How I built an AI that knows me

I run retail operations for a national chain — 120 stores, 800 people. In early twenty twenty-five I started using AI properly. Drafting updates, thinking through problems, preparing for conversations I was dreading. It was good. However every single session started the same way — with an AI that had no idea who I was. Every chat, I re-explained my role. My team. How I like things written. What I was working on. I'd get a great result on Tuesday, and on Wednesday the AI had forgotten I existed. It was like hiring a brilliant consultant with no memory — impressive for an hour, exhausting for a year. So I fixed it, and the fix was smaller than I expected. What I built is called an AI Personal Operating System and I'll show you all of it — what's in it, how each piece got built, what you need before you start, where it falls short, and what one ordinary Monday looks like now. No subscription and no new app. It works with whatever AI you're already using. And at the end I'll tell you where to get everything I built, free.

This is the whole idea. One folder on my computer, holding a handful of plain text files. Who I am. How I sound when I write. How I like to work. And what I'm working on right now. My AI reads that folder automatically at the start of every session, before I've typed a word. That's an AI Personal Operating System. Not an app, not a subscription. A folder that briefs my AI the way I'd brief a new chief of staff — and it never forgets the briefing. What that changes, day to day: it writes in my voice instead of generic AI English. It pushes back when my thinking is loose, because it knows I want that. And it starts every session already across my world, so I get to the useful part in the first message rather than the fifth.

Before the build, the shape of it. Four layers, and they go in in that order. Layer one is who I am. Identity, voice, rules — three files, and I'll spend most of this video there, because everything else leans on it. Layer two is how my work gets done. My quality standards, the processes I repeat, and the jobs that run on a timer without me asking. Layer three is what I'm working on. A folder per project, holding the context I'd otherwise re-explain every time I opened a new chat about it. Layer four is navigation — one page that tells the AI, and me, what's in the system and where. That sounds trivial until you have forty files and can't remember what you called them. Build them in that order. Layer two is worthless on top of a vague layer one, and I know that because I tried it the other way round.

So, layer one. Three files, and I didn't write any of them. I talked, and the AI wrote. Identity — who I am, my role, my goals, what I care about. Voice — how I write, so the AI stops sounding like a press release. And rules — how I want it to behave. When to challenge me, how I like answers structured, and when to stop discussing and get on with it. Each one gets built the same way. I paste in a prompt, and the AI interviews me. One question at a time. Let me show you what that looks like, because it surprised me.

It asked what I do. I gave it the rough version — retail ops, 120 stores, twenty years in the industry. It told me that was too generic and pushed for more. I added that I specialise in turning around underperforming stores. Still not enough. That went back and forth for about twenty minutes. Then it drafted my identity file, and the first draft was good, but it wasn't me. So we did a few more rounds until it was. What came out is about a page. My role and the scale of it, the two things people come to me for, what I'm trying to achieve, and the three values I don't bend on. The whole thing took half an hour and I never wrote a sentence. My only job was answering honestly, and refusing drafts that didn't sound like me.

The voice file works differently, and it's my favourite part of the whole system. Instead of describing how I write, I attached real examples — a couple of emails, a board update, a note to my team. The AI read them and told me things about my own writing I'd never noticed. That I open with the person and not the subject. That I put numbers inside sentences instead of in lists. And that I start about a third of my paragraphs with "so", which I'd have denied under oath. I corrected what it got wrong, added what it missed, and it wrote the file from that. Your writing knows your voice better than you can describe it. Let the AI read it.

The rules file rounds out layer one. Mine is short and blunt. Give me a recommendation, not five options. Don't soften bad news. Challenge me before I've decided, and execute once I have. And tell me when you're not sure, rather than sounding confident and being wrong. Four lines, and they changed the character of every answer I get. AI is agreeable by default — it will happily tell you your plan is strong. Mine argues with me now, because I asked it to, in writing, once. Three files. A few conversations. At that point the AI knew me better than some colleagues do. But it still did nothing — because the files were sitting in a folder that nothing was reading.

Two switches. Switch one is the boot file — one file saved at the top of the folder, which my AI reads automatically every time a session starts. All it says is: before you do anything, read these files, in this order. That's the ignition key. Switch two is a short paragraph pasted into the AI's settings — a compressed version of who I am. That's the backup, for sessions where the folder isn't connected. A quick question on my phone, for instance. Folder connected, I get the full system. Not connected, it still knows the essentials.

So I ran a test. Brand-new session, one I'd never used before, and one question: what do you know about me? It answered. My role. My stores. My goals for the quarter. That I want recommendations, not options. That I don't like bad news softened. It even knew how I sign off my emails. Then I asked it a real question — draft me a note to my area managers about a slipping standard — and it came back in my voice, with a recommendation at the top and no padding, because that's what the rules file told it to do. A blank session, and it already knew me. That was the point this stopped being a clever trick and became how I work. Every session since has started warm. No re-explaining, no warm-up laps.

