zmime.com
Agentic platform · Antigravity

Prompt for Antigravity

Antigravity agents work across editor, terminal and browser. Prompt them like you would brief a contractor who has to show evidence.

  • Agentic platform
  • Gemini
  • Browser tasks
  • Verification
01 — Overview

What Antigravity rewards

Google Antigravity is an agent-first development platform built on Gemini models, announced alongside Gemini 3 in November 2025. Agents can plan a change, edit files, run terminal commands and drive a browser, then hand back artefacts such as plans and walkthroughs as evidence of what they did.

That autonomy shifts the prompting job. Your prompt is a brief, not a code request: state the outcome, the boundaries, the environment details the agent cannot infer, and what proof you want back. The more explicit the verification step, the less you have to re-check by hand.

  • 01Agents that operate across editor, terminal and browser in one task.
  • 02Plans and walkthrough artefacts you can review instead of reading raw diffs.
  • 03Good fit for end-to-end work: implement, run, check in a browser, report.
  • 04Gemini long context for large codebases and specifications.
02 — Anatomy

Briefing an autonomous agent

  • Outcome What must be true when the task is finished, in observable terms.
  • Environment How to run the app, ports, seed data, test accounts, feature flags.
  • Boundaries Files, services and data the agent must not touch.
  • Verification path Exact commands plus the browser steps that prove it works.
  • Evidence What to hand back: plan, diff summary, screenshots, test output.
  • Stop conditions When to pause and ask rather than improvise.
03 — The prompts

4 prompts written for Antigravity

Paste, then replace the [bracketed] parts. Keep the wording identical for anything you want repeated across a set.

Nothing matches that word. Clear the field to see every prompt.

Prompt 01

End-to-end feature task with browser verification

Work you want implemented and actually checked in a running app.

  • agent
  • end to end
Template · plain text
OUTCOME: [user-visible behaviour] works in the running app at [URL/port].

ENVIRONMENT
- Start the app with `[command]`, it serves on [port].
- Test account: [credentials source — never a real secret in the prompt].
- Seed data: `[command]` if the database is empty.
- Feature flag: [name] must be [on/off].

IMPLEMENT
- [Requirement 1]
- [Requirement 2]
- Edge cases: [empty state], [validation failure], [slow network]

BOUNDARIES
- Only modify files under [paths].
- Do not touch [migrations / generated code / third-party config].
- No new dependencies without asking first.

VERIFY (all of these before reporting done)
1. `[build]`, `[lint]`, `[test]` pass.
2. In the browser: [step-by-step user journey], confirming [expected result at each step].
3. Check at 360px width and confirm no layout overflow.
4. Confirm the browser console has no new errors.

REPORT BACK
- A short plan before you start.
- What changed, by file, and why.
- Screenshots of the verified journey.
- Anything you could not verify, stated plainly.

STOP AND ASK IF: the change requires a schema migration, a new dependency, or touching [sensitive area].
Prompt 02

Bug reproduction and fix

Handing over a defect with enough detail to be actioned autonomously.

  • bug
  • agent
Template · plain text
BUG: [what users see] — expected [correct behaviour].

REPRODUCE
1. Run the app with `[command]`.
2. [Step], [step], [step].
3. Observe [wrong behaviour].

EVIDENCE
"""
[paste error output, console log or failing test]
"""

TASK
1. Reproduce it first and confirm you have seen the failure. Report what you observed.
2. Diagnose the root cause and name the specific line before proposing a fix.
3. Add a regression test that fails on the current code.
4. Apply the smallest fix that makes it pass.
5. Re-run the reproduction steps in the browser to confirm.

CONSTRAINTS
- Minimal diff. No refactoring around the fix.
- If the root cause sits in [area I care about], stop and explain before changing it.
- Report other defects you notice rather than fixing them.
Prompt 03

Migration or upgrade task

Dependency upgrades and framework migrations, done incrementally.

  • migration
  • upgrade
Template · plain text
TASK: Upgrade [dependency] from [current version] to [target version] in this repository.

BEFORE CHANGING ANYTHING
1. Inventory every usage, grouped by the API involved, with file references.
2. Read the upstream changelog and list the breaking changes that actually affect us.
3. Give me a step plan ordered by risk, lowest first.

THEN, ONE STEP AT A TIME
- Apply the step, run `[test command]`, and report the result before continuing.
- If a step needs an API change with no direct equivalent, stop and ask.

CONSTRAINTS
- No unrelated formatting or version bumps.
- Keep behaviour identical; note anything that cannot stay identical.
- Do not touch lockfiles for other packages.

DONE MEANS: build, lint and full test suite pass, and the app runs and serves [key page] correctly in the browser.
Prompt 04

UI review across breakpoints

Using browser control to catch responsive problems.

  • browser
  • responsive
Template · plain text
Review [page] of the running app for responsive and accessibility issues.

SETUP: start with `[command]`, page is at [URL]. Log in as [test account source] if needed.

CHECK AT THESE WIDTHS: 360px, 480px, 768px, 1024px, 1440px.

FOR EACH WIDTH REPORT
- Horizontal overflow or clipped content.
- Text that wraps badly or drops below 14px effective size.
- Touch targets under 44px.
- Images without dimensions causing layout shift.
- Navigation usability, including whether the mobile menu opens and closes.

ALSO CHECK
- Keyboard-only navigation: focus order, visible focus, no traps.
- Whether every interactive element has an accessible name.
- Console errors and failed network requests.

REPORT: a table of issue, width, severity, location, suggested fix, with a screenshot per issue. Say clearly which items need manual testing with assistive technology rather than being verifiable from the browser alone.
04 — Craft notes

What separates a good prompt from a wasted run

Do this 05

  • Give the agent everything it needs to run the app itself.
  • Ask for a plan and approve it before a long autonomous run.
  • Define verification as commands plus a browser journey.
  • Request evidence: test output, screenshots, a change list.
  • Set explicit stop conditions for risky areas.

Not this 05

  • Putting real credentials or secrets in a prompt.
  • Letting an agent run migrations against shared data.
  • Broad goals like "modernise the frontend" with no boundaries.
  • Accepting "done" without reading the evidence.
  • Multiple unrelated objectives in a single task.
05 — Questions

Asked often, answered plainly

How much autonomy should I give?

Scale it to reversibility. Local, easily reverted work in a branch is fine to run end to end. Anything touching shared state, production config or data deletion should require approval mid-task, which is what stop conditions are for.

What makes agent tasks fail most often?

Missing environment detail. If the agent cannot start the app, seed data or log in, it cannot verify anything and will report unverified work as done. Put the run instructions in the prompt or a project rules file.

Do I still need to read the diff?

Read the plan and the evidence at minimum, and the diff for anything security-related or user-facing. Walkthrough artefacts speed review up, but you remain accountable for what ships.

06 — Sources

Read the primary documentation

Vendors change flags, limits and defaults often. Where this site disagrees with official docs, the docs win.