The Two Files That Turn Claude Into a Coworker

Most people use AI like a vending machine.

You walk up, punch in a question, take the answer, walk away. Come back tomorrow and the machine has no idea who you are. Same stranger, every time.

I did exactly this for over a year. Every session I’d re-explain who I am, what I do, who’s on my team, how I like things written. Then I’d close the tab and all of that context just… evaporated. Next morning, type it all again.

The thing that finally fixed it wasn’t a smarter model. It was two text files.

Here’s the whole setup in plain English. No code. If you can write a grocery list, you can do this.

Two plain text files: one for your standing instructions, one for what it learns and keeps. That’s the whole gap between a chatbot and a coworker.

Start with the desktop app, not the chat window

The browser chat is fine for asking questions. But there’s a different mode that changes the whole thing.

In the Claude desktop app there’s a feature called Cowork (Claude’s agent that can actually do work on your computer, not just talk about it — open files, write files, take real actions). That’s the part that stops feeling like a chatbot and starts feeling like a teammate sitting at your machine.

First move: make a folder wherever you already keep your stuff — Documents, Drive, Dropbox, doesn’t matter. Call it something like Cowork OS. That’s home base. Everything you do with Claude lives here.

Now the two files.

File one: CLAUDE.md — the standing instructions

Make a text file called CLAUDE.md. The .md just means markdown, which is a fancy word for the simplest kind of text file there is.

This is your global briefing. Five things:

• Who you are — your name, your role
• What you do — the business, in a sentence or two
• Your team or family — the names that come up
• The tools you use — your CRM, your email, your calendar
• Your tone of voice — how you want it to write

Claude reads this at the start of every single session. Think of it as the one-page rundown you’d hand a new hire on day one so you never have to repeat yourself again.

You don’t even have to write it yourself. Tell Claude “set up my CLAUDE.md, ask me whatever you need to know,” answer the questions, done. It evolves over time. Keep the first version short.

Then it stacks. Each project gets its own subfolder — personal finances, a website, a specific client — and each subfolder gets its own CLAUDE.md. The root file is the company handbook everybody follows. The project file is the specific quirks of that one job. Claude reads both. The rules stack.

File two: the memory file — what it learns and keeps

Second file: a memory file. Call it memory.md.

CLAUDE.md is the stuff you tell it up front. The memory file is the stuff it figures out and holds onto. It grows as you work. You can also just say it out loud — “remember that I never want beige on the walls” — and it writes that fact down so you never have to say it twice.

That’s the difference between a chatbot and a colleague. A chatbot forgets you the second you close the tab. A colleague takes notes and actually remembers them next week.

One model tip that’ll save you a headache

Quick aside, because this trips people up: don’t reach for the biggest, most powerful model for everything.

Match the model to the job. The heavyweight model is built for genuinely complex work — three, four, five steps that all depend on each other. For most day-to-day stuff, a lighter model is faster, plenty smart, and won’t burn through your usage by lunch. Save the big one for when the task actually earns it.

Then give it hands

Last step, and this is the one that flips it from interesting to useful.

Go into the settings and connect your apps — Gmail, Google Drive, Google Calendar, whatever’s in the connector list. Now it’s not just talking about your work from the outside. It can read your calendar, pull a file, draft the email, look at the actual thing instead of guessing.

That’s the jump. From “an AI that knows about business in general” to “an AI that knows about yours.”

Why this is the actual unlock

Here’s the part worth sitting with.

The model is the same for everyone. Anthropic ships the same Claude to me as it does to you and to the smartest person you know. Some people prompt a little better than others, sure. But the raw intelligence is a commodity now. Everybody’s drinking from the same well.

The edge isn’t the model. The edge is context — your context, written down where the AI can read it.

Without it, you’re talking to the internet’s AI. A brilliant stranger who knows the whole world and nothing about you, starting from zero every single conversation. With those two files, you’re talking to your AI. It already knows the players, the tools, the way you like things. It doesn’t need to be re-briefed.

Two text files. That’s the whole gap between a vending machine and a coworker.

I wrote a longer piece a while back on why this matters so much — the real edge is your data, organized. This is the smallest, most practical version of that idea. Start with the two files. Build from there.

The model isn’t your coworker. The context is. The model is just the part that reads it.

If you’ve set this up your own way — different folders, a memory file that’s gotten huge, a CLAUDE.md you’re weirdly proud of — I’d genuinely like to see it.