I usually struggle a lot to complete the things I start. As a matter of fact, this blog is quite a good example: very little content gets published, and lots of drafts pile up.

Similarly, I have an entire folder of unfinished side-projects that really hurts my brain.

Nothing crazy or revolutionary in that folder, obviously - Just small projects that only needed time and focus to get them done or decide to drop them (evaluating needs effort).

The point has always been that: experimenting and gathering conclusions.

Then, AI came along, and got better at basically everything computer connected, especially in the last 6/8 months of 2025.

As a software developer with a very bad focus and momentum issues, I needed to understand 2 things about AI

1. Hot or Not

Is AI sustainable on the long run? What’s the direction it is getting and how it affects/will affect my life? Is it a scam?

2. How AI can be useful to me?

As it evolves fast, and some premium (aka paid) tools and models exists, how can I use those tools to help me?

I figured that the second point may be the only benchmark to know the answer the the first point.

Prompting my way out

The only thing left to do was trying it out, using the pile of unfinished projects and draft ideas as a benchmark.

First of all, most of my needs are code oriented. General Purpose AI Chatbots (ChatGPT, Gemini, Perplexity) are useful for spot on questions and things that doesn’t need a long conversation to be fulfilled (IMO).

Since I’m a cheap f**k, and given the fact that most of my code is only useful to me and may end up being pushed in a public repo on my GitHub page (so, little privacy concerns), picking up which model/provider was pretty straight forward: Z.AI with GLM-4.6/4.7

I needed a more specialized tool than a chatbot. I tried Droid, Open Code and then settled to Claude Code. I was baffled by the amount of additional knowledge a user (not a developer) needs to use these tools: SubAgents, Skills, Hooks, MCP Servers, Plugins, Context window etc…

By the way, If you are interested in knowing more spending less time, here is a very good video by The Gray Cat which explains it all magisterially

Then, thanks to my brain, I had to set a deadline to keep myself accountable: 3 projects (2 from an existing codebase, 1 green field, only drafted idea), 2 weeks.

Project 1: Home Assistant Integration for my city local public transport

This was an existing project I tried to kickstart several times in different ways, including a standalone API in Typescript and a partially working setup for the Home Assistant Integration (Python) that connected to that wrapper API.

Since reversed the Public Transport Company API part myself first (no public API is available) and the reversed logic (quite simple actually, but I don’t think the “stock” Claude Code may know how to find that way - Maybe an Agent with skill could help?) was documented in the Typescript API, I decided to use both of them to get the model up to speed and well aligned during the first /init

After that, I was literally 4 simple prompts away from a working prototype (with issues of course, but I blame Home Assistant for the lacks of a good developer oriented documentation to how to write an extension)

I had to perform several fixes manually of course, but it removed the edge of it.

Repo is available here, working and still rough but… that’s what I needed months ago and now I have it. Plus I can always improve it if I need it (or somebody else could - Contributions are always welcome)

Read more: https://github.com/LBRDan/anm-hass

Project 2: MP4 Atom/Box visualizer webapp

At Spreaker we are working a lot with mp4s. You know, video container format that has a standard with other standards extended by other standards. And you may want to stick with one. Unfortunately, I need to see clearly what am I working with. Like literally. I needed to see the MP4 file composition to ensure I wasn’t screwing up something in my code by accident (like reading the version field instead of the reference_ID one from tfhd boxes)

This was a brand new project for a tool, with a clear purpose. It was easy to compose a PRD, since I was the original owner, and Claude helped there too.

I also tried a different thing. Since I was already using 3 different existing tools (MP4Box.js Filereader, hexed.it and Online MP4 file parser), I tried to feed them as a reference for Claude (which got it partially). Each tool has its own limitation, basically I wanted to squash them into one, removing all the extra stuff I didn’t need.

Here is the result (Github repo is still private - Gonna open it as soon as I am sure it doesn’t affect/harm anybody in any way):

Read more: https://boxxy-indol.vercel.app/

Again, still iteration and some known bugs to fix (not to me, since I reached the utility point I needed) - Like navigating a huge 1GB+ file can get slowish or unstable

Some of you may think that this could have been a simple CLI tool written in `insert_language_here` that actually highlights with terminal colors the bit in the input file buffer, but then it was (at least for me) harder to distribute and adopt. Plus, I’ve always had an hard time building interactive UI inside a terminal window (last time, it was during my last university “try” and I had to use C + ncurses lib… I’m sure now it’s way easier)

Project 3: Business info website for a family member

Not gonna share the source code or website link, but this was special to me.

First, It was the first project I made “for others” to use. Sure, the fist 2 can be useful for others, but they serve my needs first.

In addition to that, I had to try something “different”. In my local area, you can get a simple business website (you know, contact form, what we do, customers reviews, where are we etc…), without any SEO optimization (sell as an optional) for somewhere around 600 to 1200 euros. Most of the time, you end up having a Wordpress website (not in a bad way), on a undersized shared VPS instance (that’s bad) performing 1 req/decade, without compliance at all (no Policies whatsoever), cookies being collected and values stored for tracking reason without any consent etc…

Since I don’t want to learn how to configure and use Wordpress, and since I am still the same cheap f**k as always, I wanted to see if I can get out with a fast, “hand-crafted”, localized website with a lot of caching involved as content rarely changes and interactivity is limited

That was the longest project because I tried to bootstrap the UI using https://lovable.dev/ but I was disappointed when I saw that the website (a pretty nice one I’d say) was a Single Page App 100% rendered on the client

So… you feel better now, right?

Using Claude Code felt like a superpower at first, allowing me to focus on high-level requirements rather than getting stuck in technical rabbit holes.

But the novelty eventually wore off. I realized that while AI is an incredible utensil for efficiency, it often hides the 'how' behind the 'what.' For non-experts (like me or anybody else in a new context, this lack of friction removes the struggle necessary for learning, potentially trading long-term skill development for short-term speed.

Conclusions

Especially for side projects, the 'product' isn't just the code, it’s the curiosity we satisfy and knowledge we build along the way.

While AI tools like Claude Code are incredible for clearing the graveyard of half-finished repos, we must be careful not to automate the very things that made us start the project in the first place.

If we use AI to skip every difficult integration or 'useless' side-quest, we might find ourselves with a finished project but a hollow experience. The secret to using AI in our personal time is to let it handle the chores, while we still need to take on the challenge ourselves.

As my wife always says: "Habere non haberi". we should possess our tools, not be possessed by them.

At work, the equation changes. The industry is almost entirely product- and result-oriented, which aligns perfectly with the "let the AI do it" scenario. Yet, even in a professional setting, AI isn't a silver bullet.

It can quickly become counterproductive if applied blindly. Where it truly shines to me is to use it as "tool to build tools", creating the scaffolding and meta-utility that allows us to focus on higher-level problem solving.

So, is this AI thing hot or not? I’d say it’s boiling (in all the possible meaning ways). Just make sure you’re the one holding the handle, or you’re going to get burned.