Surrendering to the “English” Compiler

Andrej Karpathy called it back in 2023: “The hottest new programming language is English”

Two years later, he was right. Software Engineers faced a silent revolution and that’s just a beginning. They aren’t just writing syntax anymore; they are managing intent. This shift – colloquially dubbed “vibe coding” – is about surrendering the implementation details to the AI while you focus on the architecture. But as we close out 2025, the tool landscape has split into three distinct “vibes,” and choosing the wrong one can turn your flow state into a maintenance nightmare.

Here is my take on the current stack, from the reliable workhorses to the bleeding-edge agents.

1. The Daily Driver: Cursor

If you want to actually ship software today, Cursor is still the first choice. At least, It remains my personal go-to.

While other tools are flashy, Cursor’s “Composer” feature (especially the multi-file editing) feels like a natural extension of your brain. It indexes your local codebase so you don’t have to copy-paste context, and it runs on a fork of VS Code, so you don’t lose your extensions.

• The Vibe: “Flow.” You type, and it predicts the next logic block, not just the next word. The perfect solution for developers who still want to drive and treat AI as an assistant, not a ghostwriter.
• The Risk: It’s easy to get lazy. The “Accept All” button is dangerous if you stop reading the diffs. You can try to ensure quality by adding a mandatory code review stage to every small feature implementation, but you are still at great risk of losing that vital mental connection with your code.

2. The New Challenger: Antigravity

I’ve always liked Google’s engineering, and their new Antigravity platform (launched November 2025) is finally giving me a reason to look away from Cursor.

I tested it on release day. My brother, a professional HEMA fencer, asked me to figure out who was showing better performance: the US or Europe. That request turned into a tiny web app using Python to fetch the data and native PHP for the frontend.

The verdict? The new Google Gemini 3 model doesn’t deliver a miracle. Without a programming background, Antigravity – just like Cursor or Claude – is pretty useless. The initial implementation on my app was broken on several levels, and it took me a few hours of back-and-forth about the project’s core logic and debugging to make it work right.

On the flip side, Antigravity is actually another fork of VS Code, just like Cursor. This means your extensions, themes, and muscle memory migrate in about 30 seconds. The main pro today is that it’s free, meaning the casual developer won’t hit free tier token limits. While it still lacks a few useful Cursor functions, it is a strong underdog. Knowing Google’s resources and traction in the AI battle, it’s going to be a fierce competitor.

• The Vibe: “Orchestration.” It uses the new Gemini 3 model, which is currently topping the leaderboards for agentic reasoning.
• Interesting Feature: Artifacts. Instead of scrolling through logs, the agents generate clear plans and screenshots of what they did. Fun at first, though a bit annoying after a day.

3. The “Fresh” Option: Claude for Desktop

I love Anthropic’s models – Opus 4.5 is arguably the smartest “brain” available – but Claude for Desktop feels like a beta product. Actually it’s unfinished wrapper around Claude CLI lacks plan mode and other useful stuff. There’re tricks – you are still able to use it in CLI mode and the core Claude functionality returns. But you spend too much time managing the tool and not enough time building the app. It’s a great research assistant, but for building a product, it feels disjointed from the code.

• The Vibe: “Friction.” Great intelligence, but the “hands” are clumsy compared to Cursor or Antigravity.

4. The Trap: Web-Based Prototypers (Bolt, Lovable)

Tools like Bolt.new and Lovable are magic for the first hour. They run entirely in the browser and let you “vibe” a React app into existence in seconds.

• The Verdict: Avoid for serious work. They suffer from the “Ejection Problem.” Once your app needs a real backend or complex logic, you hit a wall and have to export messy code. For now, they are toys, not tools.

Final Ratings (Ivan’s Scale)

1. Cursor: 9.5/10 – The professional’s choice. At least when programming is your instrument, not an actual profession.
2. Google Antigravity: 8.0/10 – The exciting new future and the strongest underdog of 2025.
3. Claude Desktop: 8.0/10 – Smartest brain, unfinished body.
4. Web Prototypers: 4.0/10 – Great for demos, terrible for products.