Source of truth

What you'll walk away with: the test that tells you whether your work has a home a next window can actually read, and the decision that keeps that home cheap — what you load every time, and what you look up.

You need: both tabs, about 20 minutes, one finished document you own and could forward outside the firm — the checklist from lesson 13 if you built one — and somewhere you already keep things: a folder, a notes app, a drive. The demo runs in your browser. Nothing to install.

Three weeks from now somebody asks for the current version of the thing you finished today, and what you have is a chat — a wall of text to re-read for what you decided, beside two other copies you are not sure about. Thirty seconds of retrieval first, from memory: what does a new chat start with? Say it, then check it against lesson 05: a window with nothing of yours in it, and nothing from an old chat rides along unless something puts it there. Everything until now happened inside one window; this lesson opens the stage beyond it. What survives when the window closes? Not a feature — a place you keep, a source of truth, and a habit of loading from it.

What the next window can read

The chat that produced a document is a script of how you got there; the document is a thing you can use. Lesson 05 told you to put it somewhere it will live — this lesson is the test of "somewhere". A home is not where you saved it. A home is text a window with nothing of yours in it can read and answer from: paste the text into such a window and ask what this is, where it stands, and what is next, quoting the line each answer comes from and writing UNKNOWN where the text does not say — lesson 08's check and lesson 10's word, doing one job.

Two copies is the same failure as no copy — the next window reads one, and you do not know which. So: one current version, dated at the top, the other copies renamed or moved out.

Lesson 11 made a project a pre-loaded window, and that is one home, not the general case. A folder of plain files and a page in a tool you already live in are others; source of truth has the list. Pick the one you already open every day, and stop choosing.

What you load every time, and what you look up

A home grows, and the whole of it cannot ride in every window — lesson 02's x-ray priced it and lesson 07 showed it going wrong. Predict first: of everything in your home, how much would you want in front of the machine every single time?

It opens with the two panels whole and nothing asked yet. On the left, HOT — RE-READ EVERY SESSION: three blocks — an identity card, the working rules, the open items — under a chip that prices them, ~1,200 tokens — rides in every window. On the right, COLD — LOOKED UP ON DEMAND: eight rows under the line not in the window until a question needs it — four dated days and four fact files. Read that column before you press anything, because its shape is the point: cold is an index, not contents.

Press Why is the coach offline? The cold panel is tagged ANSWERS THIS; one row lights, tagged FETCHED — the 2026-07-12 day — and the strip at the bottom, WHAT GOT LOADED INTO THE WINDOW, reads FETCHED ON DEMAND · Journal — 2026-07-12 and prints the entry it read: one row, not the archive. Press What are you working on right now? and nothing is fetched: the strip reads NO FETCH — ALREADY IN THE WINDOW, and the open-items block on the hot side is tagged ALREADY LOADED — and What database do we use? pulls a fact file instead of a day.

This agent is built to re-read its hot set at the start of every session — its own rule, not something a chat app does for you (what the app's own notes do is lesson 12's). The rule costs it about 1,200 tokens of every window before anybody types, and buys back a fetch on every question the hot set already answers. The grey line underneath is the honest one, and it opens with the register: SIMULATED — static data, no model and no network behind it, a token figure that is the size of this example set and not a measurement of anything you own, and a 2026-07-12 entry that is a real day, lightly anonymized.

Nobody hands you that split; it is a decision you make about your own text. Three or four lines carry the standing cost, and everything else waits to be asked for.

Do this in ChatGPT and Claude

Bring the finished document from lesson 13 — or any document you own and could forward outside the firm. Four steps, the same two prompts in both tabs.

No document from lesson 13? Any document you own works — a plan, a brief, a handover list. If nothing exists, spend five minutes writing lesson 05's three lines — what it is, where it stands, what was decided — into a file, and use that. It is a small home and it passes or fails the same test.

Predict first, in writing: if a chat with nothing of yours in it got only this text, what is the first question it would have to answer UNKNOWN? Write it down; step 3 checks it.

Step 1 — give it a home, in under a minute. A folder of plain files is the home the rest of this path builds on, so if you are choosing, choose that: a folder on your own computer, named for the work, with the document in it as one plain-text file — copy the text into a new file if what you have is anything else. A page in a tool you already live in, or a note you open every day, is a home too; source of truth has the list. The point is not which one; it is that there is one.

The document lives in a panel that will not export. Select the text in the panel and copy it out; paste it into the home. What has to survive is the text, not the panel it was written in.

Your home is a paper notebook. Then the home for this lesson is one text file you type once. A window can read text you paste; it cannot read paper, and neither can the next person.

Step 2 — date it, and make it the only current one. Today's date at the top, and one line under it: what this is, and where it stands. Then deal with the other copies — rename them, or move them out of the home — so the next window cannot read the wrong one by accident.

Step 3 — the stranger-window test. In the first tab, open a chat with nothing of yours in it and paste in the home's text and nothing else. No project, no attachments from the old conversation, no explanation typed above it.

Your home is the project from lesson 11. Then run this step outside it: a chat started inside a project starts with the project, so it is not a window with nothing of yours in it. Open an ordinary chat and paste the text in by hand — the test is whether the text carries the work, not whether the pipes are connected.

The text below is everything I keep about [the thing]. Using only what it says, tell me three things: what this is, where it stands right now, and what the next action is. Quote the line each answer comes from, word for word. Where the text does not say, write UNKNOWN — do not infer it.

Observe: read the quotes against your own text, character for character, then count the UNKNOWNs and look for the one you predicted. Explain: every UNKNOWN is a line your home does not carry — not a failure of the machine, which had exactly what you gave it and nothing else. A wrong answer with a quote is worse news than an UNKNOWN and better news than a fluent one with no quote: it tells you the line exists and says something you did not mean.

Step 4 — the placement decision. Back in the home. Mark the three or four lines you would want in front of the machine every single time — who this is for, what the goal is, where it stands — and put them at the top under one heading. Everything else stays where it is, to be asked for. That is the split the demo made for you, made by you.

Here are the four lines I intend to carry into every chat about [the thing], and below them the rest of what I keep. Tell me which of the four earns its place — for each one, name the question it answers that the rest of the text cannot — and which line I left out of the four that answers more questions than one of them does.

Then steps 3 and 4 in the other tab. The home does not move — one home, two tabs. Paste the same text and send the same two prompts, and read the two UNKNOWN lists against each other: what both tabs could not find is a line missing from your home; what one found and the other did not is one product on one day, and neither is a rule for next month.

---

Check your understanding: six weeks from now you open a chat with nothing of yours in it, paste nothing, and ask it to carry on with a piece of work you finished today. What does it have — and what would have to be true of your home for it to have what it needs?

Check yourself first — then open one good answer+−

It has the app's own hidden instructions, today's date, and whatever notes the app keeps about you — and nothing else of yours. The chat that did the work is a script that stopped riding anywhere the moment you closed it; six weeks later is just a different window.

For that, the home has to be readable as text by that window: one current version, dated, in a place you can paste from — and split, because the whole of it will not fit and should not. A few lines you load every time, carrying who this is for, what the goal is and where it stands; everything else looked up when a question needs it.

Your answer should have hit that nothing carries over on its own, and that the home is judged by what a next window can read out of it, not by where you filed it.

Use this the moment a chat produces something you will need again. One current version, dated, in a place you can name — then the next window starts by reading the work instead of being told about it. Next: The agent workspace — what an agent is made of, and the two things you read on every run.