How to Use ChatGPT Projects: One Project per Client, Set Up Right

If you do client work with ChatGPT, you probably start too many chats with the same paragraph: who the client is, what tone they like, what was agreed last week. Knowing how to use ChatGPT Projects fixes most of that. A Project keeps the background, the files and the rules for one body of work in one place, so every new chat starts already briefed.

This guide covers how we set them up: one project per client or workstream, instructions that hold up, a small set of files, and a habit that keeps the project current. At the end we cover when a Project is the wrong choice, which matters more now that OpenAI is retiring custom GPTs.

What a ChatGPT Project holds

A Project is a workspace with its own chats, files and instructions. Every chat you open inside it starts with that context. Project instructions take priority over your global custom instructions while you work inside the project, so a client's style rules won't clash with your personal defaults.

As of October 2026, according to OpenAI's Help Center:

• Projects are available on every plan, including Free, and you can create as many as you want.
• File limits are per project: 5 files on Free, 25 on Go and Plus, 40 on Pro, Business, Enterprise and Edu.
• You choose between default memory, which can use your saved memories from across your account, and project-only memory, which keeps the project separate from the rest of your account. Set this when you create the project.
• Projects can be shared. Shared projects always use project-only memory, and everyone in the project can see its chats (release notes).

One project per client or workstream

The rule we use: a project is a body of work with its own context that you will come back to for weeks. Good examples are a client account, a recurring monthly report, a product launch, or a hiring round for one role.

Two setups usually go wrong. A project called "Marketing" is too broad: three clients' tone rules end up in one set of instructions and the answers get muddled. A project for a single email is too small: setting it up takes longer than writing the email.

Setup takes about 20 minutes:

1. Create the project and name it the way you name the client folder on your drive.
2. Choose memory. For confidential client work, or anything you plan to share, pick project-only.
3. Write the instructions (template below).
4. Upload three to six core files, not everything you have.
5. Drag your existing chats about this client into the project.
6. Open a kickoff chat and check that ChatGPT actually understood the context.

Project instructions you can copy

Instructions are the part people skip, and it's why their projects feel no different from a regular chat. Put rules and context here and keep reference material in files. Replace the bracketed parts:

You are working with me on the [client name] account. I am a [your role].

Context:
- Client: [one line on what they do and who their customers are]
- My scope: [what I deliver for them, e.g. monthly content plan and reporting]
- Main contact: [role, not name], prefers [short emails / detailed reports]
- Current priority: [the goal for this quarter]

How to work:
- Use the files in this project as the source of truth. If the files and my message conflict, ask which is current.
- When you state a fact about the client, name the file it comes from.
- If information is missing, ask one question instead of guessing.
- Write in [US English], plain style, no hype words. Match the tone in brand-voice.pdf.
- Default output: a draft I can edit, then a short list of what you assumed.

Never:
- Invent numbers, dates, or quotes.
- Promise anything on my behalf that is not in the scope document.

The "name the file" and "ask one question" lines matter most. Without them, a project answers confidently from half-read files. For more patterns, see our custom instructions examples for work and the four elements of a work prompt.

Which files to upload

Files are what make a project useful, and the per-project limit is the main constraint. A short set of current files works better than a full archive:

File	Why it belongs	Tip
Scope or statement of work	Keeps drafts inside what you actually sell	Remove fees and bank details first
Brand voice or style guide	Tone without repeating yourself	Add two short examples of approved copy
Latest report or brief	Current numbers and goals	Replace it monthly, delete the old one
Project log (your own doc)	What was decided and when	Update it at the end of each session
Glossary of client terms	Product names, banned words	Half a page is enough

What to leave out: contracts with personal data, anything under NDA that your agreement doesn't allow you to process with AI tools, and raw exports with customer records. Our guide on what not to upload to AI goes through these in detail. Project-only memory controls which context ChatGPT uses. It is not a privacy setting, and OpenAI's usual data policies still apply.

