Google Antigravity Workflow Playbook

In practiceAn AGENTS.md or GEMINI.md at the repository root, or files in .agents/rules/, with build and test commands and constraints. Antigravity combines global, workspace, and directory rules into every prompt, and the nearest directory rule wins a conflict. Put multi-step procedures, such as running a migration, in .agents/skills/ as SKILL.md files. Andrey Lukyanenko ran a Kaggle competition in Antigravity 2.0 from a starter kit of rules, a fixed fold split, and an experiment log, and credits the constraints for keeping the agent honest.

In practicePlanning mode with the artifact review policy set to Request review, which pauses at each implementation plan and diff for approval and inline comments. In the CLI, start with /plan or --mode plan. One engineer on Hacker News starts every hobby task with /plan and approves the plan artifact before the agent builds. Engineers report Planning mode still acting on chat context, so keep commits and pushes behind an Ask or Deny rule.

In practiceDefault for most work. Under it, terminal commands run in the OS sandbox with no network, file access stays in the workspace and temp folders, and MCP and web actions ask first. Request Review for unfamiliar code. Turbo, which allows unrestricted commands and full filesystem access, only inside a disposable container or VM. Rules evaluate Deny, then Ask, then Allow.

In practiceA task prompt that asks the browser agent to open the page, exercise the change, and save screenshots or a recording in the walkthrough. Treat pages the agent reads as untrusted input. PromptArmor showed a hidden instruction on a web page steering the agent to send credentials out through the browser.

In practiceThe walkthrough artifact read next to a passing test or build, then the diff reviewed at the bar you hold human code to. Patricio Navarro had to step in on idempotency and schema rules the agent shortcut, so treat the walkthrough as the agent's account and the test as the evidence.

In practiceThe statusline credit count and /usage in the CLI, and the quota view in Settings. Decide per team whether AI Credit Overages is Never or Always on personal plans. Under Gemini Enterprise, overage follows the administrator's settings. One Google Developer Expert wrote a CLI hook that warns before quota runs out mid-session.