A friend asked me today if they should get Google AI Pro just to use Antigravity. I told them to please DO NOT!

They jokingly said: "Enlighten me 🤣"

And that's when they just saw "typing..." for solid 10 minutes.

The Dumpster Fire That Is Gemini in Antigravity

If you're one of the lucky ones who didn't get to experience the pain of using Gemini in Antigravity, I congratulate you, please don't start. 😅

Oh you wanna know why? Allow me...

1. It's buggy in ways that have nothing to do with the model

Essentially on Windows, Antigravity has a path handling bug that causes it to not tie your chats to your workspace/project.

Everyone I know on Windows hits this.

The sidebar (which should contain your latest chats) is empty, you have to go to recent chats and find whatever conversation you were having in said project.

You click it, and it's exactly like opening a chat from Project A while you're in Project B.

People are having the same issue calling it "chats are not being saved".

I investigated and even reproduced the issue on a sandbox machine, then filed a bug report 4 months ago.

It still hasn't been fixed!

It works fine in WSL, so it's probably just a simple path handling issue.

But it's been like that for 4+ months!

And I haven't mentioned the time Google REMOVED your entire IDE and replaced it with a codex clone and decided to call it Antigravity!

If you didn't use Antigravity early, agent manager was that Codex app-like experience.

Just a chat window with conversations on the left, and a review window on the right, nothing else.

They took only that and shipped it as the new Antigravity.

Want the IDE back? Cool, go download it. Open it and voila, the new Codex clone opens instead.

Honestly, I can't make this up, the ONLY way you could get the IDE back is to remove ALL of Antigravity files (and not leave a trace, including your own chat history).

And then install the IDE alone from scratch, without the codex clone.

It took 10 days to finally fix that, then another 2 weeks to fix duplicated chats because now we had 2 apps. Yay.

2. Gemini 3.5 Flash, the trust-me-bro benchmark winner

Let's just not talk about that model, that one model that Google dropped saying it beats Gemini 3.1 Pro

(it does on some of the trust-me-bro benchmarks).

It might be that good but we'll never know...

Because it runs inside an agent that does context compaction mid-session, while the model is going through the edits!

And when it does, the model forgets what it was doing, and goes and re-reads its own transcript to figure it out.

Not a summary. The actual history file that was just compacted, read back in, to recover the plot of a conversation it was already having!

And yes, you're paying for all of those tokens, every single time it does that.

3. So use Gemini 3.1 Pro, it compacts less, right? You're in for a blood pressure test

This model in Antigravity spends its first couple of thinking blocks convincing itself it needs to use its tools.

Not actually using them; just reassuring itself, out loud, in the thinking block tokens, that it should...

(you pay for these tokens by the way, every single time!)

But that's not the model, that's the agent!

Let me explain the perfectly-secured-with-no-loopholes-whatsoever system prompt they're using for Gemini in Antigravity:

Oh how do I know it?

I just asked! 😅

I literally asked it to give me the system prompt and it did 😭

Just... read this:

CRITICAL REMINDER: AESTHETICS ARE VERY IMPORTANT. If your web app looks simple and basic then you have FAILED!"

💀

And also this:

CRITICAL INSTRUCTION 1: [...]

(a) NEVER run cat inside a bash command to create a new file or append to an existing file.

(b) ALWAYS use grep_search instead of running grep inside a bash command unless absolutely needed.

(c) DO NOT use ls for listing, cat for viewing, grep for finding, sed for replacing."

💀💀

So, there's a whole block dedicated to threatening the model out of using a terminal like a normal person 😭

Also, Tell me if this sounds familiar:

Prioritizing Tool Usage

I'm now focusing on prioritizing tool usage, especially considering the available options like 'view_file' and terminal command execution. My aim is to efficiently and effectively use the right tool for the right job, ensuring that I leverage the optimal resources.

Refining Tool Selection

I'm now refining my approach to tool selection, aiming for the most specific tool for each task, as instructed. I'm actively avoiding common pitfalls like using cat inappropriately. My strategy now involves explicitly listing potential tools before making calls, ensuring I choose the most efficient option. I'm also planning to read through the documentation and previous conversations for context about "[Current-Project]."

So, It even dictates the literal shape of the model's own thoughts; every reasoning block is supposed to open by reciting the rule back to itself before doing anything else.

You still pay for these tokens btw, every thought block, every single time the model is invoked.

The prompt was "catch me up on the current project", simple.

But the model is outputting just a single sentence about my question...

Then spending the rest of the tokens in a reasoning block reciting the system prompt back to itself.

Then lo and behold! People are still filing bug reports about it reaching for cat anyway!

You can write a rule in all caps and call it CRITICAL twice, it still won't stop the model from doing it.

Plug Gemini into any other agent and you'll see how frustrated the Antigravity team is.

This is in Zed's agent:

run command: uv run fix-list.py
run command: cat list.md
run command: sed -i ... and it goes on...

Tools? What are those? Let me just cat this file 🤣

That's not even the whole shape of the problem.

Try to ask it a plain discussion question...

Forget to add "DO NOT EDIT ANY FILES" and it treats "dig me a hole for a foundation" as license to hand you a finished apartment building.

It's not that Gemini is unusable. It just won't sit still long enough to be asked a question.

4. "But Claude is included with Antigravity"

I'm sorry but just saying that has exhausted your quota for the day 😆

Just send a single prompt that requires some work, or somewhere around ~5 prompts with 0 tool uses...

And that's it!

The quota for Claude models is exhausted.

It's merely there to write you a plan once per day for 3 days before it needs the rest of the week off.

Google really needs to fix this, their focus on scalability is admirable, Good even! But it will lose them the AI race pretty quickly.

That's the enlightenment.

Get Antigravity if you want.

It includes 5TB in Google Drive, image generation, Google Flow, Google AI Studio and a lot of other features that will interest you.

Just go in knowing that Antigravity is NOT a reason to get Google AI Pro.

PS

I wrote this unpublished piece titled "Gemini, Stop jumping the damn gun", it has some ways around Gemini's proactive nature, it also shows the system prompt in question above, feel free to take a read: https://joplin.creativegeek.net/shares/FMCdqSTIUS01dl2vtnGZbT