Worth saying what this costs, because the answer is almost nothing. You need an AI assistant — whichever one you're already using. The whole system is plain text files, so none of it is locked to one tool. I've moved mine between assistants twice and it carried over intact. It works best if your AI can connect to a folder on your computer, because then it reads the files itself and you stop thinking about it. If yours can't do that, you're not shut out — you paste the files in at the start of a session, or keep them in whatever project or workspace feature your tool has. Less elegant, same result. And time. Half an hour for the first file, then a few more sessions for the rest. That's the whole bill.

Layer two. Knowing me is one thing; knowing my work is another. Three kinds of file, teaching the AI how my recurring work gets done, so I never explain the same thing twice. Reference files are my quality bar. Mine says lead with the point, numbers are specific or absent, everything ends with clear next steps. I never mention it — the AI checks against it automatically, every document. Skills are repeatable processes. Anything I do more than once gets one: written down once, triggered with a phrase forever after. And scheduled tasks run on a timer, without me asking at all. I've got two. One assembles my week on Monday morning. The other, on the last Friday of the month, pulls together what changed across my stores and asks me the three questions I'd otherwise skip.

Every Monday I send a trading update — same structure, same standards, new numbers. It used to take me forty minutes. Now it takes one phrase: draft my weekly trading update. The skill already knows my structure. Headline first. Numbers inside sentences. What's working, what needs attention, actions with owners. It asks me for this week's figures, I paste them in rough, and it writes the update in my voice, to my standards, ready to send. Forty minutes becomes five. And the five-minute version is better, because the structure never has an off day. Writing the skill took about ten minutes, and again I didn't write it. I described the job out loud — what goes in it, what order, what good looks like — and pasted in three old updates as examples. The AI turned that into the skill. Same trick as the voice file: show it, don't specify it.

I didn't build the engine in a weekend, and you shouldn't either. One reference file first — whatever you produce most. Then a skill, the next time you catch yourself explaining the same task twice. This is what it adds up to. Three moments from one ordinary Monday. Seven in the morning, before I've properly opened my laptop. Nine-thirty, the weekly update. And two in the afternoon, a conversation I'd been putting off for two weeks.

Seven a.m. I haven't asked for anything — this ran on a timer. My calendar scanned, the three meetings that count flagged, my priorities checked against my goals, and the one thing most likely to slip, named before it slips. Two minutes with a coffee and I know my week. That used to be forty minutes of Sunday night, and I'd do it badly, because it was Sunday night.

Nine-thirty. One phrase, this week's numbers pasted in rough, and the trading update is written. My structure, my voice, honest about the week we actually had. I read it once, change one line, send it. Five minutes — and the most tedious forty minutes of my week is gone.

Two in the afternoon. A performance conversation I'd been putting off for two weeks. My prep template asked me the questions I'd been avoiding asking myself. What's the real issue. What's my part in it. What outcome do I need from this. Then it built the prep. An opening line. The two or three messages that had to land. The reactions to expect, and how to hold steady through them. I walked in prepared and calm instead of anxious and improvising. She's still on my team, and she's doing well. That's not a time saving. That's a better outcome for a person, and it's the part of this system I'd give up last.

Before you go and build one, three things it won't do — I'd rather you heard them from me than found them out in week three. It repeats your habits back to you, faithfully, including the bad ones. If you write in long hedged paragraphs, your voice file teaches it to write long hedged paragraphs. It goes stale. My goals changed in March and I didn't update the file for six weeks. For six weeks I had an AI confidently briefing me against priorities I'd already dropped. It still gets some facts wrong, and knowing you well makes it sound more convincing when it does. Checking is still your job. This saves you the explaining. It doesn't save you the thinking.

If you want to build one, the roadmap runs in five phases, and the first two are the ones to do this week. Foundation — identity, voice, rules. Then switch it on — the boot file and the settings paragraph. That's the moment your AI stops being a stranger, and it's an afternoon's work. Phases three to five grow it from there. The engine. Your projects. And the habits that keep it current — re-read your identity file whenever something big changes, and again every few months whether it has or not. Spread it across a few sessions. There's no prize for rushing. And if you only do one thing after this video, make it the first conversation. Open your AI, paste in the identity prompt, and answer its questions properly for twenty minutes. You'll know by the end of it whether the rest is worth your time.

None of this is magic. It's a folder of plain text files and a few habits. That's all it is. Hours back every week, and better work at the end of it — which was the part I didn't expect. Everything I built is free. Go to mel-and-mo dot com and look for Build Your AI POS. You'll find the guides, the blank templates that become your files, and my complete finished system as the worked example, so you can see where you're heading before you start. Start with the identity file. One conversation, half an hour. Then see whether your AI still doesn't know you.

That's the whole system. It took me a handful of sessions across a few weeks, and it's changed how I work every day since. The AI didn't get smarter — it stopped starting from scratch, because now it knows me. Good luck. I mean it — an AI Personal Operating System is worth building.