Antigravity: Beyond the Basics of AI Coding

Writing software with AI can be as simple as it is complex.

In this series that I called Antigravity: What Others Don’t Tell You, I want to explore concepts that are essential for our daily workflow, using Antigravity as our primary tool for agentic development.

We often think we can just write a medium-sized list of requirements, feed it to an AI, and get a perfect result. While this might work for simple solutions, it’s a bit like a junior developer from ten years ago that writing, copying, and pasting untested code that “just happens” to work.

Today, I want to show you how to manage development more efficiently by reusing patterns that actually work. I recently had the opportunity to take an agentic coding course at my company, and I’ve translated that knowledge into the Antigravity ecosystem. However, these principles are universal; you can apply them to any AI tool like Codex, Claude Code, or GitHub Copilot.

Antigravity

If you haven’t installed Antigravity yet, you can do so via this link or by following the official Google documentation. Antigravity is essentially a VS Code-like IDE, integrated with advanced features Google has developed for us.

The standout feature is the integrated chat, sitting right next to your code. It’s not just a generic chat; it’s a contextualized chat. This brings us to a crucial topic: Context.

Google provides us with free models that we can use:

• Gemini 3 Pro (high)
• Gemini 3 Pro (low)
• Gemini 3 Flash
• Claude Sonnet 4.5
• Claude Sonnet 4.5 (thinking)
• Claude Opus 4.5 (thinking)
• GPT-OSS

Obviously, these have usage limits and must be used intelligently depending on what we want to achieve, the speed required, and the quality of our data.
I asked Gemini to compare them and generate a table:

Note: These models have usage limits. Use them intelligently based on the required speed, quality, and complexity of your data.

Understanding Context

Context is the set of information we provide to a model alongside our query to get a relevant answer.

Naturally, the larger the context, the longer it takes for the model to respond. It is best to minimize context whenever possible to ensure fast, coherent responses. A bloated context can lead to hallucinations, a phenomenon often called Lost in the Middle – where the LLM loses focus or gets confused by the volume of data.

How do I reduce context?

Some agents handle this automatically. Others use “compact” instructions to summarize previous history, as the more we chat, the more the context grows. This concept is called Tokenization.

A token can be a character, a word, or a part of a word (like -tion), depending on the model. For example, GPT-4, a model behind ChatGPT, breaks the phrase “I can’t wait to build AI applications” into nine tokens, as shown below

The JSON vs. Markdown Trick

A great way to understand how many tokens we are using is by using this tool tiktokenizer or a similar one.

Tiktokenizer shows us how many tokens we are using, but not how we could save them.
Suppose, for example, we have a JSON file

[
  {"id": 1, "status": "active", "label": "low_risk"},
  {"id": 2, "status": "pending", "label": "high_risk"}
]

This format counts as 44 tokens.

Convert to Markdown

By converting to .md (specifically tables or lists), you almost always save tokens because you eliminate the syntax overhead (braces and repeated keys).

Avoid “TOON” Formats

While newer compression formats like “Toon” offer significant token savings, they aren’t always your best ally. LLMs are primarily trained on vast datasets of natural language and established standards. Moving away from these “standard” patterns in favor of niche formats can actually degrade the model’s reasoning performance.

Tips and Tricks

As a final tip for this first installment of Antigravity: What others don’t tell you, I highly recommend the Antigravity Cockpit extension. It provides an intuitive dashboard with infographics to monitor your remaining usage limits for each model.

See you in the next one! 👋🏻