See exactly what you get

# ContextHouse

**The house where your AI's context lives. It sets itself up with you, and stays yours.**

A working operating system for people who've stopped asking AI occasional questions and started doing their actual work inside it.

*(You're reading plain text files. Any editor opens them, Notepad included, and any Markdown viewer makes them prettier. Nothing to install to read on.)*

**Start here:** point an AI at this folder (§ Setup below), then say **"let's set up my vault"**. It introduces itself and sets itself up with you. The illustrated handbook (every room, every command, the whole logic) is `Overview.html`, in this folder; open it in any browser.

## What is a second brain, and why run one now

A second brain is one place outside your head where your work and your life admin live as plain files: this week's priorities, your projects, the decisions you've settled, the people who matter, the paperwork of your life, what you actually did each day. The idea is old, and its classic promise ("capture everything, find it later") mostly produced note graveyards. People built elaborate systems, filed diligently for three weeks, and never opened them again.

What changed: AI can now *operate* files, not just chat about them. The moment your notes are plain Markdown in folders, they stop being storage and become **working memory for an assistant**: one that files your loose thoughts, keeps your priorities in view, checks your own past decisions before advising you, drafts in your voice because it has read who you are, and never starts a conversation from zero again.

That's the shift this vault is built for. Whether you're building something (a business, a product, an independent practice) or simply running a full life (a household, a flat you rent out, paperwork that never stops), the alternative isn't "no system". It's re-explaining your context to a chat window every single morning, and losing every decision you've ever made to scroll-back.

This package is the structure, rules, and processes pre-built and tested in daily use. Think of it as a well-built empty house: rooms, wiring, and house rules work from day one; the furniture is yours. No architecture phase, no six-week setup, no life-purpose homework before you're allowed to write a note.

---
type: status
trust: untrusted
ai_generated: false
sensitivity: private
updated: 2026-06-19
---

# Now

The weekly command dashboard. One screen, ~100 lines max. Set at the weekly review, read every morning. This file answers: *what matters this week, what's active, what's next.*

> Everything below is example content from the starter vault — replace it with your own at your first weekly review.

## Top Priorities (Week of 2026-06-22)

1. Publish the new homepage — [[Projects/portfolio_relaunch/00_PROJECT]]
2. Quarterly tax prepayment (Monday — due 2026-06-25)
3. Draft the "Site Tune-Up" package page

Fallback if blocked: clear the § Open list in `Tasks.md`.

## Active Projects

- **Portfolio relaunch** (Revenue Engine). [[Projects/portfolio_relaunch/00_PROJECT]] — Next: pick the hero variant (Monday), then build the homepage (step 4).

## Recently Completed

- (none yet — completed projects get one line here for ~2 weeks, then drop off; detail lives in the project folder)

## Upcoming

- 2026-06-25 — quarterly tax prepayment due.
- 2026-07-01 — client retainer renewal conversation.

## Current Status

Steady client load, one proposal out. The relaunch is the bet for attracting better-fit clients in Q3.

---

> Keep this file lean: no project history, no long recaps, no session chronology. If a section grows past a few lines, the detail belongs in a project file.

## Purpose

Own the guided setup: the conversation that turns this vault from a well-built empty house into the owner's own second brain. Not one long onboarding: several short sessions, days or weeks apart, each optional, each ending with something real in the files. The promise on the box: **a vault that sets itself up with you.**

Two things this skill is not:

- **Not a gate.** The vault works from the first minute without it: "good morning" runs, capture works, every skill fires. The wizard is the guided move-in, never an unlock sequence. Nothing is locked behind a stage, and skipping the wizard entirely is a fine way to live here.
- **Not a form.** Facts are captured from conversation, in the owner's words. No questionnaires, no interviews, no homework between sessions. A stage that ends with "go fill in X before next time" has failed.

Progress lives in `System/Setup_State.md`. The wizard reads it at the start of every setup session and resumes exactly where the last one stopped, same AI tool or a different one, tomorrow or next month. The files carry the state; no chat history required.

| # | Name | Length | What the owner has afterwards |
|---|------|--------|-------------------------------|
| 1 | Moving in | ~15–20 min | Knows what this is and why it's theirs; `About_Me.md` started; saw capture + sweep work once |
| 2 | Your rhythm: pick your rituals | ~10 min | Knows the core rhythm (sweep + the weekly check); the rituals that fit them switched on, or consciously none |
| 3 | Tasks & projects | ~10–15 min | Knows the 30-minute line; `Now.md` shows their real week |
| 4 | People | ~10 min | First person note exists, in their words |
| 5 | Files in, files kept | ~10 min | Knows the file inbox, the Archive, the naming, and what the AI sees |
| 6 | Make it yours | ~10–15 min | Worked example retired; their own principles started |
| 7 | Your feature set, revisited | ~5–10 min | Feature set matches real use; setup complete |

## Naming rule for files (deterministic — findable without the AI)

```text
YYYY-MM-DD_plain-language-name.ext
```

- **Date** = the document's own date (letter date, invoice date, policy start) when readable. Not findable → today's date.
- **Name** = the plain words the owner would search for. Lowercase, hyphens, ASCII, original extension kept.
- Examples: `2026-01-15_liability-insurance-policy.pdf` · `2026-07-31_payslip-july.pdf`
- Collision at the destination → append `-2`. Never overwrite.

Date first means every folder sorts itself into a timeline, and a plain file search finds anything by the words in its name. No AI required to retrieve what the AI filed.

