ChatGPT Projects: How to Use Them Properly

TL;DR: A ChatGPT Project groups chats, files, and shared instructions for one ongoing effort. Project instructions override your global custom instructions, not add to them. File limits and memory behavior vary by plan and by mode, and two of OpenAI's own help pages disagree on the numbers. Projects organize your chats. They don't give you a reusable prompt library.

What Is a ChatGPT Project, Actually?

A Project is a workspace that bundles three things for one piece of ongoing work: the chats you have inside it, any files you upload as reference material, and a set of project instructions that apply only there. OpenAI's own "Projects in ChatGPT" help article frames the point plainly: group chats, files, and instructions together so ChatGPT stays on-topic across sessions instead of you re-explaining the setup every time (help.openai.com, checked 4 September 2026).

Two facts from that same article matter before you build anything. First, "Projects are available to all free and paid subscription types globally" — this is not a paid-tier feature. Second, "Users can create an unlimited amount of projects", so there's no reason to cram unrelated work into one Project to conserve a quota that doesn't exist.

A Project is not a chat with a folder icon. It's the one place in ChatGPT where instructions, files, and a growing set of related conversations stay linked, and where moving an existing chat in changes what that chat can see going forward. That last part surprises people, and it's the reason using Projects properly is a real skill rather than just clicking New Project.

How Do You Set Up a Project Properly?

Five steps, in order, and the order matters because instructions should exist before you start chatting inside the Project.

1. Create it and name it well. Click New Project in the sidebar, give it a name specific enough to tell it apart from a similar one six months from now, and pick an icon and color so it's identifiable in a crowded sidebar at a glance.
2. Write project instructions before your first chat. Open the three-dot menu on the project, choose Project settings, and add instructions scoped to this specific effort: role, tone, constraints, and anything the model should assume every time.
3. Restate anything from your global custom instructions that still applies. This is the step almost everyone skips, and it's the one that actually matters, covered in full below.
4. Add files as reference material, not as a dump. Upload only what chats in this Project will actually need. Only 10 files upload at once, per OpenAI's article, so batch large sets rather than trying to drop in fifty at once.
5. Move existing chats in deliberately. Drag a chat onto the Project or use its menu to move it in. After moving, a chat inherits the Project's instructions and file context, which is exactly the behavior you want if the chat belongs there, and exactly the behavior that surprises you if it doesn't.

Two features go beyond plain file uploads and are easy to miss entirely. You can paste a link to a Slack channel or a Google Drive file or folder straight into a project's sources area, and ChatGPT treats it as project context without you downloading and re-uploading anything, though the Google Drive app doesn't sync automatically once added this way, so a source file that changes later won't update on its own. And if ChatGPT produces something worth keeping, a summary, a decision note, a draft, you can save that exact response as a project source from the message menu instead of copying it out to a separate file first.

If you do hit your file limit, OpenAI's own guidance is simple: remove older or unnecessary uploads, combine file data, or split the work into multiple projects rather than trying to force everything into one.

Here's a template for step 2 that covers the parts people forget:

ROLE AND SCOPE
This project is for <specific, ongoing effort>. Nothing outside this scope.

TONE AND FORMAT
<how responses should read, structured lists vs prose, length preference>

FACTS THIS PROJECT SHOULD TREAT AS FIXED
<the 2-3 things that stay true across every chat in this project>

CARRIED OVER FROM MY GLOBAL CUSTOM INSTRUCTIONS
<restate anything from your account-wide instructions you still need here —
a project does not inherit these automatically>

WHAT TO NEVER DO IN THIS PROJECT
<the mistake you'd have to correct most often if you didn't say this>

That fourth block exists because of one sentence in OpenAI's own documentation, and it changes how every Project should be written.

What Does Project Memory Actually Remember?

Projects have their own built-in memory on top of whatever account-level memory you have enabled, and OpenAI lets you choose between two modes per project: default memory and project-only memory. Understanding the difference is what stops one project's context from silently bleeding into another, or a shared project from exposing more than you intended.

With default memory, a Project's chats can reference your other conversations, both inside and outside the Project, provided the other one isn't itself set to project-only. With project-only memory, chats inside the Project can reference only each other. Nothing from outside gets in, and nothing inside leaks out.

Two behaviors are easy to miss. First, there's no account-wide switch for this. OpenAI's FAQ states plainly: "Each project’s memory must be set individually, and also respects your personal and workspace memory settings." A habit you set on one project says nothing about the next one you create. Second, sharing a project changes its memory mode for you, not just for guests: "Shared projects are automatically set to project-only memory and cannot be switched to default memory." If you've been relying on a personal project referencing your other chats and then invite a collaborator, that access is gone the moment you share it, permanently for that project.

One more detail worth knowing before you rely on a mode change: switching a project's memory setting is not instant. OpenAI's article notes it "may take a few hours to take effect", so don't test a sensitive boundary and assume the old behavior is already gone.

Can You Share a Project With Your Team?

