If you only chat with AI, context evaporates every session. The real edge is putting conversations inside a project — a box that holds rules, files, and history. Tokens, time, and results accumulate only there.
A few months into using AI tools, most people hit the same wall. You catch yourself explaining the same thing you explained yesterday. Re-uploading yesterday's PDF. Re-typing yesterday's style rules. Hunting for yesterday's output in a scroll of chats. Why? The answer is simple. You've been chatting. You haven't been projecting.
This essay walks through the difference from start to finish. If you're just starting with AI, come along at your own pace. Today's example happens to be Claude's project feature, but the principle applies to ChatGPT, Gemini, or any tool that comes next. The names will change. This spine won't.
Let's begin with the principle itself.
A chat with AI is like water. You pour it, it runs off. Yesterday's water isn't here today. A project is a bowl. With a bowl, the water stays. Next time you sit down, you start on top of what's already there.
Humans have been doing this for a long time, by the way. Offices have drawers. Each drawer has a label. Tax drawer, contract drawer, receipts drawer. Someone asks where last November's contract is, and you just open the right drawer. Now picture an office with no drawers — paper stacked on the floor, everything flipped through every time. That's what "chat only" with AI actually looks like.
Somehow this common sense disappears in front of AI. Why? Because the interface looks like one window. A blinking cursor, an Enter key. No drawer in sight. But the drawer is there. It's just hidden behind a menu called "Projects."
A project, concretely, is a box that holds three things.
First, instructions. Rules like "use polite register, no slang, end with a question." You write them once. They apply automatically to every chat inside the project. No more pasting the same preamble.
Second, knowledge. PDFs, docs, work-in-progress manuscripts. Upload once, and the AI keeps reading them across every conversation in that project.
Third, a bundle of chats. Every conversation you've had on this topic lives in one folder. A question you asked three weeks ago is findable again.
The idea comes from tools like Notion. Hundreds of notes, grouped by project. That mental model just walked over to AI. "Group your chats into folders." That's the whole thing.
Here's a simpler picture. The first time you order at a café, you explain everything: latte, oat milk, less sweet, to go. You do that every visit. But once you become a regular, you walk in and hear "the usual?" Because the barista has your personal instructions already loaded.
A project is that "regular" state for AI. Chatting only is walking into a new café every morning and starting from scratch. Three months in, the AI still treats you like a first-timer. Wasteful.
Numbers time. Imagine you paste 500 characters of instructions and two small files at the top of every chat. That adds about 3,000 input tokens per conversation. Ten conversations a day means 30,000 duplicate tokens a day. A month gives you about 900,000 tokens of pure repetition.
| Mode | Upload frequency | Monthly duplicate tokens | Search time |
|---|---|---|---|
| Chat only | Every single chat | About 900,000 | Scroll each time |
| Project | Once, up front | Near zero | Straight from folder |
Here's the first aha.
Chat only turns you into someone who keeps re-explaining themselves. Projecting turns you into someone who designs once and harvests forever.
This isn't about saving a few cents on tokens. It's about where your time and attention flow. If you spend your mornings restating context, the hard work never gets done.
So when should a chat become a project? One question.
"Am I going to do this again within a week?"
Yes? Make it a project. No? A plain chat is fine.
"Plan our one-time family trip next weekend" — one-shot. A chat is enough. "Write in the tone of my blog" — every week forever. Upload tone instructions and a few past posts as a project. Three months from now, the voice will still match.
Concretely, here's how to build a blog project.
Step 1. Open Claude.ai, left menu, Projects, New Project. Call it "My Blog." Thirty seconds.
Step 2. Under Custom Instructions, type your rules. Example: "No slang. Three to four sentences per paragraph. No emoji. End with a question." Write this once.
Step 3. Under Project Knowledge, upload five past blog posts as PDFs. The AI now references your voice every reply.
Step 4. From now on, only chat inside that project. Type "give me three topic ideas for this week" and the AI already knows who you are, how you write, and what you've written. No preamble.
Try this for a week. The lightness of not typing "you know my style, right?" is the real payoff.
One warning. Don't split projects too small. "Monday blog," "Tuesday blog" — no. One "Blog" project is enough. Start a new project only when the domain truly differs. Blog, taxes, family calendar — three is plenty for most people.
Also, some AI tools don't let you cleanly delete projects, only archive. Name them carefully the first time. Naming is cheap; cleanup later is not.
To close.
Chat flows away. Projects hold. Which one you choose is what divides people who run AI from people AI drags around. Today the example was Claude's Projects, but the labels will keep changing. ChatGPT calls it GPTs. Gemini calls it Gems. Next year's tool will call it something else. The principle — "build a vessel for your conversations" — transfers cleanly to every one of them.
Carry one question in your pocket: "Am I going to do this again within a week?" That one question automates the moment you promote a chat to a project.
The people who use AI well aren't the ones who talk best. They're the ones who build good drawers. That instinct doesn't come from AI. It comes from the much older craft of organizing work. Three years from now, when these tool names have all rotated, this principle still works.
Chats ask. Projects remember. Management wins.