All pages in this guide
Making Projects
A Project is a permanent workspace where the AI already knows who you are and what you're working on. You explain yourself once, in a short instructions box and a few files, and from then on, every chat inside that workspace starts with the AI fully briefed.
This is the single biggest upgrade you can make. A regular chat forgets everything the moment you close it. A workspace remembers, so you stop re-typing the same background at the start of every conversation.
This page is a little long, on purpose. It's the lesson that pays off the most, and it's the one most people skip. Take your time.
1. The same idea, four names
Every big AI tool has this feature. They give it different names, and the details differ, but the shape is identical: a place to store standing instructions plus a place to store your files, so a fresh chat begins already knowing them.
| The idea | ChatGPT | Claude | Gemini |
|---|---|---|---|
| A saved workspace with standing instructions | A Custom GPT, or a Project for grouping chats and files | A Project | A Gem |
| The box where you say how it should behave | The instructions field when you build the GPT / the Project | Custom instructions | The instructions field when you make the Gem |
| The files it always has on hand | Uploaded knowledge files | Project Knowledge | Attached files / knowledge |
| A stricter cousin that answers only from your sources, with citations | NotebookLM (from Google, works alongside any of them), covered at the bottom of this page | ||
The exact button names move around and the menus get redesigned. Don't memorize the wording. Look for the place that says "instructions" and the place that lets you "add files" or "knowledge." That's the whole feature, in every tool.
A workspace feels like the AI "remembers" you. It doesn't, really. Under the hood it takes your instructions and your files and quietly pastes them in front of every new chat, every time, so you don't have to. It's the same as if you re-typed all that background at the start of each conversation, except it does so reliably and never forgets a piece.
2. What's inside a workspace
Every workspace has two parts you set up once. After that, every new chat you start inside it inherits both automatically.
Instructions box vs. files: what goes where
People get these two boxes confused. The rule is simple:
- The instructions box is for how the AI should behave. Who you are in one line, the tone you want, standing rules ("always work from the files," "never invent a fact," "be brief," "quiz me when I ask"). It's short, a paragraph or two. The AI reads it before every reply.
- The files are for what the AI should know. The actual material: your syllabus, your resume, your past emails, your notes. Facts, not personality. You upload the real thing rather than paraphrasing it into the instructions box, so the AI reads the exact words.
A good test: if it's a fact you'd have to look up, it's a file. If it's a preference about how you want to be talked to, it's an instruction.
One workspace for your course, a separate one for your job hunt, another for the thing you run on the side. Don't pile them together. A workspace holding three unrelated topics gives muddier answers on all three, because the AI can't tell which files are relevant. Making a fresh one takes a minute, so do that instead of mixing.
3. Three workspaces you could build this week
The setup is the same every time: an instructions box and a few files. Only the contents change. Here are three worked examples. Pick whichever is closest to you and copy the shape.
A. The student studying for a course
- What you put in: the syllabus, your lecture notes, and a couple of past problem sets (screenshots or PDFs are fine).
- Instructions box: "You're my study partner for Calculus 201. Work from the files I've uploaded, the syllabus, my notes, the past sets. Explain each step in plain language. When I ask, quiz me on a topic and mark my answers. Never invent a formula; if you're unsure, say so."
- One real question: "Make me a five-question quiz on the section covered in problem set 2, in the same style as the questions there. Wait for each answer before giving the next."
B. The person job hunting
- What you put in: your resume, the exact job posting you're aiming at, and a page of notes about the company (what they do, why you want in, anything you learned from a call).
- Instructions box: "You're helping me prepare for interviews. Use my resume, the job posting, and my company notes. Write in my voice, direct, specific, no buzzwords. Base every answer on things that are actually on my resume; don't invent experience I don't have."
- One real question: "Using my real experience, draft answers to the three most likely interview questions for this role. Keep each under a minute spoken, and point to a specific line on my resume for each."
C. The person running a small thing on the side
- What you put in: your website copy, a handful of past emails or posts that sound like you, and your current price list.
- Instructions box: "You help me write for my small business. Match the voice in my past emails, warm, plain, no corporate words. Use my real prices and services from the files. Never make up a number or a policy; ask me if it's missing."
- One real question: "Draft a short reply to a customer asking whether we can do a rush order, in my usual tone, using my actual pricing."
Notice the pattern in all three: a one-line "who I am and how to behave," a few real files, and a question that leans on those files. That's a workspace. The topic is the only thing that changes.
It's maybe an hour, once, for something you'll use daily. And you don't have to finish it in one sitting. Start with one or two files (the syllabus, the resume, the price list) and add more whenever you notice the AI missing something it should have known. The workspace gets better every time you feed it another piece.
4. The limits, so you're not surprised
A workspace is reliable, but it isn't clever about staying current. Three things to keep in mind:
- There's a size cap. Every tool limits how much you can upload, a generous amount, but a limit. If you try to stuff in every document you own, you'll hit it. Upload short, focused files that matter, not your entire hard drive.
- Stale files go stale silently. If your price list changes and the file still says the old price, the AI will confidently quote the old price. It trusts whatever you gave it. Old files don't warn you they're old.
- It will not notice you updated a document. This is the one that catches people. Editing the original file on your computer does nothing to the workspace, it's still working from the copy you uploaded. When something changes, you have to delete the old file and upload the new version. There's no automatic sync.
None of this is a flaw to fix; it's just how the feature works. Treat the files like a folder you keep tidy: when a fact changes, re-upload.
5. NotebookLM: the stricter cousin
A NotebookLM notebook is similar to a workspace but stricter, and the strictness is the point. It only answers from the sources you've uploaded, and it shows you the exact sentence every answer came from.
Use it when you want the AI to tell you what your documents actually say, rather than brainstorm from them. A PDF you exported, a report, a set of notes, even a long YouTube video: drop them in and ask questions.
- Add sources: Google Docs and Slides from your Drive, PDFs from your computer, URLs, YouTube videos (it reads the transcript), or pasted text. You can add a lot of sources; most people never use more than a handful.
- Ask grounded questions: "Pull every place these notes mention a deadline, and quote each one." or "What themes come up most across these three documents?"
- Every answer comes with citations: little numbered markers you click to jump to the exact sentence in the exact source, so you never have to guess where a claim came from.
There's also a Studio panel that turns your sources into a summary, a study guide, or even a short audio overview you can listen to. We come back to citations, and why they matter so much, on the Checking page.
A workspace is for when you want the AI to make things: draft, rewrite, brainstorm, think a decision through.
A NotebookLM notebook is for when you want the AI to tell you what your documents say, with citations you can check.
They sit happily side by side. Plenty of people keep both for the same topic.
Quick check · 1 of 2
You've set up a workspace with your notes uploaded. Later you fix a mistake in the original document on your computer. What does the workspace now know?
Answer: a workspace works from the copy you uploaded, not the live file on your machine. Editing the original does nothing until you delete the old copy and upload the new one.
This is why stale files quietly stay stale: the AI trusts whatever it was handed.
Quick check · 2 of 2
You're setting up a job-hunt workspace. Your resume and the job posting go in the files. Where does "write in my voice, direct, no buzzwords" belong?
Answer: the instructions box is for how to behave: tone, standing rules, who you are in a line. Files are for what to know, the facts. Voice is a behavior rule, so it goes in instructions.
And you only write it once. Every chat inside the workspace re-sends it for you.