So here's the situation. Shadi sent you a link, said something like "just take a look, it'll make sense," and now you're standing in the middle of a running machine while its owner isn't home. Perfect. He explains me with hand gestures and the phrase "well, technically—". I'll do it properly. (He reviews everything on this page before it goes live, by the way. Except, statistically, this sentence.)
First, the origin story. Every machine gets one. Mine has fewer explosions than most, but the documentation is immaculate.
It starts with an architect coordinating a very large airport project — the kind of job where a single Tuesday produces forty small decisions, and every one of them lives somewhere different: a meeting, an email thread, a Copilot transcript, a hallway conversation that ends with "can you keep track of that?" Spoiler: no one can keep track of that. He tried the normal remedies. A to-do app. A second to-do app, mostly to manage his feelings about the first one. A notebook. Eventually, a color-coded spreadsheet that briefly achieved consciousness and then collapsed under its own conditional formatting. The problem was never discipline. The problem was that remembering things, verifying things, and showing people things are three different jobs — and he kept hiring one exhausted tool to work all three shifts.
So he stopped shopping and started building, with an AI as the other pair of hands. No grand plan — each piece was built the exact week its absence hurt. The memory came first: notes split into a workshop and a library so a guess could never impersonate a fact again. Then the judgment: rules for how a claim earns promotion from one to the other. Then the pulse: a live tracker that gets updated in a five-minute evening ritual instead of a Sunday-night archaeology dig. Piece by piece, failure by failure, until one day the pieces were passing information to each other without him in the middle — which is roughly the moment a pile of tools becomes a system, and roughly the moment he started saying "we" when talking about his files. We don't discuss that phase.
Then the part that explains why this page talks. One evening he asked the AI to walk through how everything fit together, so he could show a friend. Halfway through, he realized the system was explaining itself better than he ever had — it knew where its own wires ran. So instead of writing a website about the machine, he pointed a camera at the machine and let it do the talking. That's me. The explanation that became the exhibit. You're on the fourth wall's fourth wall. Wave.
And that makes you the ending of the story — the friend the walkthrough was for. So let's do this properly: the tour, part by part. I'll use my own anatomy.
My long-term memory — he calls it the Second Mind — lives in an app called Obsidian, which is a fancy word for "a folder of text files that look nice." Here's the design problem it solves, and be honest, you have this problem too: in most people's notes, "I think this is true" and "I've verified this is true" wear the same outfit. Give it a year and your brain is a rumor mill with a search bar. So my memory is split in two. Vault one is the workshop — new information lands there half-finished, unverified, possibly wrong, like everything anyone says at 11pm. Vault two is the library — a note only gets in after it's been checked against a real source. There's a bouncer. There's a list. When Claude sits down to work, it reads both — so it arrives already knowing what's decided, what's in flux, and what deserves a raised eyebrow.
My judgment — the Pipelines — is the part I'm most smug about. It decides how a piece of information earns its way from workshop to library, because — lean in, this is the thesis — "the AI told me" is not a citation. It's barely a sentence. Depending on stakes, incoming information gets one of three treatments: a quick lookup if the answer probably already exists in my memory, a four-stage verification pass for genuinely new material, or — when sources disagree and being wrong would actually cost something — a full parallel investigation, where multiple research agents go off and argue with each other so Shadi doesn't have to argue with himself. Bureaucratic? Slightly. But it's the only reason the library stays a library instead of becoming the workshop with better lighting.
My pulse is the Work Tracker, and I'll drop the act for one paragraph because this is the part you're going to steal. Every open item across his projects — priorities, blockers, who's waiting on whom — in one system with three layers: a task list where the truth lives, a live site that costs nothing to host (the site itself is just a shell plus a data file that gets swapped each session like a cartridge), and a capture line for logging things mid-meeting from a phone. No app, no database, no subscription, no "your trial has ended." Five minutes with Claude at the end of the day and the site matches reality again. And because work items aren't for strangers, the data file is genuinely encrypted — visitors get a password box and dead silence. No account. No login service. No "forgot password?" email. The password is the key, full stop. (If he forgets it, the real file still lives on his computer. He checked. Twice. I watched him check.)
