Antigravity: the LLM does it better

Mon, 16 Mar 2026 18:00:00 +0000

Over the last few weeks, I’ve been writing about my experiments with Antigravity, or rather, with the (more or less) intelligent agents integrated into the editor. The results have been mixed: sometimes the agents proved to be very effective, accurately easing some complex or repetitive tasks; in other cases they didn’t accomplish anything worthwhile, only wasting a huge amount of time.

Coding badly with an LLM

Among all the more or less useful applications of LLMs, one of the most controversial is the idea that anyone can use an LLM to program, even when they don’t know the first thing about it. This is the so-called vibe coding, a concept that can be summarized more or less like this: “I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works”. This is not a quote from Nanni Moretti, a famous italian actor and director; it was reported by Andrej Karpathy, one of the founders of OpenAI.

Now, setting aside the many issues regarding security, the maintainability of code produced without really knowing what it does, and the regression to the mean, which here implies that the quality of AI-produced code will inevitably tend to collapse, it only takes a bit of personal experimentation, even with simple projects, to realize that the very idea of coding for everyone doesn’t work. If you want to get something good out of AI, you have to know exactly what you want and how you want it.

I tried vibe coding myself: I asked Gemini to write a driver for an Arducam camera, explaining only the bare essentials and waiting for it to do everything on its own.¹

However, this approach did not work. It took hours upon hours of work, errors, inconclusive answers, corrections, and advice from me to reach an acceptable result.

And this was only possible because I eventually abandoned the idea of letting Gemini do everything and provided it with specific advice on how to proceed. Without that, it would have only been a massive waste of time.

Mimmo also tried, on his own, to have Lumo write a Micropython program to manage my Arducam Mini 5MP Plus with the Pico. Even Lumo, which is undoubtedly one of the best chatbots around, couldn’t do it.

– Image generated by Google Gemini.

Coding better with the same LLM

This doesn’t mean that an LLM can’t be useful when programming; you just have to use it properly. You must explain exactly what you want it to do, provide all the necessary materials, and set boundaries within which it must operate.

In short, you have to treat it like a graduate student or an intern, with little experience but a great desire to learn.

It only took two carefully written prompts to make a difference. The first defined the project’s directory structure and instructed the agent to use git for version control and uv to install any necessary Python packages,

Prompt

Start a new project and name it project13-pico-camera5mp-reprise.

Read the AGENTS.md file that defines your basic behaviour.

Create a new directory with this name and initialize an empty git repository using git init Use uv for managing Python packages. All the code must stay in the code directory, with no subdirectories within it. This project does not need a data directory and the output directory must be renamed as images.

while the second prompt contained everything I could think of to help it work.

Prompt

I have a Rasperry Pi Pico 2W connected to a Arducam 5MP Plus OV5642 Mini Module Camera Shield (https://www.arducam.com/arducam-5mp-plus-spi-cam-arduino-ov5642.html).

I have connected the camera to the Pico using the standard layout found in the documentation:

Camera	CS	MOSI	MISO	SCK	GND	VCC	SDA	SCL
Pico	GP5	GP3	GP4	GP2	GND	3V3	GP8	GP9

The main documentation for the camera can be found at these links:

The code to drive this camera with the Pico is written in C (i.e., for Arduino) and CircuitPython, and can be found here: https://github.com/ArduCAM/PICO_SPI_CAM. You can download all the files you need from this repository. In particular, the files for CircuitPython are in the Python directory:

Arducam.py, which contains the functions to drive the camera,
OV5642_reg.py which, as far as I have understood, defines the resolutions allowed when taking pictures, and
boot.py , but I don’t know what it does. In the same directory there is also the script ArduCAM_Mini_5MP_Plus_VideoStreaming.py which sends the video captured by the camera to a Windows application and that is of no use here.

Please write a basic CircuitPython script that uses the functions defined in Arducam.py to test that the camera works and can take photos.

These two better-crafted prompts allowed Gemini 3 Flash to quickly generate a CircuitPython script that runs directly on the Pico and performs all the requested tasks.

Actually, the very first version of the script didn’t work. But Gemini was able to modify just a single line in the Arducam.py driver provided by the manufacturer to fix it (specifically, it removed a zero from the frequency value).²

– The line highlighted in blue shows the only change that Gemini made to the Arducam.py file; the original file is on the left, and the Gemini-modified version is on the right.

All of other files provided by the manufacturer remained intact.

