In the middle of June I opened a fresh Claude session, typed one lazy line, what should I work on tonight, and the answer named the exact thing I had been avoiding for weeks: a lead list of 6,523 rows whose Sent? column is still completely blank. I never told that chat about the list. The fact was sitting in a plain text file on my disk, inside my personal AI operating system, and the session read it before it read my question. So instead of a polite generic suggestion I got called out by my own filing system at eleven at night, tea going cold beside the keyboard.
That is the honest pitch for this whole piece. I work a full day at Huawei as a business analyst, I freelance evenings on Upwork where I am Top Rated with 100% Job Success, I ship small paid tools, I run four X accounts, and since spring every bit of that runs through one folder of files that Claude reads at the start of each session. People searching for this idea call it a Claude AI OS or an AI second brain. I call it the folder. Let me show you the shape of what is inside mine, and the one file that does most of the work.
Every new chat is a stranger
Before the folder existed, my mornings with AI had a shape you will probably know. Open a chat, and before asking for anything useful, explain myself. Who I am, what I sell, my rates, which client is waiting, how I like my proposals to sound, which product is mid launch. By the time the model was ready to be useful I had spent 20 minutes typing context, every single session, because a chat has no yesterday.
The time was the smaller loss. The bigger one was inconsistency, which is already my weakness and did not need the help. On Monday a proposal sounded like me, on Wednesday it sounded like a brochure, because I had described my voice a little differently each time. Decisions evaporated too. I would settle a price for a tool in one chat, and another chat would happily argue for a different one, and both were right by their own tiny memories.
The part that stings is that I had half solved this without noticing. I kept my introduction paragraphs in a notes file and pasted them into every new chat by hand. That is a context file with a human loader. The day I saw it that way, the rest was obvious. Put the files where the AI can read them itself, and stop being the loader.
What a personal AI operating system actually is
An AI operating system is not software you install. It is a set of plain files an AI reads at the start of a session, so it stops being a stranger every morning. Mine is markdown in ordinary folders. No database, no plugin, no subscription beyond the AI itself, and if Claude vanished tomorrow the files would still open in Notepad and still describe my working life.
Four layers do the work. A context file holds who I am, what I sell, my numbers, and my rules, meaning the actions the AI must ask about before taking, like messaging a client or spending money. A wiki holds knowledge, one page per topic, with a small hot cache file on top that always gets read first. Skills are written procedures for jobs that repeat. Memory is where decisions get logged, so future sessions inherit them instead of arguing them again.
Files beat chats for a boring reason. Chats forget, and everything you typed is unreachable next session. Files version, so every change is a diff, and mine sit in git like any code. And a human can still read them, which matters more than it sounds, because the whole point of an AI second brain is that it stays your brain, open and yours, never a box that answers questions about you in ways you cannot check.
The hot cache, the small file that makes it fast
The first version of my wiki had one problem. It was good, so it was big. If a session read all of it before starting, I paid time and tokens for pages about Fiverr while asking about an ebook. If it read none, I was back to the stranger problem. The fix is the one piece I would keep if I threw away everything else. A single small file, around 500 tokens, always read first.
Mine lists three things. Active threads, current numbers, and the one next action. Right now that is an ebook waiting for my review and the frozen lead list, the row count of 6,523, and a sent count that still reads zero. A session opens, reads this card in seconds, and orients, the way you glance at the sticky notes on your monitor before the first sip of tea. Most questions never go deeper than this, which is exactly why the whole thing feels fast.
Under the card there is a reading order, a fixed path so the AI never wanders, and a rule that caps how many pages one question may open. That order, the exact templates for the context and hot cache files, the skill files, and the memory pattern are the parts I keep as one assembled kit rather than loose paragraphs, because the value is in how they fit together, not in any single file on its own.
Skills for repeated work, memory for decisions
A skill is a procedure written down once. The steps, the rules, an example of good output. I have skills for Upwork proposals, Gumroad listings, weekly X reviews, and yes, articles. The piece you are reading came through that pipeline, drafted by a session that read my voice rules and my fact files before it wrote a word. I wrote separately about how I trained Claude to write like me, and the short version is that the whole training lives in one plain file. Most of the ways I make money with Claude are really skill files run on a fresh evening.
Memory is the quiet layer. When I decide something for a reason, the decision gets logged where future sessions look, like my rule that product source stays in private repositories, which went in after a release script nearly attached a paid zip to a public page. And one growth rule keeps the whole thing lean. Nothing becomes a skill until I have done it by hand three times. Twice can be chance. Three times is a pattern, and a pattern is worth thirty minutes of writing down.
That rule also blocks the classic failure, building a beautiful empty system. The setup grew backwards, from real repeated work into structure, and that is the only direction that survives contact with a busy week.
Anything you have done three times by hand is a skill file waiting to be written.
So you have the shape now. Every new chat is a stranger until you hand it files. Four plain layers fix that, a context file, a wiki with a hot cache on top, skills, and memory. The hot cache is the piece to build first, a single small file read before any work, and the growth rule keeps it honest, write a skill only when a job has repeated three times.
If you would rather not assemble it from scratch, the full folder structure, the file templates, the skills, and the memory pattern are packed ready to copy in Build Your Own AI Operating System with Claude, and members on my site get it, and every other product I make, free. Keep the rest in plain files you can read without any AI at all. Models will change and subscriptions will lapse, and your folder will still be there, holding your working life in a form you own.