An Operating System for Judgment
What each part of one executive's AI operating system does, and why it helps — a walk through the map, box by box.
Claude is an AI assistant. On its own, it answers a question and forgets it the next day. The system below is the scaffolding a senior executive built around it so that it could carry real work: remember decisions, find them again, reason with help, and produce things that hold up. The parts look technical, but the design rests on one plain idea worth holding onto as you read. The system keeps a broad record of everything the executive has thought strictly apart from the small body of thinking they actually trust, and a person stands at every gate between the two. The machine is never allowed to promote itself to truth.
Read the map top to bottom. Three workspaces sit at the top, where the work happens. Eight numbered layers sit beneath them, and they exist to make that work faster and to keep what comes out of it honest.
The surfaces — where the work happens
A Project in Claude is a saved workspace with its own instructions and its own memory. The executive keeps three, each for a different kind of thinking.
Esme is for thinking a problem through. They bring a position and Esme argues back, and because it remembers where they landed last time, the conversation moves forward instead of repeating itself. Mrs Harris is for learning. It is a teacher built to do the opposite of a search engine: it withholds the quick answer, gives the established explanation first, and makes the executive predict and restate it, because something understood in passing is gone in a week. Presentation Planner turns a finished position into a deck, following set rules for how a story and its evidence should be built. These three are the cockpit. Everything below makes sitting in them cheaper and safer.
I — Capture: so nothing is thought once and lost
A good idea you can’t find again may as well never have happened, so the system grabs them as they occur. Two phrases do it inside any Claude chat, the main place ideas get made: “archive this chat” saves a plain record of the conversation (the archive-capture skill; a skill is just a short routine Claude runs on command), and “capture for Esme” keeps only the conclusion that got settled (esme-capture). Spoken ideas come in too, transcribed from a pocket voice recorder (Plaud), alongside ordinary working files. Thinking is saved at the moment it happens, in the executive’s own words, with nothing to file by hand.
II — Transport: getting it home automatically
Captured notes have to travel from the chat to the executive’s own computer. They go first to a private online drop-box (a GitHub repository kept off the public internet), and a small program on the laptop (collect.sh) checks it every five minutes and copies anything new into a holding folder (intake). Because this runs on a timer, capturing a thought on a phone in a parking lot works exactly like capturing it at the desk. There is no step to remember.
III — File and split: the most important layer
This is where the one big idea lives. New notes are sorted automatically each morning by a filing agent (the Archivist) into the Archive: the broad record of everything thought, including positions later abandoned. Separately, when the executive wants to update what they actually believe, they run a reconcile step that folds new conclusions into the Curated layer, the short and heavily edited set of things they have decided are true. Reconcile proposes each change, points out anything that contradicts what is already there, and writes nothing without a yes. Keeping these two stores apart is the whole purpose of the system. The record can inform a decision, but only a deliberate human act moves anything into what gets trusted. That, in turn, is why the Esme workspace upstairs can be relied on to argue back: it reads from the curated layer, not the noise.
IV — Retrieve: finding any of it by meaning
A big archive only helps if you can find things in it. A small search program running entirely on the laptop (the MCP server) finds anything in the archive by meaning rather than exact words, so a half-remembered decision surfaces on a plain question. Personal material is filed separately (a per-class index) so the version reachable from the phone, through a guarded gateway (a Cloudflare Worker), can never expose it. The same search also covers the executive’s working files (a projects index) without moving them. A choice made last quarter comes back inside today’s conversation, in seconds.
V — Reason: a second and third opinion
One mind, even a capable one, is a single point of view. A small command-line tool (askmodel) lets Claude put the same plan or draft to other leading AIs, GPT-5.5 from OpenAI and Gemini from Google, and bring back what they say. Each call is priced and logged (cost + fallback), and the executive still decides what to do with the answers. It is a way to pressure-test a piece of thinking against other models without leaving the desk.
VI — Produce: turning thinking into things
Output runs through set pipelines instead of being rebuilt from taste each time. Essays grow from the executive’s own argument through a fixed drafting method (the LinkedIn system) and get published, with the tools behind them, on a website whose build automatically rejects anything off the house look (SBRetrieval — the site you are reading). That look is itself a written rulebook everything inherits (the Design Language). Quick tasks and slides come off the same line (Notepad and decks). The result is a steady standard: things leave looking and reading the way they are meant to.
VII — Assess: the part that only judges
This is the layer most people never build, and for a leader who cannot personally check everything, it matters most. Four skills read finished work and report what is wrong, and not one of them rewrites a word. slop-check reads any piece of writing in five passes and names its emptiest sentence, the polished line that says nothing. linkedin-review runs an essay through four tests, including a check for the tics that mark AI-written prose and a blunt question about whether the writer actually learned anything. check-voice measures a product’s writing against six defined voice rules. check-spec-coherence compares a plan against the software built from it and flags where the two have drifted apart. The reason this works: the standard is written down, the machine measures against it and flags failures, and the human keeps the pen. What these catch is the hardest thing to catch in your own work, output that looks finished and is not.
VIII — Substrate: the floor it all stands on
Underneath sits the plumbing that makes the rest dependable. A standing instructions file (the global CLAUDE.md) loads into every session, so the assistant always knows how the executive works. A file-memory holds settled facts so nothing has to be re-explained. A reusable scaffold (agentic-OS) drops the same structure into any new project in minutes. And a library of skills plus scheduled background jobs (cron) keep the couriers and indexes running untended. None of it is glamorous; all of it is what lets the layers above run without babysitting.
Why it holds together
Look back at the map and the same shape repeats on every layer. The record is kept apart from the trusted thinking. Diagnosing a problem is kept apart from deciding what to do about it. The filing agent may read and sort but not write; a plain, dumb script does the writing, so an automated run can never be tricked into damage. Everywhere, a broad and automatic process is fenced off from a narrow and human one, and a person stands at the gate between them.
That is what makes a system like this safe to lean on, and it is the answer to why any of it helps. The parts that run on their own are exactly the parts that cannot change what the executive believes or sends out. The parts that can, wait for a human. The machine carries the memory and does the mechanical work; the judgment stays with the person. The small amount of friction it leaves (running the captures, approving the reconcile, invoking the checks) is not a rough edge to be automated away. It is the seam where judgment enters, and it is the most valuable part of the design.