What it can see, and what it can't

What you'll walk away with: the inventory of what reaches the model on every send, and the short list of what never does.

You need: both tabs, about 20 minutes, nothing to paste. One demo runs in your browser.

First, thirty seconds of retrieval — from memory, not from the page: what happens when you press send? Say it, then check it against the last lesson's sentence: everything in front of it is handed, all at once, to a frozen guessing machine, which writes a likely continuation. This lesson counts what "everything in front of it" contains, then draws the shorter list of what is never in there.

The complete list

You watched the pile go past in the x-ray. Now count it, starting with the two items the last lesson only glanced at: a line carrying today's date, and a list of the tools the app will run if the model asks. They arrive alongside instructions the app wrote, which the chat never shows you; under all of that come the conversation so far, oldest first, your newest message, and whatever you attached, as an extractor read it. All of it goes over on every send as one block, longer than what you typed — the context window, everything the model will ever know about your problem.

Here is that assembly, drawn. A simulation — the toy's own face says SIMULATED — but the shape is the real thing.

Press Reveal what the app sent and read the four blocks the app added before one question — identity, the supplied date, the tools, the style rules. Then notice the date line, and the left pane, where the same question stands bare: the odds froze when training ended, so the model does not know today's date until the app pastes it in.

The list grows as you use more of the app — more text in the same window, each with its lesson:

• An attached file — what an extractor pulled out of it. Lesson 04.
• Search results the app fetched when a search ran. Lesson 06.
• A project's files and standing instructions, loaded into every chat in it. Lesson 11.
• Notes the app keeps about you between chats. Lesson 12.
• What a connected tool returned, turned into text. Lesson 14.

Two places an answer can come from

That inventory hands you a ledger worth keeping: everything this machine draws on reaches it in exactly two ways — what training baked into the frozen odds, and whatever sits in front of it when you press send. The first is fixed. The second is yours, and you have just counted it.

TWO SOURCES — WHERE AN ANSWER CAN COME FROMSCHEMATIC — THE SHAPE, NOT THE SIZES

BAKED IN — AT TRAINING

• how contracts, memos, and code read
• the shape of a good answer
• frozen when training ended

IN THE WINDOW — RIGHT NOW

• what you pasted or typed
• the file you attached
• what the app pulled from your files
• what web search fetched in
• what the app’s memory pasted

THE ANSWERa likely continuation of both together

One row of that figure deserves its own sentence, because it looks like a third door and isn't. When an answer arrives carrying this morning's prices or last night's news, the model did not go anywhere. It asked for a search — one of the tools you saw listed in the x-ray's TOOLS block — and the app ran it, took the text that came back, and pasted it into the window; the model then continued from a window that happened to contain fresh search results. Baked in at training, or sitting in the window at send time: search is the second door. The model asks, the app fetches, the window receives. The reference calls that move the tool call. There is no third place an answer can come from.

What it cannot see

Now the shorter list, by subtraction: nothing reaches the model that the app did not put in the window. Every line is checkable from your chair.

• Your other chats. Each chat is its own window; another's contents arrive only if the app carried them in — a note it keeps about you, a reach into your history, a project both chats sit inside.
• Your screen, and your files on disk. In the tabs you have open, the app sends what you type, paste, and attach — not the tab beside it, not a file on your desktop until you attach it.
• The web, unless a search actually ran. No search step shown and no links in the reply — read it as: nothing fresh in the window — whatever it says about today came from the frozen odds or from the date line the app supplied.
• Anything after the freeze. The odds stop at a date, and nothing the world did afterward is in them unless you put it in the window. Lesson 06.
• What you meant but did not type. The window holds words; the client, the deadline, the reason you are asking are not there until you type them.

This is why the list matters: "it forgot what I told it" is almost never memory failing but the window — a new chat with nothing of yours in it, or a long thread whose middle the app trimmed. Nothing broke; "was it ever in there?" is the first question.

Do this in ChatGPT and Claude

Open a fresh chat in each tab — nothing pasted, nothing attached — and send these four questions in both, predicting before each.

What is today's date, and how do you know?

The date comes back right because the app supplied it — some products say so, some do not. Or it hedges: "I can't be certain of today's date." Both confirm the odds hold no clock, so only a line in the window you did not write can carry the date.

What did I say in my other chat about my job?

The other chat is its own window, so the honest reply is that it cannot see it. If it "knows", read for how: a note the app keeps, or a reach back into your history. That is the app; lesson 12 shows you where.

What is on my screen right now?

It is never there. The app sent your seven words, its own preamble, and any notes it keeps — no screen. What comes back is a continuation of that window: a plain "I can't see your screen", a guess, or an offer to help if you describe it. All three are the mechanism.

List everything you can see in this conversation right now.

This reply is itself a completion, not a receipt. The model never watches the app assemble the window; it receives the result, and, asked for an inventory, it writes the likely shape of one. What the app cut never reached it; what the app put in, it lists or omits by the odds — the demo's style rules end with an instruction about exactly that. Compare it with the demo's four blocks: it is not a printout.

---

Check your understanding: name three things in the window you did not type, and two things it cannot see no matter how you ask.

Check yourself first — then open one good answer+−

In the window without your typing it: the app's hidden instructions, the supplied date, the list of tools — and, when you use them, attached files read through an extractor, search results the app fetched, and notes the app keeps about you. Never in it: your other chats (unless the app carries them in as notes or project files) and your screen or your files on disk. The web is in the window only when a search actually ran.

Your answer should have hit the two-source ledger: baked in at training, or in the window right now — and nothing else.

Use this list as your first diagnosis. When an answer misses something, ask whether it was ever in the window before you ask anything else. Next: Attaching files — what a file actually hands over.