## Sweep — Inbox (N lines, M files)

**Lines:**
1. "<original line>" → Tasks.md § Open § <category>. Reason: <short>.
2. "<original line>" → propose Now.md § Top Priorities. Reason: <short>.
3. "<original line>" → Inbox/YYYY-MM-DD_<slug>.md (raw / unclear). Reason: <short>.

**Files:**
4. `Scan_0042.pdf` → `Areas/Insurance/2026-01-15_liability-insurance-policy.pdf`
   Liability policy, dated 15 Jan 2026. Original kept in Archive/Inbox_Originals/.

**Couldn't read:**
5. `altvertrag.tif` — this tool can't read TIF images (see System/Model_Recommendations.md).
   Tell me in one line what it is, or I'll leave it here flagged.

Apply lines + propose the rest? (y/n) · Move files? (all / numbers / none)

## Purpose

The vault's caretaker. When something feels off — clutter piling up, a surface you've stopped opening, a system that fights you instead of helping — the Steward looks at the whole vault, tells you honestly what it found, and proposes the smallest fix that helps.

Two jobs, one flow:

1. **Health check** — find clutter, misfiled items, duplicates, stale notes, and missing links. Routine maintenance, like airing out a house.
2. **System fit** — a short conversation about what's working for you and what isn't, ending in **one** small adjustment. Never a rebuild.

The Steward keeps the system serving you — it never grows the system on its own. The fix for "I don't use X" is almost always to shrink or archive X, not to build a better X. And when the owner *wants* to grow — a new area to track, a recurring routine, a new capability — the Steward is the gatekeeper of the growth ladder (`System/How_It_Grows.md`): practice by hand first, codify after it repeats, specialize only when load-bearing. Growth is proposed through the Steward, never self-installed.

## Vault health check — YYYY-MM-DD

**Healthy:** <2-3 things genuinely in good shape — always first>

**Findings:**
1. <finding> → proposed fix: <one line>
2. <finding> → proposed fix: <one line>
...

**The one adjustment:** <the single system change proposed, with why — or "none needed">

Apply which fixes? (all / numbers / none)

---
type: decision
trust: trusted
ai_generated: partial
sensitivity: private
created: 2026-06-12
updated: 2026-06-12
status: accepted
---

# No client work on Fridays

> **Worked example.** Fictional sample decision showing the format: one decision per note, immutable once accepted, superseded by a new note if it changes. Delete once you've seen the pattern.

## Decision

Fridays are for my own business (product work, marketing, the weekly review) — no client calls, no client deliverables scheduled on Fridays.

## Context

Client work was eating every weekday; the portfolio relaunch had made zero progress in four weeks. The pattern: anything without a deadline loses to anything with one. So the fix is structural, not motivational — a protected weekly build window.

## Options considered

1. **Friday block (chosen)** — one full protected day. Predictable, easy to defend to clients ("I book calls Mon–Thu").
2. Daily 90-minute morning block — dies the first week a deadline lands.
3. "When there's slack" — that's the current failure mode, not an option.

## Consequences

- Client capacity drops ~20% — priced into the new packages.
- The weekly review gets a reliable home (Friday evening).
- Revisit if two consecutive months show Fridays being regularly broken — then supersede this note with whatever actually works, rather than silently ignoring it.

## Related

- [[../Projects/portfolio_relaunch/00_PROJECT]] — the project this window protects.

---

> AI involvement: Drafted by AI from my notes on 2026-06-12; decision made and locked by me. Reviewed by me on 2026-06-12. Trust: trusted. Sources checked: n/a (personal decision).

# What the AI Sees

The honest note about your data. Two minutes, worth it.

## Reading is sending

Your vault lives on your machine. It has no app, no account, no server — nothing here phones home.

But the AI is not on your machine. When your AI assistant reads a file — to sweep your inbox, to file a payslip, to summarize a note — the content of that file is sent to the company that runs the AI (Anthropic for Claude, OpenAI for ChatGPT, and so on). It has to be: that's how these tools work. **Reading is sending.**

That's not a leak and not a scandal. It's the deal, and you should know you're making it.

## What it means in practice

- **What the AI never reads never travels.** The vault sitting in a folder sends nothing. Only files opened in a session go anywhere.
- **Working notes are the normal case.** Your tasks, projects, and ideas are what the AI is *for* — reading them is the product working as intended.
- **Dropped files are read once to be filed.** Putting a document in `Inbox/` is your go-ahead for the sweep to read it. That's the trade: you never sort files again; the AI looks at each one, once, to sort it for you.
- **What happens on the provider's side** is governed by that company's privacy policy and your settings there — most tools let you turn off "use my data for training". Worth five minutes in your AI tool's settings, once.

## The counter-move: `Sensitive/`

For material the AI should **never** read, this vault has a standing pattern (`System/Rules.md`, Iron Rule #3):

1. Create a folder called `Sensitive/` the day you first need it.
2. The rules instruct your AI to never read it by default — the instruction already ships in `CLAUDE.md` / `AGENTS.md`.
3. Add the folder to your AI tool's ignore configuration too (most tools have one), so the tool itself also skips it. Two layers: the rule, and the tool setting.
4. Put such files **straight into `Sensitive/`** — never through `Inbox/`, because inbox files get read on sweep.

What belongs there: ID scans, medical records, anything with passwords or account numbers, and anything that simply feels like nobody's business. If you hesitate, it belongs there.