About

The Website #

This site can change quite frequently.

The following sections contain recommendations for the code I want to share, without positioning myself as an expert.

I’ve had mixed results with LLMs (Large Language Models), and the bad surprises often came from poor prompts or bad tool choices, but rarely from the limitations of the technology itself.

Test #

My personal websites serve as playgrounds for experimenting with AI while progressing on the frontend side.

Good practices are generally the same as on the backend.

Some of these may seem obvious, but it’s easy to overlook them.

Be Precise #

Prompts like “improve the code” or “change the style” are way too generic and will often worsen the results, even causing nasty errors.

Small Goals #

Work in small iterations on well-defined features like “add a registration form and authentication with a Google account.”

Such approach also helps you write better prompts in general.

Good Vibes #

Vibe Coding is not evil.

However, in its current form, the code produced by these platforms is often hard to evaluate, especially if you’re a beginner, and can even hide significant errors under the appearance of a result that “almost” works for you.

In this sense, it’s sufficient for a quick prototype, but quite risky for a final production product.

It’s worth noting that there’s a clear quality difference between the free offerings and paid plans.

Not all tools are equal either, depending on the goals you’re trying to achieve, and it’s the same for tools specialized in coding.

Automate #

We didn’t wait for AI to script everything we could, saving us from tedious manual tasks.

Some teams even develop using automation platforms that existed long before the hype.

However, these features are now integrated directly into coding platforms like GitHub (or others), and they can perfectly recognize your scripts, improve them, create new ones, and do all of this much faster than you.

Tools like Copilot or Jules integrate mini-script creation into their detailed plans.

This allows them to test and automatically fix errors generated by their own modifications without us having to copy-paste the error into the client (e.g., Claude Desktop).

You must supervise and validate all operations, of course, and errors are quite frequent, but with smarter use, the gains may become significant.

Some Resources #

The list is limited, for now, and may change significantly, but here are some links you can bookmark:

• Prompts GPT for Engineers
• Llama Models
• Shai