How to Use Claude Projects: The Workflow That Walls Off Your Context

To understand any product, the first thing I do is work out the problem it solves. In my opinion that's the quickest route to actually learning it. And with AI it matters more than ever, because the whole thing moves so fast you can sink weeks into a feature you never needed. Work out the problem first, and you know straight away whether it's even for you.

So let me do that with Claude projects. If you've used Claude at all, you've probably noticed you can create a project, drop some files in, and chat inside it. But seeing the buttons isn't the same as knowing what they're for. So what problem does a project actually solve?

The problem Claude projects solved

Before projects existed, it was just chats..

The trouble with chats is that the context gets mixed up. While you're talking to one chat, the memory is quietly pulling in all the others.

For example. Say you're having a chat about a new business idea. But a few weeks back, you had a completely separate chat about how to decorate your bedroom. When you start asking questions about the business idea, believe it or not, the AI is leaning on what it picked up about your bedroom to answer you.

Now, you won't see the damage most of the time. The AIs are genuinely good at grabbing the right information regardless of all the noise getting injected from old chats. But it does throw it off a bit. And you never quite know when an answer is slightly wrong simply because the context got muddled.

Then there's the other half of it, which is old context. And this is my pet hate.

Just this month I was trying to talk to Claude about this brand, James Tandy. And because of an old brainstorming chat way back in the history, it kept calling the brand "Lucid Momentum." No matter how many times I corrected it, it refused to budge. That old context was sitting there somewhere I couldn't see or reach.

So I made a project for the blog. One clean space where I decide what the AI knows, and everything else stays outside the door.

The four parts of a Claude project

There's not too much to understand here, fortunately. It's one of those things where, once you get used to it, you just use it automatically at the right moments.

There are four parts to a project. I'll walk through each one the way I actually use it, using a real project of mine as the example: a little bodybuilding app I built to track and analyse my own growth.

Files: the stuff that never changes

Files is where you upload the things the AI should always have in front of it. The whole trick is what you choose to put in.

For my bodybuilding app project, I loaded three things. The codebase, the branding docs, and my coding stack. The reason they earned a spot is simple. They're the things that aren't going to change.

Think of Files as the reference shelf that stays within arm's reach. Settled stuff, the things you'd be annoyed to re-explain in every single chat. If something shifts constantly, it doesn't belong here. If it's fixed, in it goes.

Instructions: the part I leave empty

This one might surprise you. I don't write any instructions.

Most guides will tell you to fill this in with a detailed standing brief. And if your project has fixed rules you repeat constantly, fair enough, that's what it's for. But I'd rather write my instructions inside the chat itself, as I go. It keeps everything dynamic. I'm not locked into a brief I wrote weeks ago and half-forgot about. I steer in the moment, every time.

So don't feel you have to fill this in to "do it properly." Empty is a valid setup. Mine's been empty the whole time and the project works fine.

Chats: one per idea, and name it

Inside the project, the chats are where the actual work happens. I treat them as brainstorming sessions.

The rule I stick to is one chat per idea. The moment a new question or idea turns up, I open a fresh chat for it rather than dumping it into whatever I had open. Then I rename that chat to the idea itself.

That second bit matters more than it sounds. A month later, when I want to pick an idea back up, it's sitting right there with a name I recognise, instead of being buried under a wall of "New chat, New chat, New chat." I just open it and carry on where I left off.

Project Memory: the part you have to police

This is the one to keep an eye on.

Project memory is the memory Claude builds up on its own after you've had a few chats. You don't write it. It writes itself, from whatever you've been talking about. Which is handy.. and also exactly why you need to check it.

Every week I read through what it's captured and make sure it's accurate. And I'm glad I do.

On the bodybuilding project, the memory had quietly decided I was building for a UK audience, on the grounds that I'm based in the UK. But I'm actually aiming at both the UK and the US. Left alone, every bit of strategic advice it gave me would have leaned towards the wrong market. It could have sent me down a properly wrong path without me ever realising why.

Fortunately, you can fix it. There's an edit button. I corrected the entry, and the AI was working from the truth again.

That weekly review takes a couple of minutes. It's the cheapest insurance you'll buy against confidently wrong answers.

A project's real job is simple. What's inside it is what the AI knows. Nothing else gets a vote.

How I actually run a project, start to finish

Pulling it together, here's the whole workflow in order. This is the version I'd hand a friend who'd never opened a project before.

That's really all there is to it. You don't need it set up perfectly on day one. Make the project, drop in the few files that matter, and start a chat. The rest you'll pick up automatically, the same way I did. And once your context is clean, you'll notice the answers sharpen up almost straight away.. which, when the whole game is getting real leverage out of this stuff, is rather the point.