Antigravity / VS Code / Claude Code / Cursor
The last three episodes were all concepts.
Next episode we start building the first web page.
But before that,
you need the right tools on your computer.
I'll be straight with you here:
these are the tools I've used myself,
lined up worst to best in five stages,
plus where a complete beginner should start in 2026.
You can, but it hurts.
Every single change
means another copy and paste.
Once you have a lot of files, you lose track of which piece is in which file.
So here is the core of this episode:
pick a tool that lets AI work directly inside your computer.
These five get talked about the most.
I've used or followed all of them,
so here's what each one is like.
Google's AI dev environment, launched in early 2026,
built around running several agents at once.
The built-in model lineup is serious:
Gemini 3.1 Pro (High / Low), Gemini 3 Flash,
Claude Sonnet 4.6, Claude Opus 4.6, GPT-OSS 120B,
pretty much every top model today is in there.
What makes it different is that agents get three surfaces to act on:
the editor, the terminal, and the browser,
and you can send five agents on different tasks at once,
then watch who is doing what in the Manager view.
2026-03 v1.20.3 added AGENTS.md rule files and MCP integration,
so it can connect to GitHub, databases, and external APIs.
Pricing (changed since 2026-03, straight talk):
・Free: rate limited, fine for a light trial
・Pro: $20/month (same as Cursor)
・Ultra: $249.99/month (heavy users)
・Pay-as-you-go: $25 / 2500 credits
The community complained about the increase after the 2026-03 change,
so check antigravity.google/pricing for the current numbers before you decide.
Cursor is Anysphere's AI IDE.
Cursor 3 shipped 2026-04-02,
with the interface rebuilt from scratch
and the core idea moving from writing code to managing agents.
What's new:
・Agents Window: several agent chats open side by side
・Parallel runs, and cloud agents pick up the work without breaking the thread
・Design Mode (a built-in design tool)
・Built-in git and multi-repo support
・Improved diff and PR workflow
Pricing didn't change, Pro is still $20/month.
It's the hottest tool right now, the biggest community, the most tutorials,
and nine out of ten vibe coding videos on YouTube use Cursor.
VS Code is Microsoft's code editor,
free, and the most used one in the world.
Claude Code is Anthropic's official AI coding tool,
and it runs in the terminal (the command line).
Paired with VS Code,
it's the combo behind my first three products (Crispy Kingdom 脆脆王國 / NaLi Match / Flame Man).
The biggest difference from Cursor:
Claude Code is terminal-driven,
so you get a black window with white text
where the AI talks with you and edits your files.
Still shipping updates in April 2026:
a new /powerup interactive tutorial,
the MCP tool output cap raised to 500,000 characters,
/cost breaking spend down by model,
and file diffs 60% faster.
The learning curve is steeper though,
because you have to get used to talking to AI inside a terminal.
OpenAI's answer to Claude Code.
CLI, macOS desktop app, VS Code extension, and web are all connected,
so whichever end you use,
sessions and long tasks carry over.
The main model is GPT-5.2-Codex (released 2026-01),
SWE-Bench Pro 56.4%, Terminal-Bench 2.0 64.0%,
the reference model right now for long tasks and cross-file refactors.
There's also GPT-5.1-Codex-Max,
which uses 30% fewer thinking tokens at the same strength,
so it costs less.
Pricing change, this one matters:
From 2026-04-02 Codex bills by token instead of per message (Plus / Pro / Business).
Short tasks got cheaper,
but watch your usage on long tasks and big files.
Some users reported a single prompt eating 7% of their weekly limit.
The 2026-04 CLI also added a Windows sandbox,
ChatGPT device code login,
and codex exec support for prompt + stdin.
I used the Manus cloud agent version myself.
That version could run long tasks,
but two things made me switch:
1️⃣ It couldn't touch every file on my computer
Output still had to be moved back into the project by hand.
2️⃣ Credits burned fast, and it got expensive
Manus was credit-based back then,
one complex task ate a big chunk of the balance,
so experimenting meant topping up again and again,
which doesn't work long term for a solo developer.
But Manus turned things around after that:
Manus also launched a personal agent inside Telegram.
Scan a QR code and it joins your chat.
WhatsApp / LINE / Slack are officially very soon.
Pricing also moved from pure credits to a mix of monthly fees and subscription tiers,
which is much friendlier if you plan to run it long term.
Still, check the official page before you commit,
because the pricing may move again.
In other words,
the thing I complained about back then, not being able to touch files on my computer,
got fixed after Meta took over.
