From AI noise to a method

I'm planning to write more about AI, mostly technical content, so it's documented for my future self (or for anyone else who's interested). I wanted to start with how it all began: this first piece is the most personal of the bunch — how it started for me, how I gradually got into it, and how it ended up woven into my day-to-day, both at work and outside it, well beyond the usual question to ChatGPT.

Almost starting from zero every time

I'm usually caught up in some idea or tinkering with something in my free time: if it's not reading about some technology or tinkering with some code, it's something else I'm interested in and digging through documentation on for days — in short, buried in personal projects.

Organizing each project was another matter. I'd create a folder with whatever name came to mind first and start poking around in there, without thinking about whether I'd need to find anything in there later. As long as I only had one thing open, it worked. As soon as there were several at once, I'd forget which folder was which, which container belonged to which test. Picking something back up weeks later was almost starting from zero.

At work it was much the same. I work in support, which means juggling tickets, reproductions, tests, code, and I'd spin up environments I never shut down. I had no real system either; it was basically just folders and little else, just enough to get by.

This needed some organizing: both in my free time with my own projects and at work, I was sick of reinventing the wheel every time.

A chat that felt like magic

I remember showing Spotify to a coworker ages ago, long before it became the giant it is today. I told her: no more mp3s, how convenient, just search and go. The same thing happened with AI: it showed up as a curiosity, with no one imagining what it would become.

At first it was ChatGPT: a chat that felt like magic simply because it answered you. It was good for fixing up an email, polishing some text, little more. The boom of tools we have now didn't exist yet, and nobody was quite clear on what a model even was, let alone an agent, a RAG, tokens, and so on.

I'd tinkered ages before, back in university, with neural networks — perceptrons, semantic search, that sort of thing — and had a rough idea of what it was all about. But I was nowhere close to connecting that with what it's become today.

All of this felt a bit far from me. I was busy with APIs, web development, SDKs, and hadn't looked into that world. I had no idea what was coming.

Suggested replies, and little else

Back then I was working at MicroStrategy — today's Strategy, yes, the bitcoin one — and it was the early days of ChatGPT. We started using it on our own, until the company decided to ban it: privacy, confidentiality, that sort of thing. But they never really explained the reason or the actual risks. Everything was so new that they themselves didn't seem too sure what they were banning.

In 2023 I changed jobs: I moved to Auth0 — part of Okta. There, it was already starting to show, though timidly: AI-generated text as possible replies to customer issues, a list of suggestions in case one of them fit. We could also search internal information using natural-language questions, and results came back ranked by relevance — I wonder now whether that wasn't already basic semantic search, a full-fledged RAG, though back then it never crossed my mind to call it that. It was useful, but it fell short: it didn't change how I worked, it just helped me along a little.

From being able to use it to having to

Things got serious when I changed jobs again, in 2025 — where I am now. A much smaller and more organic company than the last one, more agile, more willing to try new things than the corporate dinosaur I came from.

There I work with things similar to before — access control, IdPs, authorization, APIs, infrastructure, containers — though with newer technology, the kind that's a step ahead, so to speak. But deep down it was more of the same: if there was something I didn't know, I'd look up the docs, learn it, reproduce it, and that was that.

Until the company decides to pivot the product it already had to fill a gap in the AI space. That's when it stops being something we could use and becomes something we HAD to use, no turning back. They set up a basic tool for us employees to use, and from there there's no choice but to get on with it.

The same old folders

Here I was in no-man's-land. I had no idea, and suddenly we were handed that tool: I had to figure out how this stuff actually worked, for real this time. And the worst part wasn't that — it was not knowing where to even start looking on my own. I started reading here and there: models, agents, SDKs from different companies, tokens, semantic search. Theory first, then playing around, trying out different SDKs with simple tests — in the same messy folders as always.

The more I read, the more I wanted to dig into it. I gradually understood that what they'd set up at work was mainly for engineering and development. Support was left out of it, as usual. So I got to work.

The magic, again

At first, the company hadn't settled on any of the well-known options — OpenAI, Anthropic, Google — and we could use whichever we liked. At a kickoff, I tried Antigravity for the first time, investigating a real issue, and felt that same sense of magic as the first time: starting from a plain support ticket, it figured out what was wrong, pulled the repo, fixed what was broken, and opened a pull request with the fix ready to go. The same WOW.

Later they settled on one option, Claude, and we all started using it, with a shared company plan. That's when I got serious about it: my own skills, projects, MCP servers. I kept shaping it into my way of working until it became fully part of my day-to-day.

Over time, to cover the gap in support, I adapted what they'd set up and pulled together everything I'd built into a plugin, which covers both the routine side — handling and replying to tickets — and the technical side — reproducing, investigating, checking logs.

The umpteenth time I wanted a site

Given what was happening at work, at home I was already caught up in the same thing on my own. AI had been pulling me in for a while: I'd looked at providers, prices, how each thing worked — API or subscription, cloud or local model — and I was tinkering with different SDKs. All of that was turning into the same old mess of folders with cryptic names.

For the umpteenth time I felt like having a personal site, and for the umpteenth time I couldn't be bothered to build it all by hand from scratch. I'm pretty lazy about things I already know how to do. It's the exact opposite when I have to learn something I like. So I told myself: let's try a Claude Code subscription, which was the option that appealed to me most. I got started with 11ty, and the rest is already told in the series of articles that kicks off this site.

The same mess, fixed in a weekend

One thing led to another, and I realized this, just like at work, could also fix the usual mess of folders. I had it all organized in half a weekend: an idea lab, tidy sandboxes, and a structure by domain — web, AI projects, and so on — each with its own CLAUDE.md. There's one general one, at the user level, with the basic rules; and each domain has its own, more specific one, layered on top of the general one — the guidelines for a web project aren't the same as those for an AI experiment.

As a bonus, I also cleaned up my GitHub repository, which was a mess of things with no real structure — not that there was anything especially interesting in it, but at least now it's clean and presentable.

And from there I didn't stop: I use it for almost everything I need to organize or document, whether it's tech-related or not. A savings plan, for instance — understanding what funds and ETFs are, how to compare them. If there's something to learn or something to put in order, I turn to the same thing.

All of this, by telling Claude what to do. It does it, asks me questions, offers alternatives, and in the end I'm the one who decides, of course — though other times I can't be bothered and just tell it, "Fine, go ahead." And that's where things stand. Whatever comes next, I'll tell it as it happens.