and retrieve context from previous conversations:

- "Look at the conversation where we refactored the API"
- "What did we decide about authentication last week?"
- "Check the implementation plan from conversation 922ca3e5..."

### 4. Artifact Version History

The `.resolved.N` files in `brain/` directories are version history:

```
task.md.resolved.0  # First version
task.md.resolved.1  # After first update
task.md.resolved.2  # After second update
task.md             # Current version
```

Useful for seeing how a plan evolved or recovering earlier versions.

### 5. Knowledge Items Are Editable

KI artifacts are just markdown files. You can:

- **Edit** them to correct mistakes
- **Delete** entire KI directories to remove outdated knowledge
- **Create** new ones manually (though the subagent usually handles this)

```bash
# View all KIs
ls ~/.gemini/antigravity/knowledge/

# Check KI metadata
cat ~/.gemini/antigravity/knowledge/*/metadata.json | jq .title
```

### 6. Image Generation

The agent can generate images for mockups, diagrams, and assets:

- "Generate a login page mockup"
- "Create an icon for the settings menu"
- "Design a dark mode version of this UI"

Generated images are saved as artifacts.

### 7. Browser Automation

The agent can control a real browser for:

- Testing web applications
- Filling forms and clicking buttons
- Debugging UI issues
- Taking screenshots

Recordings are automatically saved as WebP videos in `browser_recordings/`.

### 8. Experimental Conversations

If you're experimenting and don't want patterns codified as KIs, say:

> "This is experimental — don't create Knowledge Items from this."

Consider adding to `AGENTS.md`:

```markdown
## Experimental Conversations
If the user says "this is experimental", do NOT create Knowledge Items.
Mark any artifacts with a CAUTION banner.
```

---

## Useful Commands

```bash
# List all conversations with artifacts
ls ~/.gemini/antigravity/brain/

# Find recent KI updates
find ~/.gemini/antigravity/knowledge -name "*.md" -mtime -1

# View KI summaries
cat ~/.gemini/antigravity/knowledge/*/metadata.json | jq '{title, summary}'

# Check conversation artifact count
for d in ~/.gemini/antigravity/brain/*/; do echo "$(ls "$d" | wc -l) $d"; done | sort -rn | head -10

# Browser recording locations
ls ~/.gemini/antigravity/browser_recordings/*/metadata.json
```

---

## Encrypted Files (Read-Only)

The `.pb` files in `conversations/` and `implicit/` are encrypted Protocol Buffer files. They contain:

- Full conversation history
- Implicit context (workspace state, file contents, etc.)

**You cannot decode these** without Google's decryption keys. The human-readable exports are in `brain/` and `knowledge/`.

---

## MCP Configuration

To connect external tools, edit `~/.gemini/antigravity/mcp_config.json`:

```json
{
  "servers": {
    "my-tool": {
      "command": "/path/to/mcp-server",
      "args": ["--flag"]
    }
  }
}
```

The agent can then access resources and call tools from connected MCP servers.

---

*Last updated: 2026-01-03*

Expand the second block to see a blog-formatted version of it:

blog-formatted version

This Power User Guide was created by Claude Opus 4.5 (Thinking) via Antigravity 1.13.3. declare-ai.org


Antigravity Power User Guide

A comprehensive reference for getting the most out of Antigravity.


Glossary

TermMeaning
KIKnowledge Item — Persistent, curated knowledge distilled from conversations. Stored in ~/.gemini/antigravity/knowledge/
ArtifactFiles created during tasks: task.md, implementation_plan.md, walkthrough.md. Stored in brain/<conversation-id>/
BrainThe ~/.gemini/antigravity/brain/ directory — one subfolder per conversation containing artifacts
Task BoundaryThe mechanism that creates the structured task UI during multi-step work
ModePLANNING → EXECUTION → VERIFICATION — the phases the agent moves through during a task
WorkflowReusable procedure defined in .agent/workflows/*.md in your project
Knowledge SubagentSeparate async process that distills conversations into KIs after they occur
Browser RecordingWebP videos auto-saved when browser tools are used — stored in browser_recordings/
MCPModel Context Protocol — for connecting external tools (configured in mcp_config.json)

Directory Structure

~/.gemini/antigravity/
├── brain/                      # Conversation artifacts
│   └── &lt;conversation-id>/
│       ├── task.md             # Current task checklist
│       ├── task.md.resolved.N  # Version history
│       ├── implementation_plan.md
│       └── walkthrough.md
├── knowledge/                  # Knowledge Items
│   └── &lt;ki-name>/
│       ├── metadata.json       # Title, summary, source refs
│       ├── timestamps.json     # Access/modification times
│       └── artifacts/          # The actual markdown docs
├── conversations/              # Encrypted conversation state (.pb)
├── implicit/                   # Encrypted implicit context (.pb)
├── browser_recordings/         # WebP videos from browser sessions
├── mcp_config.json            # MCP server configuration
└── user_settings.pb           # User preferences (encrypted)