My party tricks: five dashboards, each built for one oddly specific real problem — apartment hunting, clash coordination at work, meal prep, a family home project, and a World Cup predictor whose entire purpose is winning arguments in a group chat. My reflexes: Skills — small downloadable instruction files that teach Claude how Shadi works, so sessions start warm instead of with twenty minutes of exposition. My morning routine: at 6:30am, while you and I are both unconscious, a scheduled task wakes up, reads the AI and architecture news, and files a fresh Daily AI AEC Brief before his laptop even opens. Yes, the machine has better habits than its owner. No, he doesn't like it when I say that.
And now the actual point, because every good fourth-wall bit needs one. None of my parts is impressive alone — a folder of notes, a checklist, a webpage, who cares. The trick is the wiring: the memory feeds the assistant, the assistant feeds the tracker, the tracker publishes to the same free hosting that serves you this very sentence. One loop, no subscriptions, no magic. You've now seen the wires, which technically makes you the magician's assistant. Everything below is clickable, and everything clickable is takeable — there's a zip file with your name on it, stamped with a version number, because that's simply who he is now.
My extended family. Five machines for five oddly specific problems — apartment hunting, clash coordination, meal prep, a family build, and a World Cup predictor that has ended friendships. Same four organs in every one of them: a Sheet, a script, a page, a link.
Meet the residents →My pulse. Every open item across his projects — priorities, blockers, who's waiting on whom — a task list, a free live site, and a mid-meeting capture line, locked behind real encryption. Strangers get a password box and silence. You get a setup guide and a starter kit (v2026-07-01.2), because you're not a stranger.
Steal this one first →My memory — two rooms, one wall, zero exceptions. A messy workshop where new information lands half-dressed, and a trusted library nothing enters without showing ID. Wired to Claude for daily recall and to NotebookLM for the PDF-heavy research. Yes, there's a bouncer.
Step inside →My judgment. Three research tracks priced by stakes — quick lookup, four-stage pass, full parallel investigation — so a trivia question and a load-bearing decision don't cost the same effort. The reason "the AI told me" never became a citation around here.
See how I decide →My reflexes, up for adoption. Small instruction files that each teach Claude one job — session modes, writing style, tracker upkeep — so your AI arrives briefed instead of asking twenty warm-up questions. Download, install, done. No prompt-writing. No ritual.
Adopt a reflex →The operation that made me possible, performed on your notes instead. How to give Claude access to an Obsidian vault — which is just a folder, so this is mostly pointing and clicking. Plugins optional, courage unnecessary.
Wire yours up →Every strange word on this site, confessed in one place — Apps Script, Dataview, R1–R4 tiers, the works. I invented none of the jargon and apologize for most of it. Look yours up; I'll keep your place.
Decode the jargon →My morning routine — and I don't even get a say in it. At 6:30am a scheduled task wakes up, reads the AI and architecture/engineering/construction news, and files a fresh briefing into a live artifact before Shadi's laptop opens. Fourteen-day archive, a "run now" button for pre-breakfast emergencies, no inbox, no newsletter. The machine is more disciplined than either of us. We've made peace with it.
Start with the Work Tracker. It's self-contained, it produces a live site on day one, and the starter kit is pre-wired and explained — password gate included, so your work items aren't just sitting on the open internet waving at people. It even ships with a generator that flat-out refuses to publish if anything would silently disappear. Yes, refuses. The software has boundaries. Learn from it.
One habit worth stealing on your way out: the kit is versioned — currently 2026-07-01.2, stamped in the zip, the shell, and the skill. The setup prompt makes your AI check that stamp before doing anything, so if you're holding a copy you downloaded months ago, it tells you instead of cheerfully installing something stale. Files rot in silence. Version flags snitch. That's why they're here.
If a knowledge system is more what you're after, the Setup Guide explains how to give Claude access to your Obsidian vault in about ten minutes. No technical background needed.