My recent honest thoughts with Google Antigravity

Source: Dev.to

I know the recent quota change with Antigravity pissed off many people (based on countless Reddit posts). I first thought I took a hit when I could barely use the “Gemini High” model, but when I was forced to use “Gemini Flash”, it actually started to work better for what I’m doing.

I’m not saying this is true for everyone, but for me (and perhaps a few others—some Reddit posts mention similar experiences) this is the case, and I want to share how I’m getting the most out of Google Antigravity (with a Google AI Pro license). I think it’s still one of the best values out there (I have ChatGPT Plus and Google AI Pro and I’m pretty happy with this setup).

Ask Antigravity to generate documentation

You might think, “here we go again—Mr. obvious,” and I know there are many posts about having good agent markdown files, instruction files, skills, etc. My approach is slightly different. Whenever I want to implement a new feature (I’m creating a game using Phaser as the engine), I first ask Antigravity to document some prerequisite knowledge I want the AI to have.

This has two benefits:

1. The document is useful for me because I can’t possibly remember every detail.
2. The document gives the AI a quick understanding of what I want to do, allowing me to switch “rooms” when the context gets too full (more on this later).

Don’t be afraid to create a new chat

A popular Reddit topic about Google Antigravity is the error:

The model’s generation exceeded the maximum output token limit.

I hit this hard today. I was trying to keep the conversation going, and Antigravity attempted to break the response into smaller chunks, but it was brutal. My hypothesis is that at some point it ingested a very large JSON file (enemies.json). I don’t touch that file directly; instead, I have smaller files that a Node script combines into one.

What I did:

1. Asked Antigravity to generate a concise summary of what I was trying to do and what the new “room” needs to know.
2. Started a new conversation with that summary.

The experience was a night‑and‑day difference. If you encounter that error, first make sure you’re not feeding the entire context—humans can’t work like that either. Then create a new chat; it can save you a lot of time.

Conversation‑based approach works really well

It took me a while to realize, but Google Antigravity can really understand your intent. I used to craft lengthy prompts in ChatGPT and then feed them into Antigravity. When I saw how well “Nano‑Banana” (the internal name for Gemini Flash) grasped my intent, I tried the same directly in Antigravity—and it worked. I usually start by asking it to document how it works, which helps a bit, but for the most part it understands the intent quite well.

Don’t just say “do it” — review the implementation plan

It’s easy to say “looks good, go ahead and implement,” but the implementation plan often contains questions that need your input (similar to how Cursor asks a few questions, though Cursor does it in a more flow‑like manner). Antigravity includes these questions in the plan document itself, and they can be easily missed unless you pay close attention.

When I started reviewing the implementation plan and answering the questions it raised, the quality of the output improved noticeably. If you haven’t tried this before, give it a shot.

AI can only get more expensive

To summarize, I’m getting a lot done with Google Flash. My game isn’t rocket science, but it’s also not trivial. So far, Gemini Flash has assisted me well. I sense AI pricing could rise, and I hope the Google AI Pro license stays within its current price range. Considering it also gives more prompts to Nano‑Banana Pro and even includes 1,000 credits for Google Flow, I still think it offers great value (personal opinion).

Not the most recent demo, but an idea of what I’m building

(I made quite a bit of progress since this video, but this is the best English‑language demo available. The translation isn’t complete, so you may still see some Japanese characters. If I create another English demo video, I’ll post it here.)