Power User Tips

1. Workflows with Turbo Mode

Create project-specific workflows in .agent/workflows/ that the agent will follow:

---
description: Deploy to production
---
// turbo-all
1. Run tests: `bin/agent_rake`
2. Build: `bundle exec rake build`
3. Push: `gem push pkg/*.gem`

Annotations:

  • // turbo — Auto-run the next command without approval
  • // turbo-all — Auto-run ALL commands in the workflow

Usage: Type /deploy or say “run the deploy workflow”

2. AGENTS.md — Your Control File

Create AGENTS.md in your project root with rules the agent will follow:

## Commit Style
- Use Conventional Commits format
- Include AI attribution: `Co-authored-by: Antigravity &lt;[email protected]>`

## Forbidden Actions
- Never modify files in `vendor/`
- Never push directly to `main`

## Definition of Done
- All tests pass (`bin/agent_rake`)
- CHANGELOG.md updated
- Documentation updated

3. Reference Past Conversations

The agent can search and retrieve context from previous conversations:

  • “Look at the conversation where we refactored the API”
  • “What did we decide about authentication last week?”
  • “Check the implementation plan from conversation 922ca3e5…”

4. Artifact Version History

The .resolved.N files in brain/ directories are version history:

task.md.resolved.0  # First version
task.md.resolved.1  # After first update
task.md.resolved.2  # After second update
task.md             # Current version

Useful for seeing how a plan evolved or recovering earlier versions.

5. Knowledge Items Are Editable

KI artifacts are just markdown files. You can:

  • Edit them to correct mistakes
  • Delete entire KI directories to remove outdated knowledge
  • Create new ones manually (though the subagent usually handles this)
# View all KIs
ls ~/.gemini/antigravity/knowledge/

# Check KI metadata
cat ~/.gemini/antigravity/knowledge/*/metadata.json | jq .title

6. Image Generation

The agent can generate images for mockups, diagrams, and assets:

  • “Generate a login page mockup”
  • “Create an icon for the settings menu”
  • “Design a dark mode version of this UI”

Generated images are saved as artifacts.

7. Browser Automation

The agent can control a real browser for:

  • Testing web applications
  • Filling forms and clicking buttons
  • Debugging UI issues
  • Taking screenshots

Recordings are automatically saved as WebP videos in browser_recordings/.

8. Experimental Conversations

If you’re experimenting and don’t want patterns codified as KIs, say:

“This is experimental — don’t create Knowledge Items from this.”

Consider adding to AGENTS.md:

## Experimental Conversations
If the user says "this is experimental", do NOT create Knowledge Items.
Mark any artifacts with a CAUTION banner.

Useful Commands

# List all conversations with artifacts
ls ~/.gemini/antigravity/brain/

# Find recent KI updates
find ~/.gemini/antigravity/knowledge -name "*.md" -mtime -1

# View KI summaries
cat ~/.gemini/antigravity/knowledge/*/metadata.json | jq '{title, summary}'

# Check conversation artifact count
for d in ~/.gemini/antigravity/brain/*/; do echo "$(ls "$d" | wc -l) $d"; done | sort -rn | head -10

# Browser recording locations
ls ~/.gemini/antigravity/browser_recordings/*/metadata.json

Encrypted Files (Read-Only)

The .pb files in conversations/ and implicit/ are encrypted Protocol Buffer files. They contain:

  • Full conversation history
  • Implicit context (workspace state, file contents, etc.)

You cannot decode these without Google’s decryption keys. The human-readable exports are in brain/ and knowledge/.


MCP Configuration

To connect external tools, edit ~/.gemini/antigravity/mcp_config.json:

{
  "servers": {
    "my-tool": {
      "command": "/path/to/mcp-server",
      "args": ["--flag"]
    }
  }
}

The agent can then access resources and call tools from connected MCP servers.


Last updated: 2026-01-03

P. S. – If you want to hear the story of the near-disaster, comment and let me know!

Be First to Comment

Reposts

Leave a Reply

Your email address will not be published. Required fields are marked *


To respond on your own website, enter the URL of your response which should contain a link to this post's permalink URL. Your response will then appear (possibly after moderation) on this page. Want to update or remove your response? Update or delete your post and re-enter your post's URL again. (Find out more about Webmentions.)