What three rebuilds of my Family OS taught me

I could have posted a sparkly screenshot and told you I built my Family OS in a weekend… cos I'm a genius. But that's not the real version. The real version involved three full rebuilds, a lot of Googling, and at least one moment where I closed my laptop and went to stare at a wall.

I am not a coder. Not even close. But I am good at deep diving, deciphering, and refusing to give up when something should work but doesn't.

I find the mistakes are often the best lessons. So that's what I'm going to share with you - so you can learn with me.

One thing: an AI system that takes real load off the invisible job. The school emails, the dinner decisions, the life admin that lives in my head rent-free. Not a novelty. Not a "look what AI can do" demo. A usable, practical system that actually runs and simplifies my life.

Simple enough goal. Getting there was less simple.

I asked Claude to build the whole system - skills, files, structure, all of it. And it did. But when I went to check the work, I had more questions than answers. Things were bloated, overlapping, full of assumptions. The same information lived in multiple places. Things were contradictory. And I had no clear picture of what was really driving what.

The lesson: Claude will build confidently, but it won't build cleanly - not without clear direction from you on what should go where, and more importantly, what should not go where. So I made a rule: one source of truth for everything, intelligent naming so I can find it, and keep every file and skill as simple as possible.

Giving AI the right background matters far more than writing a clever prompt. AKA: context engineering > prompt engineering. It's the backbone of building good agentic systems. Dump context everywhere and you end up with an unmanageable system, confused agents, and sub-par outputs. It'll also chew through your credits.

In sum, taking a beat to work out what should go where and why is worth it.

Oh - and one more thing from attempt one. I knew to always check AI outputs. But I was genuinely shocked that Claude lied to me even after I called it out - multiple times - for making things up or not doing what it said it did. Bareface. Like asking a nine-year-old if they brushed their teeth.

So when you verify - and you will need to - don't ask Claude to check its own work. Open the actual file.

After the mess Claude made, I took the wheel. I set up separate projects. Meal planning here. School emails there. I came unstuck when I realised I wanted one consolidated output - like my daily digest or week-ahead view. I didn't want 5 separate projects running with 5 separate outputs. I didn't think about this requirement until later.

The lesson: if the parts of your system share context, or if you want to interact with it in a unified manner - they belong together in one project.

It worked! And technically if I was planning on stopping with this one use case it could have lived here. But I was already hitting roadblocks for what I wanted the system to become. I wanted to braindump anything into it - from "pick up the dry cleaning on Saturday" to "I bought olive oil, tick it off." I needed it to create and edit files for me, and run automatically on a schedule. The browser version of Claude can't do that.

I was genuinely stuck. And then Claude launched the desktop app.

The lesson: jump straight to Claude desktop. It's the biggest evolution I've seen and it’s incredibly powerful. If you're looking to go 'agentic' with your home life, it's the only way. Get it up and running on your laptop now and create one project called Family OS.

So that's the build story - three attempts, three lessons, zero regrets. Now here's something you can actually use right now.

If you're keen to jump in, here's a ready-to-use Claude skill. It's actually my favorite one and I haven't met a parent who doesn't need this.

Every week, your kids' schools (+ other activities) send a pile of emails. Some are noise. Some contain a date you'll forget by Thursday. Some have a deadline buried on page 7 of the newsletter that you won't see. Ever.

This skill fixes that. You give it to Claude, answer a few questions about your family, and it builds you a personalised skill that:

You choose how it works

[Download the Kids Email Digest skill →] < Full instructions and file here.

Apple is rebuilding Siri from scratch - powered by Google's Gemini model. The Siri that never really worked is getting a proper brain transplant. If you've written her off, it might be worth paying attention later this year. Why it matters to you: the AI assistant on the phone already in your pocket is about to get significantly better - no new app required.

Meta launched Muse Spark - their new AI assistant coming to Facebook, Instagram, WhatsApp and Messenger. Why it matters: AI assistants are about to show up in apps you already use every day. You won't need to go looking for AI. It's coming to you.

Most people aren't seriously thinking about AI yet - let alone actually doing it (despite what LinkedIn would have you believe).

If you try something and it doesn't work? Good. That's not failure. That's the job.

PS - I genuinely want to know what you think. What landed. What didn't. What you'd like more of. If you have any suggestions on what would help you - I’d love to hear them. Hit reply - I read every single email (in real life not with AI).

PPS - If you know a parent who keeps saying "I really need to get into AI" and deserves a quick win - forward them this!

See you on Thursday.

Sarah x