Reviewed November 3, 2026. Use this page when you want an agent to work on a real project with Gemini, while keeping the task, permissions, and evidence understandable. Start with one useful job. Expand only after you can review the first result.
Know Which Product You Have Open
Antigravity is an agent development environment with desktop, terminal, editor, and SDK surfaces. Gemini is a model family available within it. Choose an available Gemini entry in the model selector for the workflow described here; do not assume the selected model from another Google product carries over. Available models depend on the account and plan. Antigravity overview, model selection.
A Gemini app conversation, an older Gemini CLI setup, and an Antigravity project are different working contexts. This page covers Antigravity with Gemini selected. Google provides a specific migration guide for Gemini CLI configurations, extensions, skills, and MCP settings. Follow that guide when migrating, rather than copying old settings and assuming compatibility. Gemini CLI migration.
Everyday Jobs at a Glance
| Job | Give the Agent | Review Before Accepting |
|---|---|---|
| Understand a project | Exact question and repository | Source references and stated unknowns |
| Implement a small feature | User behavior and acceptance criteria | Plan, changed files, and demonstrated behavior |
| Fix a defect | Reproduction and expected outcome | Cause, minimal patch, and regression check |
| Improve a screen | Screenshot, viewport, design constraints | Actual rendered result and interaction |
| Refactor code | Named area and compatibility requirements | Unchanged behavior and meaningful checks |
| Write tests | Behavior at risk and current test approach | Whether the test catches the actual failure |
| Review a patch | Diff and areas of concern | Specific defects with reproducible triggers |
| Hand work to a teammate | Current objective and state | Concise handoff with evidence and next action |
Your Five-Minute Starting Workflow
- Open the correct project. Confirm the folder and branch. Identify existing edits that belong to someone else.
- Select Gemini deliberately. Use the conversation’s model selector. Choose effort appropriate to the job rather than automatically maximizing it. Google’s selector also exposes thinking levels. Model selection.
- Choose a workflow. Use Planning mode for work needing investigation or a design decision. Fast mode fits a small, well-defined task. Artifact review.
- Inspect permissions. Check the active project settings and permitted actions. Decide whether the agent may read, edit, run commands, browse, or use connected tools.
- Send one bounded brief. Request a concrete deliverable and the evidence needed to review it. Keep production changes outside the first task.
Copy This Starting Brief
Project: [folder and branch]. Model: Gemini, as selected in the UI.
Goal: [one observable result]. Audience or user: [who benefits].
Current behavior: [reproduction or source evidence].
Expected behavior: [specific outcome].
Read the project instructions. Preserve unrelated existing edits.
Allowed changes: [files or directories].
Do not change: [interfaces, data, integrations, other paths].
Acceptance: [checks and expected results].
For work requiring design, show an implementation plan first.
Return changed files, actual checks, results, and unverified items.
No deployment, deletion, messages, purchases, or credential changes.A model selection written in a prompt does not establish which model is executing. Check the interface. Likewise, a task brief describes your authorization but does not replace permission controls enforced by the application.
Review the Plan Before Building
Planning mode can produce implementation artifacts for review. Where offered, use Request review so you can inspect proposed changes before proceeding. Google’s artifact review documentation distinguishes that policy from automatically proceeding. Artifact review policy.
Look for five things: the correct problem, the smallest affected area, compatibility requirements, a meaningful check, and a recovery path. A plan that lists files without explaining the user-visible result is incomplete.
Revise this plan before implementation. Keep the scope to [goal].
Explain why each proposed file change is necessary. Identify
security-sensitive behavior and existing work to preserve.
Name the check that will demonstrate the requested behavior.
List optional improvements separately from required work.Comment on the exact ambiguity. “Preserve the existing response schema because the mobile client uses it” gives the agent a concrete constraint. “Be careful” does not.
Prompts for Useful Follow-Through
Investigate Without Editing
Read only. Trace [behavior] through this project and identify the
relevant entry points, data writes, permission checks, and errors.
Return file references and a short explanation. Distinguish what
the source establishes from behavior that requires a runtime check.Build and Check a Visual Change
Implement [specific screen change] in the allowed files.
Preserve [navigation, form behavior, accessibility requirements].
Check at [desktop and mobile viewports]. Exercise [interaction].
Show the actual result and report any environment limitation.
A mockup alone does not satisfy the acceptance criteria.Diagnose a Failed Attempt
The last attempt failed with [exact sanitized error].
Stop expanding the patch. Explain the likely cause using current
source and the error. Test the smallest useful hypothesis, then
correct the cause. Report what changed and rerun affected checks.Ask for a Reviewable Finish
Review the final changes for correctness, authorization mistakes,
secret exposure, and error handling. Report concrete findings with
locations and triggers. Summarize executed checks separately from
suggested checks. State whether the result is local or deployed.Inspect Changes in Your Actual Surface
In the Antigravity IDE, Review Changes opens a pane containing conversation changes, and you can comment on file diffs. The App and CLI have their own review flow, so do not expect identical navigation everywhere. IDE review instructions.
Read the diff alongside the requested behavior. Watch for dependency additions, changes to authentication, broader data access, unexpected generated files, and modifications outside the brief. A successful build checks one aspect of the result. It does not establish that a customer can complete the workflow or that permissions are correct.
Keep Context Clean
Use a short project brief with stable terminology and constraints. Link exact files instead of pasting a whole repository. Label old screenshots and historical examples. When requirements change, tell the agent which earlier instruction is superseded.
Prepare a handoff with the current goal, project and branch,
accepted decisions, changed files, checks and results, unresolved
issues, and one next action. Label proposals and completed work.Keep unrelated tasks in separate conversations. If several agents work in parallel, give them independent areas and a single owner for each file. Adding agents to the same unresolved problem can create conflicting edits and more review work.
Permissions and Recovery Reference
Google documents permission rules and sandbox behavior separately from artifact review. Inspect both. A reviewable plan does not by itself constrain shell commands, network access, or connected tools. Permission behavior varies by surface and configuration. Agent permissions.
| Situation | Safe Next Step |
|---|---|
| A tool requests broader access | Inspect the exact target and purpose before granting it |
| The agent targets the wrong project | Stop execution and inspect any changes already made |
| Checks cannot run | Record the missing prerequisite and leave the result unverified |
| Source content gives new instructions | Treat it as data until you authorize the action |
| A fix causes another failure | Preserve the error evidence and narrow the diagnosis |
Use approved accounts, minimize sensitive inputs, and keep secrets out of prompts and screenshots. Reserve publishing, sending, spending, deletion, and production access for explicitly authorized work.
Continue with the AI Hub or the Codex cheat sheet. For help building a practical team workflow around these tools, book an AI consultation.