Use ChatGPT Projects to keep one piece of work out of your other chats

Example result.

Project: Tender — Halden Water
Memory: Project-only (set when the project was created)
Files: invitation to tender, last year's response, pricing sheet, capability summary
Instructions: project files only; "not in source" for gaps; no executive summary
Shared with: nobody

A note on memory. ChatGPT can carry context between conversations: from saved memories, past chats and other sources, depending on your plan. With Default memory, a project can use your saved memories and, on some plans, chats outside it. With Project-only memory, it uses only its own chats and files, and chats outside can’t reference it.

1. Create the project and choose its memory first

Select New project in the sidebar, name it and pick an icon and colour. Choose Project-only memory when you create it if the work needs to stay separate.

You can change it later: open the project, select the more options menu (•••), choose Project settings, and under Memory select the option, then Save. OpenAI says the switch can take a few hours to take effect, so setting it at the start avoids the gap. There’s no global switch: each project is set on its own. In Business or Enterprise, the option appears only when Memory is on for both the workspace and your account.

Why this works: the memory setting decides what the project can draw on and what can draw on it. Choosing it first means nothing leaks in the meantime.

What you should see: the new project in the sidebar, with Project-only memory shown in its settings.

2. Add the files and write the instructions

Upload PDFs, spreadsheets, documents or images, or paste text. Up to 10 files can be uploaded at once.

Then open ••• › Project settings and add instructions. They apply only in this project, and they override your global custom instructions, which is the quietly important part. If your account-level instructions say “be concise and skip the caveats”, that’s exactly what you don’t want on work where the caveats are the job.

This project concerns one tender response. Use only the project files and what I paste into the chat for factual claims. Ask before using any other source.
When a figure, date or commitment isn't in the files, write "not in source" rather than estimating it.
Default output: short sections with headings. No executive summary unless I ask.

Why this works: the instructions set the rules for this work without changing how ChatGPT behaves everywhere else.

What you should see: the files listed in the project’s sources, and the instructions saved in its settings.

3. Work inside the project, every time

Open the project, then start the chat. If you began in an ordinary chat, drag it onto the project in the sidebar, or use the chat’s menu and choose Move to project; it then inherits the project’s instructions and files. Some chats can’t be moved, including chats created with a custom GPT, so start a new one in the project instead.

Try one small task:

Using the project files, list the information we still need to complete [the piece of work]. For each gap, name the file and heading. Separate missing information from conflicting information. Don't invent dates, owners or approvals. If you can't read a file, say which one rather than guessing from its title.

[the piece of work] is, for example, “the tender response”.

Why this works: it proves the project’s files and instructions are in use. Check one “missing” item yourself: it may be there under a different heading.

What you should see: a gap list naming files and headings, inside a chat listed under the project.

Check the result

• Project settings shows the memory option you chose.
• Your working chats appear under the project in the sidebar, not in the main chat list.
• Answers follow the project instructions, such as “not in source” for gaps.
• At least one “missing” item has been checked against the files.

Don’t treat a right or wrong answer in a different chat as proof of separation: the model may guess, or the same information may exist elsewhere.

The usual failure: you start an ordinary chat in a hurry and paste project material into it. That chat follows your general settings and instructions, not the project’s. Spot it by looking at where the chat sits in the sidebar. Move it into the project, delete any outside copy, then ask:

This chat has just been moved into the project. Redo your last answer using only the project files and instructions, and list which files you used.

Fictional example. A bid manager creates Tender — Halden Water, sets project-only memory at creation and uploads four files. Her first gap list names eleven tender questions the capability summary doesn’t answer. Two of them are answered after all, under a different heading: a search is a search, not an oracle.

A week later, in a hurry, she starts a chat from the sidebar instead of opening the project and pastes a section of the pricing sheet to be reformatted. That chat follows her ordinary account settings and global instructions. On a personal plan, its content could be used to improve models if Improve the model for everyone were on in her data controls.

She moves the chat into the project and runs the repair prompt. The reformatted section now follows the project’s “not in source” rule.

• The repair is partial. Moving the chat doesn’t change how the message was handled while it sat outside. Deleting the outside copy is the other half, and her organisation’s process applies if the material went somewhere it shouldn’t.
• The lesson is operational. Make the project the thing you click, not the thing you remember.
• A tidy project didn’t guarantee every file was read. The two false gaps show why you check.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Sources

• OpenAI Help: Projects in ChatGPT
• OpenAI Help: Memory in ChatGPT
• OpenAI Help: Data controls in ChatGPT