The site in Antigravity: plan, 12 stages, one commit

In the previous post I wrote about the address and the choices: domain, Hugo, GitHub and Cloudflare. This part is different. The pieces were already on paper. What was missing was the site existing on the computer.

What I wanted to solve

On the morning of 3 October 2026 I opened Google Antigravity. I work during the day, and at night and on Sundays I build with AI agents. That morning the job was this: turn the diary decision into real files.

Antigravity is Google’s code editor with AI. I operate it. The model suggests and writes. I read, click, and approve. Darth, the main bot, guided and checked from the screen.

There had been an update with new models: Gemini 3.8 Flash, Claude Opus 5.5 and Sonnet 5.5, among others. I use the Gemini Pro subscription through Antigravity. I pay for the tool. I am not reviewing Google from the outside. I am telling you what happened that morning.

The request I typed was short. Do the work on your own. No administrator password. Show the plan before any code.

I did not want a huge block landing on the screen, with me hitting yes without having seen the path.

Quick summary

• On the morning of 3 October 2026, from 08:37 until around 10:31, I built the site in Antigravity.
• Claude Opus 5.5 proposed the Hextra theme and a plan of 12 stages.
• Each stage ended in a commit, a save point.
• Hugo v0.167.0 checked by checksum. Hextra v0.13.0 pinned, so the version would not change on its own.
• At stage 5 the model became unavailable. I switched to Gemini 3.1 Pro, told it to continue from where it had stopped, and nothing was lost. Around 09:56 the 12 stages were done: the site compiled with 0 errors, Portuguese at the root and English at /en/.

First: what Antigravity is

Antigravity is a code editor with artificial intelligence. You describe what you want. The model proposes files and commands. You stay on the keyboard: accept, refuse, ask again.

Picture a carpenter with a talkative helper. The helper measures, suggests the cut, even holds the piece. You are the one who starts the saw.

The real mechanism is different. The program sends your request and the state of the files to a model. The model returns text: code, a list of steps, a command. The analogy stops here: a shop helper does not invent a wall. The model invents with confidence. That is why the plan comes first, and approval stays mine.

First: what a commit is

A commit is a save point in the history of the code. Each one keeps a snapshot of the folder at that moment, with a sentence saying what changed.

Picture a save in a game. You finish the level, you save, and if the next one breaks you do not start from zero.

Under the hood, Git (the history program behind GitHub) records that slice. The analogy stops here: a game save often overwrites the previous one. The old commit stays. If a stage ruins the look, you can go back to the one before it without guessing what used to be there.

The rule I asked for was this: 12 stages, and one commit at the end of each.

First: what a checksum is

A checksum is a short signature calculated from a file. Change one letter in the file, the signature changes.

A program reads the bytes and returns a string. I used that on Hugo v0.167.0. I did not copy an installer that “looked close enough”. I checked the signature.

First: what a Hugo theme is (and Hextra)

A theme, in Hugo, is the look that comes ready: type, menu, post page, light and dark mode. Hugo reads the texts and puts those clothes on. Without a theme, you would have to draw each page.

Claude Opus 5.5 proposed Hextra. In that proposal, Hextra is bilingual, has search that runs in the browser, has light and dark mode, and is fast. PaperMod, Blowfish and Docsy were left out. I will not invent the case against each one. They were left out. What came in was Hextra, version v0.13.0, pinned.

Pinning the version, in one sentence, is nailing the number: the site does not wake up tomorrow with a new Hextra nobody asked for.

Search in the browser means the lookup happens on the visitor’s computer, not on a server digging through a database. That fits a static diary. On the site, Portuguese sat at the root (rodrigoesporto.com) and English at /en/.

What I used

I used Google Antigravity, with me at the controls. I used Claude Opus 5.5 for the plan and the start of the work. I used Gemini 3.1 Pro when the other model dropped. I used Hugo v0.167.0, checked by checksum. I used the Hextra theme v0.13.0, pinned. Darth guided and checked from the screen.

Login and password stay out of this text.

How it went

I was not typing every line of the theme. I was in charge, with a plan in front of me.

If you want to copy the logic, not the brand, the path is this:

1. Open the editor you operate yourself. In my case, Antigravity.
2. Write the request with these rules: do it on your own; no administrator password; show the plan before any code.
3. Read the plan. Accept or refuse the theme. I accepted Hextra. PaperMod, Blowfish and Docsy were out.
4. Break the work into small stages. Each one ends in a commit. Mine were 12.
5. Lock the versions. Hugo v0.167.0 checked by checksum. Hextra v0.13.0 pinned.
6. Keep a second model ready. At stage 5 the one that was working became unavailable. I switched to Gemini 3.1 Pro and told it to continue from where it had stopped.
7. Only call it done when the generator itself accepts it. Around 09:56 the site compiled with 0 errors, with PT at the root and EN at /en/.

When it works, the screen does not set off fireworks. It just stops listing errors, with a compile of zero errors and both languages in the place we agreed.

What went wrong, and what I would do differently

It went wrong in the middle, at stage 5: the model became unavailable.

The conversation stopped. If the whole job had been one block, I would have been staring at a half-file and trying to remember what was already done. Because each stage had already become a commit, what was saved was still there. I switched to Gemini 3.1 Pro. I told it to continue from where it had stopped. Nothing was lost.

The lesson I would leave for someone starting out fits in one line: ask for the plan before the code, save at each stage, and keep a second model as backup.

I would do that again. What I would not do is start with code to save time. You lose time when you have to undo a piece and nobody knows where it begins.

None of these tools replace judgment. They shorten the path between “I know what I want” and “it is live”. That morning, “live” still did not mean published. It meant the site compiling in front of me, with no error.

Costs and limits

In this part I do not have a dollar amount to report. I already used the Gemini Pro subscription. No checked figure for that morning, no number.

The cost I can measure is the clock. I started at 08:37. The 12 stages closed around 09:56. The Antigravity session ran until around 10:31. That is in the record of that morning, 3 October 2026.

What this way of working does well: a personal site, two languages, stages you can rewind, without drawing the look from scratch.

What it does not do: it does not put the site on the internet on its own. It does not replace reading the plan. It does not survive an “accept all” if the request was vague. And the model can simply become unavailable in the middle of stage 5. That happened.

This is what I did that morning, not a lab test of Hextra or Antigravity. A plan, small stages and commits let me switch models without erasing the work. It is not a published study.

Who it is for, and who it is not for

It is for you if you are going to build a site with an AI editor and you are not a professional programmer. I never programmed professionally. The method does not ask you to become one. It asks you to read the plan and refuse what you did not understand.

It is not for you if you want to send a paragraph and walk away. Without a plan, without commits, without a second model, the tool will still move. The problem shows up when it stops, or when the look comes out wrong and there is no earlier snapshot.

It is also not for you if the request includes an administrator password to make it easier. Easier, here, is opening a door you did not need to open. Everything in this build was asked without that password.

Closing

I wanted the diary to exist on the computer, without a blind block of code. On the morning of 3 October 2026 I asked for the plan first. There were 12 stages, one commit each, Hextra v0.13.0, Hugo v0.167.0. The model dropped at stage 5. The second model held. Around 09:56 it compiled with 0 errors.

In the next post I tell the rule of understanding before you approve: what it does, why, and whether it is safe.

If you want to follow the journey, I am on X: @RodrigoesPorto.

Sources and links

• Hextra: https://imfing.github.io/hextra/
• Hugo: https://gohugo.io/
• This site: https://rodrigoesporto.com/