The three layers

If you cloned the template, you're looking at twelve numbered directories and a few special folders, wondering what you actually do with them. This chapter is the map.

The harness has three layers. They sit on top of each other. You'll spend time in all three, but you'll spend the most time in the middle one.

Layer 1: The Vault

The substrate. A directory tree on a drive you own.

Twelve numbered categories represent the domains you run:

Plus two special directories:

And one operating layer (the brain — Layer 2):

That's it. Everything lives in one of these. Anything that doesn't fit goes in 12-daily-log/ for today and gets filed properly later.

What "the vault" looks like in practice

The vault is one root directory. Mine lives at /Volumes/T9/ — a 2TB external SSD permanently attached to a Mac Mini M4 base model. FileVault on, sleep disabled, auto-restart on power loss. About six hundred dollars for the machine plus a hundred and thirty for the drive. Yours could live anywhere — your laptop's home directory, an external drive, a NAS, a cloud-synced folder — but if you're going to run cron jobs, you need something always-on.

The location matters less than the discipline. The discipline is: every file you create as part of your operation goes inside the vault.

This rule sounds restrictive. It is. The restriction is what makes the harness work. When everything lives in one tree, AI tools can read across it, backups capture everything, and you can never lose anything because you don't remember which app you used.

Why numbered categories

Numbers force a sort order. Without numbers, alphabetical sorting puts "Legal" before "Marketing" before "Operations" — random from a workflow perspective. With numbers, you see your business in the order you think about it:

01-operations is what you do today. 02-finance is the money side. 03-people is who's involved. And so on.

The numbers also remove ambiguity. There is one place vendor invoices go. One place tax documents go. One place customer files go. No deliberation. The friction of "where does this belong?" disappears.

Layer 2: The Brain

The compression layer. A small set of living documents that read from the vault and represent what's true now.

Five files under workspace/:

• BRAIN.md — current state. Priorities, blockers, system health, numbers to know.
• MEMORY.md — long-term knowledge. Patterns, lessons, locked decisions.
• CHANGELOG.md — append-only audit trail. What changed, when, why.
• PLAYBOOK.md — decision frameworks for recurring scenarios.
• VOICE.md — writing style for anything you publish.

These five documents are the heart of the system. They are the ones you maintain by hand (with AI help). They are also the ones your AI tools read at the start of every session to orient.

If you stop maintaining anything else in the vault, maintain these.

Why this layer exists

The vault holds source data: invoices, contracts, emails, transcripts, dashboard exports. The brain holds compressed state: what you've decided, what's blocked, what's important now.

The compression matters. A vault without a brain is a hoarder's basement — everything is there but nothing is findable, and the burden of knowing what's in it falls on you. A brain without a vault is a sticky note covered in deadlines that you can't verify.

The two layers work together: the brain compresses what's in the vault; the vault provides the receipts that back up the brain.

The discipline that makes the brain work

Five minutes at session end. BRAIN.md gets a 30-second update of what changed. CHANGELOG.md gets one short entry. MEMORY.md, PLAYBOOK.md, VOICE.md sit quiet for weeks at a time.

That's the whole maintenance burden. Five minutes a session.

Skip it, and the brain decays within two weeks. The vault still works (it's just files), but the compression layer becomes stale and you start re-deriving things you already knew.

Layer 3: The Harness AI

The agent layer. Local + cloud AI integrated at the substrate level.

The harness AI is not a chatbot you visit. It's a set of automations and assistants that live inside your operating system:

• Slash commands — short prompts mapped to common workflows. /logs runs your session-closing protocol. /file categorizes an inbox item. /audit finds decay in the vault.
• Cron jobs — scheduled scripts that read the vault, do work, write back. The morning brief is the most important: 6 AM cron generates today's brief from yesterday's OPEN: lines, current BRAIN.md priorities, and any other inputs.
• Loop signals — automated monitors that surface decay. "Decline rate over 3%" or "VIP customer not seen in 60 days" — these post to a loop_signals table or file, and the brain reads them on next session.
• Cross-session memory — Claude Code, Cursor, and similar tools support persistent memory across sessions. The harness uses this to retain context you don't want to repeat: who's on your team, what your top priorities are, what your voice is.

Local-first, cloud-when-needed

The harness AI is opinionated about which model handles which task:

• Local (Ollama, ~$0/mo): classification, summarization, structured extraction, daily morning brief, email labeling.
• Cloud (Claude, GPT): heavy reasoning, code generation, long-context analysis, anything user-facing.

The reason: local models are infinitely cheap and always available, but qualitatively weaker. Cloud models are higher quality but cost real money and add network latency. Each task goes to the model that fits.

What the harness AI does NOT do

It does not make decisions for you. It does not act autonomously without your knowledge. It does not "manage your life."

It does the work you used to do in your head: read the daily log, compress the state, surface the open loops, generate the brief. You read the output and decide what to do.

The harness keeps the decisions in your head and offloads the bookkeeping.

How the three layers interact

Here's how mine runs on a typical Tuesday:

Three layers. One operator. A system that compounds.

The next chapter shows you what the vault actually looks like and how to fit your domains into the twelve numbered categories.