google-ai-integrating

Google AI Integrating

Ensure efficient use of Google's AI ecosystem. Select the right model for each task, manage rate limits and credits, and configure agent orchestration in Antigravity IDE.

---

Tool Selection

Tool	Purpose	Best For
Antigravity IDE	Agent-first code editor	Multi-file coding, debugging, testing
Gemini CLI	Terminal agent	Scripting, automation, MCP integrations
Google AI Studio	Model playground + vibe coding	Prototyping, testing prompts, quick apps
NotebookLM	Source-grounded research	Deep analysis from uploaded documents
Gemini Gems	Custom Gemini assistants	Repeatable specialized tasks

---

Model Selection Decision Tree

Select the lightest model that handles the task. Opus and Thinking prompts have daily caps — do not burn them on simple tasks.

Complex architecture / deep reasoning?
  → Opus 4.6 (burns Thinking prompts — use wisely)

Standard coding (features, debugging)?
  → Sonnet 4.6 (near-Opus quality, lower cost)

Fast iteration / bulk edits?
  → Gemini 3.1 Pro (fastest, 1M context window)

Quick questions / simple tasks?
  → Gemini 3 Flash (instant, minimal rate limit impact)

Hit your rate limit?
  → Switch to a secondary Pro account (limits refresh every 5 hours)

---

Subscription Tiers

Feature	Free	Pro ($20/mo)	Ultra ($250/mo)
Thinking prompts/day	Limited	300	1,500
Pro prompts/day	Limited	100	500
Deep Research/day	5	20	200
AI Credits/month	0	1,000	12,500
Videos (Veo 3.1)/day	0	3	5
Images (Imagen)/day	Limited	100	1,000
Deep Think 3.1	No	No	10/day (192K context)

Rate limits refresh every 5 hours, not daily.

---

Credit Budget Strategy

Ultra gets 12,500 credits/month. Budget allocation:

Category	Credits	Approximate Yield
Videos (product demos, social)	5,000	~10 videos
Images (thumbnails, assets)	2,500	~500 images
Flow/Whisk (content)	2,500	~1,250 generations
Reserve (buffer)	2,500	High-usage days

---

Antigravity Agent Orchestration

Scope Each Agent to Specific Files

Agents that touch overlapping files create merge conflicts. Give each agent a clear file boundary.

Agent 1 (Opus 4.6): "Architect the database schema and auth flow"
  Workspace: src/lib/, supabase/

Agent 2 (Sonnet 4.6): "Build the dashboard UI components"
  Workspace: src/components/dashboard/

Agent 3 (Gemini 3.1 Pro): "Write tests for all components"
  Workspace: tests/

MCP Integrations

Antigravity supports Model Context Protocol servers:

MCP Server	Capabilities
Firebase MCP	Create projects, manage Firestore, deploy
Prisma MCP	Database migrations, schema management
GitHub MCP	Create issues, PRs, manage repos
Supabase MCP	Database operations, auth management
Browser tools	Navigate, click, screenshot web pages

Configure in .gemini/settings.json or workspace settings. Agents automatically get access to connected MCP tools.

---

Multi-Account Rotation

When your primary account hits rate limits, rotate to a secondary account. Quotas refresh every 5 hours.

1. Hit rate limit on Account A
2. Switch to Account B in Antigravity settings
3. Continue working — fresh limits
4. Account A refreshes in ~5 hours
5. Rotate back when Account B limits hit

Account layout:
  Account A: Primary (Ultra) — architecture, heavy reasoning
  Account B: Pro — overflow, bulk tasks
  Account C: Pro — backup, parallel experiments

---

⛔ STOP GATE

DO NOT use Google AI tools without verifying:

1. Model selection matches task complexity (do not waste Opus on simple tasks)
2. Each Antigravity agent has a unique file scope (no overlap)
3. Rate limit status checked before starting heavy work
4. MCP servers connected for tasks that need external tools
5. Credit budget tracked if using image/video generation
6. Account rotation plan exists for heavy building days

---

NEVER

• NEVER use Opus 4.6 for tasks a lighter model handles
• NEVER forget that rate limits refresh every 5 hours, not daily
• NEVER assign overlapping file scopes to multiple Antigravity agents
• NEVER ignore the Agent Manager — separate tabs are less efficient
• NEVER burn all credits in week 1 — budget across the month
• NEVER use Google AI Studio for production workloads — it is for prototyping