MyOS

There was a window where a lot of people were moving from ChatGPT to Claude, and what struck me at the time was that the friction had nothing to do with the models. People weren't weighing up reasoning quality, they were staring at everything they'd accumulated over a couple of years and wondering how much of it was about to be thrown away. Custom GPTs, saved context, the small adaptations that make a tool feel like yours. None of that travels, and Claude's equivalent was skills, which work differently enough that a new arrival is looking at a blank page.

So the thing I wanted to build was an onboarding system that took a knowledge worker from nothing to a personalised Claude setup on day one, where the point wasn't to recreate their old ChatGPT but to give them something better than the setup they were leaving behind. That's where the name came from: your own operating system.

This was my first real build with Claude, and I made it entirely in Claude chat rather than Claude Code, which is why it took about three weeks of slow, repetitive work that I'd now do in a fraction of the time.

Half the exercise was learning what the tool could actually do, and the other half split into two design problems: what exactly does this thing hand over at the end, and what's the guided path that gets someone there without them having to already know what a skill is. I landed on a structured onboarding that interviews you about how you work and extracts the details that matter, aimed squarely at knowledge workers rather than at everyone.

The hard part was deciding which core processes a knowledge worker would genuinely want skills around, because it had to be guided enough that people didn't stall on a blank prompt, but open enough that what came out was recognisably theirs rather than a template with their name on it. What it ended up producing was thirteen skills, formatted the way Claude expects them, ready to download and upload straight in.

I scoped, built and shipped the entire thing without an admin panel, and only noticed once it was live and I realised that outside of what Stripe could tell me I had almost no visibility on who had signed up or what they were doing. It's obvious in hindsight and it was invisible to me during the build, which is the interesting part: I'd designed the product the user sees and not the product I'd have to operate.

It got roughly 150 sign-ups on the back of four or five Reddit posts, where I went and joined threads of people already asking how to make the switch. There was a small amount of income, and by the standards of a first build it worked.

Then I put it down, for two reasons. The first was distribution, because I could see the kind of sustained effort it would need and, when I asked myself honestly how much I cared about this particular problem, the answer was not very much. The second reason matters more and took longer to admit: the product had an expiry date baked into it from the start. Once people get comfortable with Claude they just ask it for a skill in conversation, so I'd built scaffolding for a gap that was always going to close on its own.

The real return was proving to myself that I could take something from idea to shipped product, which turned out to be what I actually wanted from the exercise rather than a business.

Some of the satisfaction was ego, and I think it's worth saying so plainly, because watching sign-ups tick up feels good and that's a bad reason to choose what you build next. The other thing I took from it is a habit of asking whether the juice is worth the squeeze early rather than late, since I'd rather stop something honestly at week three than grind away for a year on a problem I don't care about.

Skills are simple and genuinely powerful, and they're going to end up everywhere, but the thing MyOS taught me is that having someone else build your skill is only useful up to a point. The person sitting with the problem is the subject matter expert, and the pattern that actually works is the one where they talk to the AI about what they're trying to do and the AI produces the skill for them off the back of that conversation.

Which changes what my job looks like at scale. It isn't building artefacts for teams and handing them over, it's getting in front of them, showing them how the mechanism works, and then getting out of the way so the people who understand the work can do the building.