Why the harness exists

The problem

If you're reading this, you already have too much in your head.

Maybe you run a business with twenty-six employees and a household and an investment thesis on the side. Maybe you're holding a dozen vendor relationships, last quarter's tax notes, three open hires, and a recurring question about whether to renew an insurance policy this year. Maybe your morning starts with SHIFT4 and Gmail and RingCentral and Bank of America in four different tabs, and forty-five minutes later you've reviewed the data but haven't yet decided what to actually do.

The traditional answer is more tools. A better CRM. A better POS. Another spreadsheet for anything that doesn't fit. The pitch sounds reasonable — until you notice that each new tool adds another app to your morning round-trip.

The problem isn't that any of these tools is bad. The problem is that you are the integration layer between all of them.

This works until it doesn't. The thing that breaks first isn't the apps. It's you. The cognitive load of remembering everything becomes the bottleneck — not the cash, not the staff, not the calendar.

What I tried first

Here is the honest list, in roughly the order I reached for them:

Twenty-three tools. Each one solved something. Each one created two more things that didn't fit.

The pattern was always the same: a tool would fit one domain beautifully, I'd force-fit everything else into it, the misfits would pile up, and I'd reach for the next tool. Each cycle cost a weekend of migration and produced a brief feeling of control that decayed within three weeks.

The real cost wasn't the subscriptions. It was the missed deadlines. Things fell through the cracks because they lived in the wrong tool, or in three tools at once, or in no tool at all because none of them quite fit.

What broke the cycle was Claude Code. Sometime in late 2025 I stopped reaching for the Claude desktop app and the web interface and started using Claude Code instead. The model was the same. The difference was that Claude Code reads my filesystem. It operates inside the system I already have, instead of being another destination I visit.

That's when I understood. The bottleneck wasn't tool choice. The bottleneck was that the integration layer — me — didn't have a structured way to maintain itself. The filesystem was already there. Claude Code could read it. The missing piece was the discipline that made the filesystem worth reading.

That discipline became the harness.

Four principles

The framing borrows from Karpathy's notion of a "second brain" — a system that holds state alongside you — but reshaped for operators who run real businesses, not knowledge workers managing reading lists.

Four principles.

Principle 1: Filesystem over apps

The most durable knowledge layer in computing is a filesystem with text files. Apps come and go; SaaS companies pivot or shut down; export formats change. A directory of Markdown files on a drive you own will be readable in 30 years.

I learned this the long way. Before the harness I had files in Google Drive, Dropbox, iCloud, OneDrive, and Notion simultaneously. No source of truth. The same document existed in three places at three different versions, and I had no way to know which one was current.

The harness lives in a single directory tree. Every domain has a numbered subdirectory. Every document is a .md file or a .pdf with a .md companion. The whole system can be backed up by copying the directory.

This sounds primitive. It is. It's also the most reliable substrate available.

Principle 2: Read, don't store

The hardest part of any personal operating system isn't capturing information — it's keeping it current. Most knowledge systems decay because they require duplicate updates: change the source, then update the dashboard, then sync the spreadsheet, then update the doc.

The harness inverts this. Documents at the leaves hold source data. Documents at the root (BRAIN.md, MEMORY.md, CHANGELOG.md) compress what's true now by reading from the leaves. When the leaves change, the compression layer re-reads on next session.

My BRAIN.md is about 200 lines. It replaced four spreadsheets that, between them, totaled around 2,000 cells across a dozen tabs — and at any given moment, three of the tabs were stale.

You update once. The system reflects the update everywhere.

Principle 3: AI lives inside the system, not next to it

Most AI tools today are interfaces you visit. You go to ChatGPT, type something, paste the answer somewhere else. The friction is enormous and the AI has no memory of what you did last week.

The harness assumes the AI is operating inside your system: reading your real files, writing to your real artifacts, with persistent context across sessions. Claude Code reads my BRAIN.md before answering my question. ChatGPT can't — and even if it could, it would have no memory of last Tuesday.

Local models (Ollama) handle high-volume tasks: classifying inbox items, generating morning briefs, transcribing calls. Cloud models (Claude, GPT) handle hard reasoning when it's needed.

The AI is plumbing, not a destination.

Principle 4: Discipline over tooling

This is the one most operators want to skip, and the one that matters most.

The vault doesn't run itself. The brain doesn't update itself. Loop signals don't tune themselves. Without you maintaining them five minutes at a time, none of this compounds. The system amplifies your discipline; it never substitutes for it.

The operator is the creator. No discipline, no harness growth.

If you take nothing else from this chapter: the harness rewards the same person every other operational system rewards — the person who shows up consistently. The tools just change what showing up produces.

What the harness does

When the four principles work together, the daily experience is:

• 6 AM: A morning brief lands on your phone. It's 200 words. It tells you the single thing that matters today, the one risk if you don't act, the one decision sitting open.
• You open your laptop, read workspace/BRAIN.md for 30 seconds, and start working.
• Throughout the day, anything that doesn't have a home gets dropped in today's daily log with an OPEN: prefix.
• At session end, you spend five minutes updating BRAIN.md and CHANGELOG.md.
• Tomorrow's morning brief picks up the OPEN: lines and surfaces what matters next.

No app-switching. No "where did I write that thing." No forty-five-minute morning orientation.

The honest pitch

The harness converts your existing high agency into compounding leverage. Five minutes of discipline at session end pays back ten minutes the next morning. A BRAIN.md kept current saves you twenty minutes a week of "where do things stand." Loop signals catch decay before it becomes a fire.

If you're looking for a system that will make you organized when you're not, this won't work. The harness amplifies discipline; it doesn't create it.

If you're already operating at high agency — if the thing you're missing is structure to compound it — clone the vault template. Then read chapter 02.