If you want AI to work across your apps rather than just write code,
the new Manus is worth watching.
The five above all install on your computer.
If you haven't even opened a computer yet, or you just want a prototype in 5 minutes,
these four vibe coding tools run in the browser alone,
each good at something different.
A web tool from StackBlitz.
One prompt gives you a running React/Vite project,
with live preview, a terminal, and one-click deploy to Netlify built in.
Give it a text description or a design file,
and it returns production-ready React + Tailwind + shadcn/ui components.
A genuinely end-to-end tool from a European team.
UI, backend, Supabase database, and auth in one line,
with one-click push to your own GitHub.
From prompt to a site running on the public internet, all inside the browser,
with hosting built in, so you get a URL the moment it's done.
Another 10 tools, where they stand, their pros and cons and who they suit (including the ones renamed, acquired, or shut down)
are collected in this series' tool library:
/series/tools/
Beginner-friendly UI, a full model lineup (both Gemini and Claude),
and a free tier that is rate limited but enough to tell whether you like it.
Move up to Pro at $20 once you're sure you'll stay.
After installing Antigravity,
add two extensions:
If you already subscribe to Google AI Pro / Ultra,
sign in to Antigravity with the same Google account,
which gives you a higher agent quota and more models.
Three months in, once you're comfortable,
you'll have your own preference.
That's when to think about moving to Cursor, Claude Code, or Codex.
Here's the order I'd go in:
The detailed install steps for each tool
are all over the web,
so here I'll just give you the shape of it:
1. Sign up on the Google Antigravity site (log in with your Google account)
2. Download the build for your system (Mac / Windows)
3. Open it and sign in with your Google account
4. In the Extensions marketplace, install the two must-haves:
・Claude Code extension
・Chinese-language extension
5. Create a new project folder
6. Open the chat window inside it and talk to the AI
1. Download VS Code from code.visualstudio.com and install it
2. Create an Anthropic account and subscribe to Claude Pro or Max
3. Install the Claude Code command line tool in the terminal
4. Open your project folder in VS Code
5. In the VS Code built-in terminal, type claude, then start talking
1. Download and install from cursor.sh
2. Create an account (the free plan is limited, the paid one is Pro)
3. Open it and sign in
4. Open a new project folder and press Cmd/Ctrl + L to bring up the AI chat
For the detailed steps, tell the AI: I want to install Antigravity / Cursor / Claude Code on a Mac, walk me through it step by step.
It'll be more accurate than any tutorial online.
Don't do that.
The differences between the three tools,
for someone who hasn't built a first product yet,
come down to all of them being fine.
Pick one (I'd say Antigravity),
install it,
and tomorrow you can build the first web page from EP.5.
The concepts in EP.0~3 are the foundation,
the tools are just the carpet.
You can swap the carpet any time.
The foundation decides how tall the house can go.
Next episode we get our hands dirty.
In 2026 you won't be writing HTML / CSS / JS syntax yourself.
One sentence to the AI and it's there.
So EP.5 teaches something quite different:
why does the page the AI wrote look like that?
Why are HTML, CSS, and JS split into three layers?
Once you know what each layer is for,
you can point the AI at exactly the piece you want changed.
Add a button, make the button blue, pop up a message on click
are three completely different conversations.
EP.5 is the smallest set of knowledge for speaking the same language as the AI.
So when you give an instruction,
the AI gets it right the first time.
Yes. All of them are built so you give instructions in plain language and the AI writes the actual code. Beginners can use them. The difference is the learning curve and the interface style.
Both start at $20/month, so the price is the same, and both have a beginner-friendly UI. Cursor has the biggest community and the most tutorial videos, so a beginner finds answers more easily. Antigravity's edge is the model lineup (both Gemini and Claude) and running several agents in parallel.
The biggest difference is the interface. Claude Code is terminal-driven, so you get a black window with white text where the AI talks to you and edits your files. The learning curve is steeper, but it handles the most complex tasks, and 1M context plus subagents are its strong points.
Same class of weapon, different camp. Codex is OpenAI's answer to Claude Code, and its main model GPT-5.2-Codex is strong on long tasks and cross-file refactors. If you already pay for ChatGPT, use Codex. If you don't and you want Claude models, pick Claude Code.
You can, but every change means another copy and paste, and once you have a lot of files you lose track of which piece lives where. Building a product depends on letting the AI work directly inside your computer, and the ChatGPT web app can't do that.