Focus the window, or widen it

What you'll walk away with: the one decision you make before every task — widen the window or narrow it — and the thing you ask the reply for that tells you which one you needed.

You need: both tabs, about 20 minutes, one real task you owe someone this week, the two or three files that task turns on, and the project you made in lesson 11. If you connected a tool in lesson 14, have it connected; if you did not, run the wide window with memory on and the files attached and the comparison still works. Nothing to install.

First, thirty seconds of retrieval — from memory, not from the page: what is in the window when you press send? Say it, then check it against lesson 03: your message, the app's hidden instructions, today's date, whatever you attached, what a search returned, the notes the app keeps about you. The last four lessons each handed you one more way to fill that same finite window — a project (lesson 11), notes the app keeps (lesson 12), a document (lesson 13), a connected tool (lesson 14) — and every one spends the same budget. So the question left is not how to load but how much, and it runs both ways: a reply that has to draw on your whole working life needs the wide window; one that turns on two files needs the narrow one. What decides it is what the reply must draw on, and you decide it before you open the chat.

Minimum viable context

The craft you have been practicing since lesson 03 has a name, and it is the thing this whole site trains: deciding what goes into the window and what stays out.

Decide the goal first — what you want out is what tells you what must go in. Then split the load the way lesson 03's two-source ledger splits it. What training already baked in, you never paste: the odds already carry how a checklist is built, what onboarding is, how a polite email reads. What only you can supply, you always paste: the actual role, the actual constraints, the numbers, the way your firm actually runs — nothing brings any of that in but the window. And keep the window lean, because it is a budget: everything you add competes with everything else for it (you watched the odds move with the pile in lesson 02's x-ray).

Enough that it isn't guessing; little enough that nothing drowns. When an answer disappoints, before you blame the machine or the vendor, audit the window you handed it: what was missing that only you could have supplied? What was riding along that didn't need to be there? And did the load-bearing thing actually make the window — or did it sit in a big attachment the app read for you? (That question is lesson 04's.) That audit — not a different product — is the move.

Widen it, or narrow it

Before you open the chat you decide one thing: what this reply has to draw on. Write "this answer is right only if it knows …" and fill the blank before you type the task.

If the blank fills with things that live in different places — what you are in the middle of, a preference you never wrote down, a number sitting in another tool — that is the wide task. If it fills with two documents you could name out loud, that is the narrow one. The window follows from your own blank, not from how big the task feels.

Widen it: memory on, so the notes the app keeps ride in (lesson 12); the files the answer turns on attached (lesson 04); a connected tool left connected, so what it holds can be fetched into the window as text when a step runs — lesson 14's receipt says whether one did. Three switches, and you can name all three before you send.

Narrow it: a project holding the two files this job turns on and one standing instruction, and everything else left out — no other files, any note you do not want carried deleted (notes can ride into any window; deleting is the narrow move), the connected tool out of this one. Lesson 11 built it; a project is a scoped window.

Narrowing is not thrift: everything in the window competes (lesson 07, context rot), and two files are the largest thing in a small window, two of forty in a wide one (lesson 08). When a narrow answer comes back short, add the one thing you can name — not everything you have.

Widening is not sloppiness: a wide window can hand you a constraint you would never have thought to paste. When a wide answer uses something you did not supply, ask for the line it used and check it against its source — lesson 08's move, run on a loader instead of a paste.

Neither window is a limit. A wide window that comes back general is the average of everything you loaded — lesson 09's move; a narrow window that comes back thin is missing one thing you can name and add. Both are shapes with a move, and neither is answered by a different product.

Reading the two replies

What you compare is not which reply is better. It is what each one drew on, and where each one went wrong. So ask both replies to list what they used, then check one item against its source. Three honest readings, all three real:

1. Both replies are fine. Then the narrow one is the one you can check: two files and one instruction, every line it could have drawn on nameable. The wide one you would have to audit. Same answer, less to check — keep the narrow window for this task.
2. The wide one has something the narrow one lacks — a constraint out of a note, a number out of a connected tool, a line from a file you would not have attached. Find it at its source first; found, that is the widening move working: the narrow set was short by exactly one nameable thing. Add it to the project and run the narrow window again.
3. The wide one came back about you in general, or contradicts the files. It answered the pile instead of the job (lesson 07). Narrow, and the audit above tells you what to drop.

Do this in ChatGPT and Claude

One prompt, identical in ChatGPT and Claude and in the wide window and the narrow one, because the window is the only variable this exercise moves. Name the task and its files before you start.

Predict, before send 1, and write it down: which reply will use something you did not paste, and which one would you actually send?

Run A — the wide window. In the first tab, a new chat outside any project. Then load it: memory on, the files attached, the tool you connected in lesson 14 left connected. If memory is off (lesson 12), the wide window is the files and whatever you connected: reading 2 can only come from those, and reading 1 is this account's honest result, not a failed run. One send:

Here is the task: [one sentence — the real thing you have to produce this week]. Write it.

Then, under a line that reads WHAT I USED, list every thing you drew on: each file by name, each note you hold about me, each result from a connected tool — and beside each one, the sentence in your answer it produced. Where you had nothing to draw on for part of this, write UNKNOWN rather than filling it in.

Observe: read the WHAT I USED list against what you loaded. Open one file it named and find the sentence, word for word. Anything it used that you did not attach is a note, the connected tool, or a fill: open the notes screen or the tool's own step and find it — lesson 12's move; in neither, it is lesson 10's shape, and the wide window did not do it.

Run B — the narrow window. Same tab. Open the project from lesson 11, strip it to the two files this task turns on, and put one standing instruction on it. Everything else out: the connector off, any note you do not want carried deleted. No project surface in this tab? A new chat with the two files attached and the instruction pasted at the top is the narrow window built by hand. Then the same send, word for word.

Observe: the same read, and one extra thing to notice — the WHAT I USED list is now short enough to check completely. That is not a consolation prize; it is the thing a narrow window buys you. Which of the three readings above did you get?

Explain: the odds are the same odds, the request is the same request, the words are the same words. What changed is what was in the window — and the window is the complete list of what the reply had to work from.

Then both runs in the other tab — the files and instruction are already chosen, so this half is quick. Read the two apps the same way: what each drew on, not which you liked. A note one tab's wide window used and the other's did not is what each account holds today, not a ranking.

Then the line you keep. One sentence: for this task, wide or narrow, and the one thing you would add or take out next time. Keep it where the next window can read it — in the project, or wherever you keep working documents.

---

Check your understanding: two tasks land the same morning. One is "write the onboarding checklist for the new hire on the Riverside account," and the two things that decide it are the account brief and last quarter's checklist. The other is "what should I be worrying about this month?" What do you load for each — and what do you ask the first reply for?

Check yourself first — then open one good answer+−

The decision is the same one both times: what does this reply have to draw on? The checklist draws on two documents you can name, so narrow — those two files and one standing instruction in a project, everything else out. The month question draws on things that live in different places — the notes the app keeps, your other tools, the files you are working from — so widen. Neither window is the careful one. What you ask the reply for is the list: every thing it drew on and the sentence each one produced.

Stage 1 was one idea worked from every side: the window is the only thing the machine has, and each lesson handed you another way to fill it or read it. This lesson comes first in practice: every loader you met is only as good as the sentence you write before you open the chat.

Your answer should have hit the decision as one question — what must this reply draw on — and the check as what the reply says it used, not how good it sounds.

Use this decision before you open the chat. One sentence — this answer is right only if it knows what — and the window follows. The thing you keep from this stage is not an account or a product; it is that sentence, in your own words. Next: Source of truth — where the work lives once the window closes.