How to Use Claude Projects (and Where They Fall Short)

You set up the project on a Tuesday. Uploaded the brand guide, last quarter's decks, the messaging doc, a few transcripts. Wrote a line of instructions. And for about three weeks it was the best AI experience you'd had — Claude knew the accounts, used the right words, stopped asking you to re-explain the business every morning.

Then you started a second project for a different part of your work. And a third.

And somewhere around the fifth, you noticed you were uploading the same four files again.

How to use Claude Projects

A Claude Project is a self-contained workspace with its own chat history and its own knowledge base. You upload documents, code, and reference material, write project instructions that apply to every conversation inside it, and Claude draws on all of it without being reminded. Free accounts get five projects. Paid plans get as many as you want, plus automatic retrieval that expands what a knowledge base can hold by up to ten times once it approaches the context limit.

The setup that works is the one almost nobody does, and it takes about twenty minutes.

Write the instructions before you upload a single file. Project instructions are standing direction — who you are, what this project is for, how you want Claude to respond, what to avoid. Most people skip straight to uploading, then wonder why the output still sounds like a press release. Files are reference. Instructions are intent. Intent first.

Scope the knowledge base to the job. There's a temptation, once you learn about retrieval, to put everything in. Resist it. A project that holds four sharp documents answers better than one holding forty vague ones, because relevance is what retrieval sorts on, and noise makes relevance harder to find.

Name your files like you'd describe them out loud. Claude searches the knowledge base by meaning. strategy-doc-final-FINAL-v3.pdf gives it nothing to grab. Q3 pricing objections from sales calls tells it exactly when to reach for that file and when to leave it alone.

Then give it the documents that make it sound like you. Your personal constitution, your writing codex, your story bank. Project knowledge is usually where people put company material, and company material is exactly why so much AI output reads like it was written by a department. The material that makes writing feel human is the material about the human.

Do those four things and a Claude Project is genuinely excellent. I use them. This isn't a takedown.

Where they fall short

Here's the part you don't hit on day one. You hit it around project five.

The context is scoped to the project. That's not a bug. It's the entire design, and it's the reason projects work as well as they do. But it means the knowledge lives inside four walls. Start a new project and Claude arrives knowing nothing again. Your constitution, your codex, your voice: uploaded once per project, forever.

Change your job title and you don't update it once. You update it in every project that mentions it, assuming you remember which ones do.

And it doesn't leave Claude. This is the one that costs the most and shows up the latest. Project knowledge is a claude.ai feature. It doesn't travel to ChatGPT. It doesn't travel to Gemini. It isn't available to anything you build on an API, or to the AI features quietly appearing inside the tools you already pay for. The work you did teaching one platform who you are does not transfer, and you find that out on the day you want to switch.

The same trap sits inside every tool's memory system, which is why we compared them side by side rather than picking a winner. The pattern holds across all of them: the better the feature, the more expensive it is to leave.

Nothing tells you when the knowledge has gone stale. A project you built in March will answer confidently in September using March's numbers, March's org chart, and a positioning statement you abandoned in June. It won't hedge. It has no way to know. That's the quiet version of context drift: not the model wandering, but the material underneath it aging while you keep trusting it. The output looks identical on the day it goes wrong, which is what makes it expensive.

And a blank knowledge base is harder than it looks. "Upload relevant documents" sounds like an instruction. It isn't. It's a decision about what actually matters, and it's the step where most people stall, upload whatever's on the desktop, and get a project that knows their file names instead of their business. Knowing what belongs in there is its own skill, context mining, and it's the one part of this the interface can't do for you. An upload button is not a prompt for self-knowledge.

The layer underneath

So the fix isn't to abandon projects. It's to stop treating one as the place your context lives.

Write your context once, as plain markdown files you own — a constitution, a codex, a story bank, a current state of the union. Six documents, roughly. Keep them somewhere that isn't a product. Then paste them into whatever needs them: this Claude project, the next one, ChatGPT's custom instructions, a Gemini gem, the API app your team is building.

The project becomes a deployment target instead of a vault. Nothing about the setup changes. What changes is where the original lives, and whether losing access to one product costs you an afternoon or a year.

The difference shows up the first time something changes. Update the source file, re-upload where it matters, and you're current everywhere. The alternative is a slow archaeology through eight projects trying to remember which ones knew you used to run marketing.

We wrote up every method of giving AI context, compared, because this question comes up constantly and the honest answer is that projects are one of the good ones. Just not the durable one.

That's the whole distinction, and it's worth being precise about: a context anchor is a file you own. A project is a place you put files. One survives the tool. The other is the tool.

So set the project up properly. Write the instructions first, keep the knowledge sharp, name things like a human, and put your anchors in there.

Just write the anchors somewhere you can get them back. Because there will be a fifth project. And a sixth. And on the Tuesday you find yourself uploading those same four files again, you'll want them to be four files you can find — not four files you have to reconstruct from memory, one platform at a time.