Yes, and this is newer than most write-ups about Projects assume. Sharing rolled out to every plan, not just Business and Enterprise. As of an October 2025 update, OpenAI's article states that project sharing "is available to all ChatGPT users, including for Free, Plus, Pro, and Go users globally on web, iOS, and Android."

Collaboration comes with its own limits, separate from the per-project file caps above:

• Pro: up to 40 files and 100 collaborators
• Plus and Go: up to 25 files and 10 collaborators
• Free: up to 5 files and 5 collaborators

Invitees get one of two access levels. Edit access lets a member update instructions, upload or remove files, and invite others, but not remove existing members. Chat access lets them see and use the project's chats, files, and instructions without touching any of it. As the owner, you keep full control: you can create chats, update instructions, add files, invite or remove members, and delete the project outright.

One collaboration feature worth knowing about is branching. Any member can open an existing chat and continue it in a new direction without disturbing the original thread, useful when you want to test an alternate approach on top of a teammate's work rather than starting over. And if you ever leave a shared project, you're offered the option to copy your own chats out first, into a new personal project, before your access ends.

Should You Use a Project or a Custom GPT?

They solve different problems, and OpenAI's own comparison in its Projects FAQ is the cleanest way to tell them apart. A Custom GPT is essentially single-player: static content, curated once, meant to scale one piece of expertise across an organization. A shared Project is the opposite shape: multi-player, evolving as people contribute chats and files, meant for a team working on the same thing over time rather than one polished assistant everyone queries the same way.

There's a practical wrinkle that decides this for a lot of people before preference even enters into it: new GPT creation is closed on personal accounts. OpenAI's own help center states it directly: "New GPT creation and publishing are not available on personal ChatGPT accounts, including Free, Go, Plus, and Pro." Existing GPTs still work; you just can't build a new one unless you're inside a Business, Enterprise, or Edu workspace. A related fact worth knowing if you do have one: "GPTs do not use saved memory, custom instructions, or previous conversations. Each conversation starts fresh." A GPT never inherits stale context, and it also completely ignores whatever you've built up in your custom instructions.

What Do People Get Wrong About Projects?

Five mistakes show up repeatedly, and every one of them is silent — nothing in the interface flags it at the moment it bites.

Assuming a Project adds to your global setup instead of replacing it inside its own walls. This is the override behavior covered above, and it's worth repeating because it's the mistake with the widest blast radius: someone sets up a Project, and ChatGPT stops obeying tone rules or formatting habits they configured months earlier. Nothing broke. The Project replaced them, exactly as documented. This same override shows up again as one of six context mechanisms in how to give ChatGPT context without repeating yourself, which is worth reading if you want the broader picture beyond Projects alone.

Mixing unrelated work into one Project because being able to create an unlimited number of them doesn't feel like permission to make more. It is exactly that. A Project's instructions and files apply to everything inside it, so a Project doing double duty for two different clients, or a personal errand and a work deliverable, means every chat sees context it doesn't need and the model has to guess which parts apply.

Expecting a Temporary Chat to work inside a Project. It won't. Temporary chats can't be added to Projects at all, which matters if your habit is to reach for a Temporary Chat whenever you want something to not touch memory.

Not noticing the duplicate-filename prompt during a batch upload. If you upload a file that shares a name with one already in the Project, ChatGPT asks whether to upload it anyway or skip it. In a multi-file upload, skipping only drops the conflicting file, not the whole batch, so it's easy to click through without registering that an outdated file is still sitting next to the newer one you meant to replace it with.

Assuming a Project's file understanding matches what you uploaded. Plans outside ChatGPT Enterprise only support text-based retrieval for documents, meaning ChatGPT extracts the selectable text from a file and discards any images. A brand guideline PDF built mostly from color swatches and layout, or a scanned contract with no digital text layer, can arrive as almost nothing useful, with no error to tell you that happened.

What Can't a Project Do for You?

Here's the honest answer, not a sales pitch dressed as one: a Project organizes chats. It does not give you a library of prompts.

Everything a Project holds is a conversation, a file, or an instruction block scoped to one effort inside one ChatGPT account. There's no searchable collection of prompt templates with fillable placeholders that you reuse across different projects, no tagging system, and nothing in a Project travels if you switch to Claude, Gemini, or an API call. If your actual problem is that you keep rewriting the same kind of prompt, a Project doesn't solve that. It just gives the rewriting a nicer room to happen in.

That's a different, narrower job than what a Project is built for, and it's exactly what a dedicated prompt tool does instead: one saved prompt, filled in with variables, inserted with one click, and usable in whatever model you're in that day. If you've outgrown copy-pasting the same instructions into a fresh Project every time you start something similar, that's the actual signal to look at a prompt manager, or at one of the other ways people organize ChatGPT work outside Projects entirely, rather than spinning up a fourth Project. For the full list of storage strategies and where each one, including Projects, falls short, see how to save and organize ChatGPT prompts across devices. And if the answers you're getting inside a well-set-up Project still feel off, the fix is usually in the prompt itself, not the workspace around it — see why your ChatGPT answers are bad and, for a ready-made starting point instead of writing from scratch, 100+ ChatGPT prompt templates.