Keep long AI chats on your rules in ChatGPT, Claude and Gemini

Know how big your window is

As of October 2026, ChatGPT's plan table gives GPT Instant a total window of 27K tokens on Free, 54K on Go and Plus, and 128K on Pro. Reasoning (Thinking) models get 256K on Go and Plus and 400K on Pro; for Free it says "varies". The part left for your own input is smaller, because instructions, memories and reasoning use the same space: OpenAI puts Instant input at about 12 pages of text on Free, 40 on Go and Plus, and 250 on Pro.

Claude on paid plans: 1M tokens on the newest models (Opus 5 and 5.5, Sonnet 5 and 5.5, Haiku 5.5), 500K on some older ones, 200K otherwise. Anthropic does not publish a figure for the free plan. Gemini: 32K tokens without a Google AI plan, 128K on AI Plus, 1 million on AI Pro and Ultra.

To translate: 100 tokens is roughly 75 English words, so 32K tokens is about 24,000 words of conversation, including the model's replies. Long replies fill the window fastest.

Write your rules down once, as a short card

Rules typed in message 3 of a 200-message chat are the first thing to go. Write them as a numbered card instead: who it is for, the voice, the hard rules (words to avoid, formats, spelling), and what to do when unsure. Keep it under about 300 words. Short rules are cheaper to restate and easier for the model to follow.

Ask Claude or ChatGPT to turn your messy notes into the card, then edit it yourself.

Turn my notes below into a rules card for an AI assistant. Format: a title line, then numbered rules, each one short and testable ("British spelling", "no emoji", "captions under 125 characters"), grouped under Audience, Voice, Hard rules, Formats, and When unsure. Under 300 words. Keep my wording where it is clear. Flag any rule that contradicts another. My notes: [PASTE]

Put the card where every chat gets it

Put the card in a project's instructions, not in a message. In ChatGPT open the project, select the more options menu (•••), then Project settings; project instructions apply only inside that project and override your global custom instructions. In Claude, add it as the project's instructions. In Gemini, use a notebook's instructions, or a Gem's instructions (Google is turning Gems into skills for personal accounts from November 2026).

Rules that apply to everything you do go in the global setting instead: ChatGPT Settings > Personalization > Custom instructions (1,500 characters on Free and Go, 5,000 on Plus and above), Claude Settings > Your Claude, where instructions apply to all your chats.

Keep reference material in project files, not in the chat

Brand guides, product sheets and past scripts pasted into a chat eat the window and scroll away. Upload them to the project instead. ChatGPT projects take 5 files on Free, 25 on Go and Plus, and 40 on Pro and Business, 10 at a time. Claude project files can be up to 30 MB each; on paid plans, when the project knowledge nears the window Claude switches to retrieval and pulls in only the relevant parts, which raises capacity by up to 10 times. Gemini notebooks take up to 600 sources depending on your plan.

Split one big document into a few focused files (voice, products, past winners). Retrieval finds the right passage more reliably in a file with a clear topic.

Use memory for facts about you, not for the job's rules

Memory carries facts between chats: your name, your business, how you like answers. It is not a reliable place for a detailed rule set. OpenAI says memory "does not retain every detail", and every app decides for itself what is relevant.

ChatGPT: Settings > Personalization > Memory, with saved memories and chat history as separate switches; a project can be set to project-only memory. Claude: memory is on by default for Free, Pro and Max; each project has its own memory, and you can turn off "Use account memory" on a project to keep it separate. Gemini: Settings & help > Personal context, where you can add or delete saved info.

Check what each app has remembered every month or so. Stale memories (an old price, last season's campaign) cause as much drift as forgotten rules.

Restate the key rules in long chats

Models use the start and the end of a long input best. A 2023 study (Liu and others, "Lost in the Middle") found accuracy was often highest when the relevant information was at the beginning or end and dropped when it sat in the middle. So in a long chat, put the two or three rules that matter for this task at the end of your message, or ask the model to check its draft against the card before it answers.

Before you answer, re-read rules [3, 5 and 7] from the project instructions. Write your draft, then check it against those rules and fix anything that breaks them. At the end, list each rule with "kept" or "changed: [what you fixed]".

Spot drift, then hand over to a fresh chat

Signs a chat has outgrown its window: it breaks a rule it kept earlier, asks for something you already gave it, or contradicts a decision from the start. Claude on paid plans summarises earlier messages automatically as it nears the limit (code execution must be on), and those long chats use more of your usage limit. OpenAI and Google do not say exactly what their apps drop.

Don't argue with a drifting chat. Ask for a hand-over summary, start a new chat in the same project, and paste it in. The new chat gets the project instructions and files again, plus a clean summary instead of 200 messages.

We're moving to a new chat. Write a hand-over note I can paste into it: 1) the goal in one line, 2) decisions made, with the final wording of anything we agreed, 3) open questions, 4) rules I gave in this chat that are not in the project instructions, 5) the next step. Under 400 words. Don't include anything we discarded.

ChatGPT set-up

Projects are on every plan, including Free. Select New project in the sidebar, then the more options menu (•••) > Project settings to add instructions and choose Default memory or Project-only memory. With project-only memory, chats use only that project's chats and files, not your saved memories.

Add files from the project's sources list (5 on Free, 25 on Go and Plus, 40 on Pro). You can also save a good answer as a source: open the message menu and select Add to project sources. Move an old chat in with Move to project; it then inherits the project's instructions and files.

Claude set-up

Projects are on every plan; free accounts can create up to five. Add the rules card as project instructions and upload reference files to project knowledge (30 MB per file). On paid plans Claude switches to retrieval automatically when the project gets large.

Each project keeps its own memory. To stop your general memory leaking into a client project, turn off "Use account memory" under Memory on the project's page. To run one chat with no memory at all, open a new chat, click + and turn Memory off before the first message.

Gemini set-up

Notebooks gather chats, sources and instructions for one project: open the sidebar, then under Notebooks select New notebook, and use Add sources for files, Drive files, websites or pasted text. Google says notebooks are still rolling out, so you may not have them yet.

For a reusable assistant, create a Gem with instructions and files under Knowledge. From November 2026 Google is converting Gems on personal accounts into skills automatically. Facts about you go in Settings & help > Personal context.