For work tied to your facts, AI needs the source in front of it in the conversation, so the right context is the material that decides the answer: the policy, the figures, the email thread or the earlier draft the job depends on. Leave out what the job does not need, because extra material costs attention and can pull the answer off course. Never include passwords, access keys or confidential details the job does not require. And do not assume a new chat knows what an earlier one did unless you have checked that a memory or project feature is carrying it.
Start with the source that decides the answer
Before you word the prompt, ask what the answer depends on. If you want a reply about whether a client is owed something, the answer is in their agreement and in what actually happened, not in the AI’s general knowledge of how agreements usually work. Without that source, you get an answer that reads well and could have been written about any client.
A useful way to sort what you have:
| Kind of material | What it does | What to do |
|---|---|---|
| Required | Without it the job cannot be done right at all. | Always include it. |
| Helpful | Makes the answer better, but the job survives without it. | Include it if it is short and relevant. |
| Noise | Costs attention and buys nothing. | Leave it out. |
| Never | Passwords, keys, and confidential or personal data the job does not need. | Keep it out, whatever the job. |
An example: a reply about a late delivery
You work in account management at a sample office-furniture supplier. A client writes in: their order arrived a week late, and they want to know whether they are owed a credit. You want AI to draft the reply. Here is what you could attach, and where each piece falls:
- The client’s service agreement (required). The credit terms are here. Without it the AI will describe what credit terms usually say.
- The order’s shipping record (required). It shows when the order was promised and when it arrived, and whether the delay was on your side.
- The last few emails with this client (helpful). They set the tone and show anything already said about the order.
- Your full order export for every client (noise). Thousands of rows the reply never touches, full of other clients’ details.
- The company handbook (noise). Long, general, and not where this client’s terms live.
- Your login for the order system (never). It does not help the AI read anything, and it should not be in a chat at all.
Notice that the noise is not harmless. A big export adds personal data the job never needed, and a long general document gives the model more to draw on that is not about this client. Handing it everything is its own kind of mistake.
Tip: If you are not sure whether something is allowed in your AI tool, your company’s AI or data policy decides, not the tool. When in doubt, describe the situation instead of pasting the document, or remove names and account numbers first.
The context window, in plain words
The context window is everything the model can read while it writes a response: your messages, its earlier replies, any files you attached, and the response itself. Anthropic’s documentation describes it as the model’s “working memory”, separate from the huge body of text it was trained on. It is measured in tokens, small chunks of text that are roughly parts of words.
Two things follow. First, inside one conversation the model can see everything that is still in that window, including old material you have finished with. If a chat holds last month’s figures and this month’s, nothing tells it which one you mean unless you say so. Second, more is not automatically better. The same documentation notes that as the amount of text grows, accuracy and recall degrade, and that choosing what goes in matters as much as how much fits. It also notes that chat apps can drop the oldest parts of a very long conversation to make room.
In practice: label what is current, and when a chat has grown long or mixed, start a new one with only what the job needs.
What a new chat remembers
On its own, a new chat starts with nothing from your earlier chats. What it can carry depends on the product and on your settings, and these features change often, so check your own:
- Claude has a memory feature that is on by default on its individual plans; on Team and Enterprise plans the organization’s owners decide whether it is available. Paid plans can also search past chats. You can turn either off in settings, and an incognito chat is not saved to your history and does not feed memory. Projects are workspaces with their own chats, files and instructions, and each project has its own memory.
- ChatGPT has two memory settings: saved memories, details it keeps for future chats, and referencing your chat history. Either can be turned off. A temporary chat does not create memories. Projects keep related chats and files together, and can be set so their memory stays inside the project.
The habit that works in every case is to check rather than assume. When an assistant says “I’ll remember that,” that sentence is generated like any other, and it sounds just as sure whether or not anything was saved. If a rule or a fact matters for a job, either confirm that memory or a project is carrying it, or paste it in again.
Connected sources
Both assistants offer connectors that let them search and read your other tools. Claude’s connectors include Google Drive, Gmail and Google Calendar, and ChatGPT’s connectors include services such as Google Drive and SharePoint. Which ones you have depends on your plan and, at work, on your admin.
A connector changes what the assistant can reach, not what it actually looked at. Naming a file in your message is not the same as sending it: without a real connection, the assistant has only the text in the chat, and it may answer anyway. When a connected answer matters, look for which file it says it opened or cited, and open that file yourself. Connectors also reach as far as your account does, so only connect tools you trust, and only to what the job needs.