A worked example: the weekly client update

Take a typical consultant task: every Friday, a status email to the client covering what was done, what is blocked and what comes next. The notes are scattered across chats and a task list.

Inside the client project, the prompt is short because the project already holds the context:

Draft this week's status update for the client.

My notes from this week:
[paste bullet notes, task list export, or meeting notes]

Format: subject line, then three sections: Done, Blocked (with what we need from them), Next week.
Under 200 words. Match the tone in brand-voice.pdf.
After the draft, list anything in my notes that contradicts the scope document.

The first draft is usually close on structure and tone, since the voice file and instructions are already doing their job. Two problems are common. The "Blocked" section turns vague ("awaiting feedback"), and the draft repeats an item from last week as if it were new.

The fix is one follow-up, not a rewrite:

Rewrite the Blocked section: for each item, name the exact thing we need from the client and by what date. Then compare this draft to the latest entry in project-log.docx and remove anything already reported.

That second check only works if the project log is up to date, which brings us to the habit that keeps projects useful.

Keep the project current with a handoff note

Projects decay. The brief changes, a decision gets made in a call, and the files still describe last month. At the end of any session that changed something, run this:

Write a handoff note for this session, max 120 words:
- Decisions made (with date)
- Open questions and who owns them
- Anything in the project files that is now outdated
Use plain bullets. Do not include anything I did not confirm in this chat.

Paste the note into your project log and re-upload it once a week. If ChatGPT can draw on earlier chats in the same project, that helps, but we don't rely on it: a file you control is easier to check than memory you can't see.

Project, custom GPT, or custom instructions?

This is where most guides are out of date. On September 11, 2026, OpenAI announced it will retire custom GPTs and move them to plugins. Custom GPTs are scheduled to stop running on December 11, 2026, and as of October 2026 new GPTs can no longer be created on Free, Go, Plus and Pro (ADTmag). Check the notice in your own account or workspace, since timelines vary by plan.

This is how we decide today:

You need	Use	Why
Your defaults everywhere (role, tone, format)	Custom instructions	Apply to every chat outside projects
Ongoing work with its own context and files	Project	Chats, files and rules stay together and change over time
The same fixed task repeated by you or a team	Plugin with a skill (the GPT replacement)	Stable steps and format, shareable
A colleague working on the same client	Shared project	Same files and instructions for everyone

A quick test: if you will come back tomorrow and continue the same work, use a project. If the task is identical every time and the context barely changes, such as turning call notes into a CRM entry, build a reusable assistant instead. We cover that path in how to create a custom GPT for work now that GPTs are retiring. Claude has its own Projects feature that works in a similar way, and we cover it in a separate guide.

Where it breaks

• Retrieval misses. ChatGPT doesn't reread every file in full for each answer. It pulls the parts that look relevant. Ask it to quote the section it used, and spot-check anything that goes to a client. Our hallucination checklist helps here.
• Long chats drift. After many turns, instructions carry less weight. Start a new chat inside the project. The context carries over, the clutter doesn't.
• Stale files win. If last quarter's brief is still uploaded, it will get quoted. Delete old files when you replace them.
• Numbers. Totals and percentages from uploaded reports still need checking against the source.
• Shared means visible. In a shared project, collaborators can see the chats. Don't use a client project for side notes about that client.
• File limits on Free. Five files per project fill up fast. Combine the small documents into one before you upload.

Set it up yourself or hand it off

For one person with a handful of clients, the setup above is an afternoon of work. If you need projects, shared assistants and automations rolled out across a team with access rules, our AI implementation and automation service handles that end to end. If you'd rather build a reusable assistant for your recurring tasks yourself, the Build Your Own AI Assistant mini-course walks through it step by step.

Go further in the practicum

Projects are one piece of a working setup. Session 4 of our AI system in two weeks practicum covers persistent settings, projects and workspaces, and building your own assistant for a recurring task. You set them up on your own work during the session, then audit your week and map your personal AI system. If you want the bigger picture after that, our 90-day AI adoption plan shows how to roll it out over the following months.