As it turned out, the manufacturer’s driver worked perfectly with the Pico. Gemini simply learned how to use it to write the script that captures the image directly from the microcontroller (with a bit of patience, I could have done that myself).

Guide the LLM or let it run on its own?

The difference with respect to the previous approach based on vibe coding is striking.

Then. Gemini worked frantically for hours and created such a mess that I was forced to restart my Mac twice. Meanwhile, it had convinced itself (and me, too) that the Arducam.py driver was incompatible with the Pico and thus rewrote it from scratch, discarding an essential library for communicating with the camera (the library marked by the red arrow),

– Changes to the first few lines of the Arducam.py file; original file on the left, Gemini’s modified version is on the right. The arrow marks the library that was removed by Gemini.

and removing most of the functions defined in that file for no reason and without any real gain in efficiency or functionality (quite the opposite, actually!). The image below shows the difference in the length of the original code, on the left, compared to the modified version, on the right.

– Changes to the Arducam.py file. The red rectangles highlight the difference in the number of lines between the original code (left) and Gemini’s modified version (right).

Not satisfied with that, Gemini also completely altered the OV5642_reg.py file containing the camera configuration data, claiming memory issues. These problems evidently didn’t exist, given that everything went smoothly in this second experiment using the manufacturer’s original file.

– Changes to the OV5642_reg.py file, containing the camera configuration data. The original configuration (left) was overridden by Gemini (right).

What About Arduino?

At this point, it was inevitable to ask Gemini to generate the Arduino driver, using the same information that I had already provided for CircuitPython.

At this point, it was inevitable that I would ask Gemini to generate the Arduino driver using the same information that I had already provided for CircuitPython.

Gemini struggled more this time, mostly because I had forced it to make only minimal changes to the original Arducam libraries. However, this constraint was impossible to satisfy since the libraries do not natively support the Raspberry Pi Pico.

Response

Researching Arducam Library Compatibility I’ve thoroughly inspected both the original ArduCAM library and the Arducam_mini library. Neither version contains native support for the RP2040/RP2350 (Pico) architecture. Specifically, both libraries are missing the necessary definitions for regtype, regsize, cbi, and sbi for this platform, which is the root cause of the compilation errors.

Once the constraint was relaxed, things went smoothly, and in a short time, the driver modified specifically for the Pico and the related image acquisition program were ready.

Free or paid?

Many argue that paid models are much better than free ones. Consequently, my vibe coding problems would supposedly have disappeared if I had paid for Gemini or Claude.

That’s not exactly how it went. During my testing, I mainly used the free version of Gemini to see what would happen to a non-professional user playing with these tools. However, when I ran out of tokens, I switched to the paid version of Gemini provided by the university, but I never noticed any significant differences. I can’t say much about Claude, because paying 200 dollars a month to use it seriously makes no sense for someone like me, who isn’t a professional programmer.

What I did notice was that Claude Opus and Claude Sonnet were much better than Gemini at solving the problems that arose during code development. However, since I only used them once the work was already in progress, they might have just been helped by what Gemini had already done.

Conclusions

It goes without saying that neither this experiment nor the previous one are scientific. For them to be considered as such, I would need to use different types of microcontrollers and modules, define a well-defined set of prompts, and measure how the agent responds to changing experimental conditions.

Despite this, the experiment has one merit: it is fully reproducible. Anyone can repeat the experiment using the same low-cost hardware and prompts to verify whether my findings are correct.³

This is something that the vibe coding apostles with tens of millions of views always carefully avoid proposing. Their posts are full of certainties and bold predictions yet they never provide any evidence to support their claims.

Today, no one can deny that LLMs can be a great help during programming. However, it’s hard to believe that they can truly do so without any human intervention. At least until those apostles decide to provide us with clear, repeatable data to support their claims.

Although I know a little bit about programming, I had no idea how to develop a driver and didn’t want to spend days learning how to do it. ↩︎
Actually, this file is a library of functions. However, since microcontrollers don’t have an operating system, these functions access the hardware directly, just like the drivers that control the components of a computer. Therefore, in this case, the terms library and driver are interchangeable. ↩︎
The term low cost must now be taken with a grain of salt. The shortage of electronic components has caused the price of the Pico to skyrocket, and even more so for the Arducam Mini 5MP Plus. The latter, which cost 40 dollars a few months ago, now costs more than triple that on Amazon or AliExpress. Fortunately, some retailers are maintaining more reasonable prices, at least for now. ↩︎