How to keep a document in context in ChatGPT

The frustrating part of a long document session is that it works well for a while. Precise answers, real quotes, the right section names. Then somewhere around the fortieth message it turns generic and starts hedging, and nothing tells you the moment it changed.

Nothing pins a file in place

An attachment lives inside the conversation, and the conversation has a ceiling on how much can sit in front of the model while it writes one reply. Your messages count toward that total, its replies count, and so does the text pulled out of your file. When the total stops fitting, the oldest content goes first, which is almost always the document you uploaded at the top.

That ceiling is a separate thing from the file limits, and much smaller. A document can sit comfortably inside 2 million tokens and still be pushed out of a chat an hour later. OpenAI publishes no number for the conversation ceiling, because it varies by model and plan.

There is no pin, no lock, no priority flag. So every technique below is one of two moves: spend the room more slowly, or put the document somewhere that is supplied fresh each turn.

Set the chat up so the document lasts

The document survives longer when less else is competing with it.

Upload first, ask second. The file is the heaviest thing in the chat. Put it in before the discussion starts, so everything else stacks on top of it rather than arriving once the room is spent.

One document, one job. Tangents are not free. Every unrelated question, and every answer to one, is room the document is competing with.

Ask for bullets, not essays. The underrated one. Long replies consume the same budget your file does, so a chat full of six-paragraph answers fills far faster than one full of tight lists.

Send the section, not the book. A twelve page chapter stays in view far longer than a four hundred page report, and usually gets a sharper answer anyway.

Take a working extract early

Do the extraction in the first few turns, while the file is certainly present.

From the attached document, list the section headings in order, then every figure, date and named party, each with the section it appears in. Quote exactly. No commentary.

Keep that outside ChatGPT. It is a few hundred words standing in for a few hundred pages, and it is what you paste into the next chat when this one is finished.

Re-anchoring when the answers drift

The usual tells are a question you already answered coming back, a rule you set being quietly dropped, or a specific detail returning slightly wrong.

1. Ask what it can still see

   "Without opening the file again, list what you can currently quote from the attached document." Read the answer as a lead, not as evidence. The model is never told that content was removed, so it often describes the file from what the conversation says about it.

2. Check the answer with an exact quote

   Pick something distinctive that has not been repeated in the chat, and ask for it word for word. A paraphrase can be reassembled from later mentions. A verbatim line either survives or it does not.

3. Re-attach the file with a pointer

   Attach it again to a new message and say what you want from it now, naming the section. That puts the text back in view and aims it at the current question in one move.

4. Restate your constraints in one line

   Standing rules about format, tone or scope go back in at the same time, which moves them to the recent end of the chat where they are safe again.

Give the document a home outside the chat

If you keep coming back to the same material, stop relying on a conversation to hold it.

None of these enlarge the conversation ceiling. What they change is the cost of starting over: when the document sits in a Project, leaving a spent chat costs seconds rather than another upload.

When to start fresh instead

Some signals mean the chat is finished and no amount of re-anchoring brings it back. You are re-attaching for the second time. It has started confusing two sections with each other. It asks for the file after you have already given it twice. Answers get longer and less specific at the same time, which is usually the model padding around a gap.

The instinct at that point is to argue with it, correct it, tell it to try harder. That spends more room on a chat that has none left. Take the handover instead, while the file is still readable:

Summarise what we have established as a short brief I can paste into a new chat. Include the decisions made, the constraints I set, and anything still open. Do not repeat the document text.

That last sentence matters. Without it you get pages of quoted source, which is exactly the weight you are trying to leave behind. Open a new chat, attach the document first, paste the brief underneath, and you are back to where you were in two messages rather than forty.

The decision

Keeping a document in context is not a setting you turn on. It is spending the room carefully at the start, extracting what you need while the file is definitely there, and reading the drift as information rather than as a fault to argue with.

When it starts slipping, re-attach once and verify with a quote. When it slips again, take the brief and start clean. Fighting a full conversation costs more than leaving it, and the answers you get while fighting are the ones least worth keeping.