Every new chat with Claude used to start the same way: three or four paragraphs re-explaining what Brite Tech Lifestyle actually is, who I write for, and what I never say, before I could ask it to do anything useful. I did that for months before I actually read what Projects does. It took twenty minutes to set up properly, and I've never had to type that preamble again.
This is beginner level: nothing here needs code, just a Claude account and about twenty minutes to build it properly rather than rushing it. If you haven't settled on Claude over ChatGPT for your writing-heavy work yet, my honest side-by-side is in ChatGPT vs Claude: Which to Use for What — come back here once you know Claude is the one doing your drafting.
The problem: reintroducing yourself every single chat
A general-purpose chat is stateless by default. Ask Claude to write a product description on Monday and a customer reply on Tuesday, and Tuesday's chat knows nothing about Monday's — or about your business at all, unless you paste it in again. I was doing that paste-in every time: what the brand is, the tone it never uses, the products that actually exist versus the ones I've discontinued. It worked, technically. It also meant the first six lines of every conversation were housekeeping instead of the actual task.
The tell that finally got me to fix it was a genuinely good draft that quietly invented a product I don't sell. Not a hallucination out of nowhere — Claude was working from nothing but the couple of sentences I'd typed that day, and it filled the gaps the only way it could: plausibly. The fix isn't a cleverer prompt. It's giving the same context a permanent home so nothing has to be filled in.
What a Claude Project actually is
A Project is a persistent workspace inside Claude.ai that bundles two things together: custom instructions — standing context and rules that apply automatically to every chat you start inside it — and project knowledge, a set of files you upload once that Claude can reference without you pasting them in again. Every new chat you open inside the project inherits both, and keeps its own separate history, so you can run ten different conversations in the same project without them bleeding into each other — they just all start from the same briefing.
It's the same underlying idea as a Custom GPT in ChatGPT, which I've written a full playbook on building — standing instructions plus reference files, so the tool stops forgetting who you are. The difference is what each one is for: a Custom GPT is built to be published as a shareable app with its own name and icon that other people can open. A Claude Project is a private workspace for your own use, with nothing to publish and no app-store listing to maintain. If your goal is a tool for yourself rather than something you hand to other people, a Project is the simpler build of the two.
Worth saying up front: if the actual job is research — grounding answers in a specific set of documents so nothing gets invented — my NotebookLM guide is the more purpose-built tool. Projects is for repeated writing and reasoning tasks that all share the same context; NotebookLM is for staying strictly inside a source set. I use both, for different jobs.
Step 1: Create the project and name it precisely
In Claude.ai, open the Projects section in the sidebar and select "Create project." Give it a name specific enough that you'll know exactly what it's for in six months — "Brite Tech Lifestyle — Content" told me more than "Work" ever would once I had three or four projects running side by side. A one-line description in the project settings is worth filling in too; it costs ten seconds and saves you opening the project just to remember what it does.
Step 2: Write the custom instructions — this is the actual work
Open the project's instructions field and write what you'd want a genuinely sharp new hire to know before their first task: what the business does, who it's for, the tone it always uses and the one it never does, and any standing rules that apply to everything ("never invent a statistic," "UK spelling," "no exclamation marks"). Mine runs to about half a page. Resist the urge to write a novel — the goal is the handful of facts and rules that would otherwise need repeating in every single chat, not a full history of the company.
Be concrete rather than aspirational. "Write in a friendly, approachable tone" produces something different every time you ask; "short sentences, contractions, first person, one idea per paragraph" is instructions Claude can actually follow consistently. If you already have brand guidelines written down anywhere, that document is a better source for this field than writing it from memory.
Step 3: Upload the project knowledge
This is where the specifics live that instructions alone can't hold efficiently — pricing, product names, FAQ answers, past examples of output you'd actually publish. I keep four files in mine: a short style guide, a facts sheet covering what's genuinely true about the business right now, a pricing summary, and a folder of five pieces of writing I was happy with, so Claude has real examples of the target rather than just a description of one. A project has a knowledge capacity rather than an unlimited well, which is exactly why curating four focused documents beats dumping in every file you own — a smaller, accurate set outperforms a large, noisy one every time I've tested it against itself.
