How to stop ChatGPT and Claude losing the plot halfway through a long chat

Every model has a limit on how much text it can hold in one conversation, called the context window. When you pass it, the app does not stop you. It quietly drops or compresses the oldest parts. The symptoms arrive before any warning: it stops following a format you set early on, it answers a question you asked ten messages ago, it summarises instead of doing, and it gets slower.

Long pasted documents fill the window fastest. One 40-page PDF can use more room than an afternoon of typing. So the chats that "go wrong" are usually the ones where you pasted the most.

The fix is not a longer chat. It is to stop carrying facts inside the chat at all. All three assistants have a container for this. ChatGPT calls it Projects: a folder with its own instructions and files that every chat inside it can read. Claude calls it Projects too, with Project knowledge for files and instructions. Gemini's equivalent for documents is Gemini Notebook (the product that was NotebookLM until July 2026), which answers only from the sources you upload.

Create one per piece of work. Put the brief, the style rules and the reference documents in it. Then every new chat inside the project starts already knowing them, and you never paste them again. Anthropic's own usage page says the same thing for a different reason: files in a project are cached, so they also cost less against your limit.

When a chat is getting full, do not push on. Ask the model to write a handoff note: what the task is, what has been decided, what is still open, and the exact rules you set. Then start a new chat, inside the project, and paste the note as the first message.

This is the single most useful prompt I know. The model is good at summarising its own conversation, and a fresh chat with a clear 300-word brief beats a tired chat with 40,000 words of history every time. In Claude Code the same idea is a built-in command, /compact, which does the summary for you.

Instructions at the top of a chat fade. If a rule matters, say it again in the message where it matters. "Keep it under 120 words" in the message you want to be short works better than the same line set an hour ago.

Better still, move the rule out of the chat. Put it in the project instructions, or in the assistant's standing instructions: Custom instructions under Personalization in ChatGPT, the Project instructions in Claude, and the Saved info page in Gemini. Those are read on every message, so they do not fade.

If the job is a long document, do not paste it into a chat at all. Upload it. In ChatGPT and Claude, attach the file to a project. For a set of documents you will question for weeks, Gemini Notebook is built for exactly that: it answers only from what you uploaded and shows you the passage it used, which also solves the made-up-answer problem.

Then ask narrow questions. "What does section 4 say about notice periods?" gets a correct answer with a citation. "Summarise this contract" gets a summary that misses section 4.