Step 4: Test it on a real task before you trust it
Don't just ask it something abstract — give it a task you'd genuinely hand off, like drafting a reply to a specific customer question or writing one product description, and check the output against two things: did it use the facts you actually uploaded, and does the voice match the examples rather than reading like generic AI copy. The first test I ran was the exact product-description task that had invented a product before. This time it didn't, because it had something real to check itself against instead of a gap to fill in.
Step 5: Iterate the instructions based on where it's actually wrong
The first version of your instructions will be close, not perfect. When the output gets something wrong, don't just fix that one answer — go back and add the missing fact or example to the instructions or knowledge base, so the same mistake can't recur in the next chat. I added a line explicitly stating which two products I've discontinued after the invented-product incident happened a second time with a different phrasing; it hasn't come back since. Treat the first week as a tuning period, not a one-time setup you never revisit.
Step 6 (optional): Split into more than one project
Once you're using this for genuinely different jobs — content writing, customer replies, internal planning — resist cramming them into one project. A single set of instructions trying to serve three different jobs ends up vague enough to serve none of them well. I run a separate project for customer-facing replies with its own, stricter tone rules than the content one; splitting them took five minutes and immediately made both sharper.
Common mistakes worth avoiding
- Uploading every document you own instead of the ones that matter. A knowledge base stuffed with old drafts and irrelevant files makes it harder, not easier, for Claude to find the right fact — curate it like you'd brief a new hire, not like you're backing up a hard drive.
- Treating the instructions as a one-time task. The first draft is a starting point. The projects that actually save time are the ones you correct for a week, not the ones you write once and abandon.
- Writing vague style rules instead of concrete ones. "Sound professional" produces something different every time; specific, checkable rules produce something consistent.
- Building one giant project for every job. Different tasks want different tone rules. Splitting them is free and takes minutes.
Where this earns its keep: not a cleverer answer to any single question, but the six lines of housekeeping I no longer type before every real one.
This sits alongside the rest of what I actually use to run this business day to day, which I collect on the AI tools hub — including the honest verdict on Custom GPTs and the local-AI setup for the parts I'd rather keep off the cloud entirely. If you split your work across both assistants like I do, ChatGPT has its own version of the same idea, and I run mine the same curated way.
If setting this up yourself isn't how you'd rather spend a Sunday, keeping a business sounding like itself across every AI-drafted email, ad and thumbnail is exactly what my digital studio services exist to build for you.
Want more setups like this one? I send an honest, tested-at-home tech verdict to my newsletter every week — sign up here and I'll send the next one straight to your inbox.
Questions people actually ask
Do I need a paid Claude plan to use Projects?
Yes — Projects is a feature of Claude.ai's paid plans, not the free tier. If you're not sure it's worth paying for before you've tried it, run a few normal chats on the free plan first so you know Claude's actual answers suit your work, then upgrade specifically to set up the Project once you're confident it's a tool you'll use every week, not just this once.
What files can I actually upload to project knowledge?
PDFs, plain text, Word documents, spreadsheets and code files all work, and each project has a knowledge capacity rather than a strict file-count limit — which is exactly why curating matters more than uploading everything. I keep four documents in mine: a style guide, a facts sheet, a pricing sheet and a folder of five examples of writing I'd actually publish.
Is my data private inside a Claude Project?
On an individual account, project knowledge and chats stay visible only to you, the same as any other conversation on your account. On a Team or Enterprise plan a project can be shared deliberately with named colleagues — it's an opt-in setting, not a default, so nothing you upload becomes visible to anyone else by accident.
How is this different from a Custom GPT in ChatGPT?
The underlying idea is the same — standing instructions plus reference files, so you stop repeating yourself — but a Custom GPT is built to be published as a shareable app with its own name and icon, while a Claude Project is a private workspace for your own use. If you actually want to hand a finished assistant to other people, my Custom GPT playbook covers that route; if you just want your own tool to stop forgetting who you are, Projects is the simpler build.
Can I use the same project on my phone?
Yes — Projects sync across the Claude web app, desktop app and mobile app on the same account, so a project you set up on your laptop is already there with the same instructions and knowledge the next time you open the app on your phone. Nothing about it is desktop-only.
