Part 1 — Setting Up Your Workspace

A chat tab forgets. A Cowork workspace remembers, because its memory is a folder of files on disk — an Instructions file at the root, called CLAUDE.md, and a handful of files alongside it that you named, placed, and approved one at a time. Cowork re-reads CLAUDE.md at the start of every turn, and the Instructions tell it which of the other files to pull in. Part 1 is the hiring conversation: you decide what this workspace is for, what it should know, and what each piece (other than CLAUDE.md, which is Cowork's fixed convention) should be called. By the end you have a workspace on disk you can hand work to.

What that buys you:

• A workspace tuned to your function — every name, every path, every line traces back to a choice you made.
• When you come back tomorrow, the workspace behaves the same way it did today, because the behaviour lives in files you can read in a text editor, not in a setting buried in a chat.

What you're actually building

The Instructions are the briefing Cowork re-reads at the start of every turn. Everything else lives in its own file alongside them — the way you want it to sound, the rules it should follow, the kinds of work it does over and over — with one line in the Instructions telling Cowork to read each file when the topic comes up.

That shape has a name: a branching tree. The Instructions are the trunk. Each pointer is a branch the trunk explicitly directs Cowork to follow.

A short trunk, a few branches, each one pointing at a named file:

flowchart LR
  I["Instructions<br/>(trunk)"]:::trunk
  I -- "ALWAYS read voice.md<br/>for how it should sound" --> V["voice.md"]:::branch
  I -- "ALWAYS read audience.md<br/>for who's on the other side" --> A["audience.md"]:::branch
  I -- "ALWAYS read ready-check.md<br/>for what counts as done" --> R["ready-check.md"]:::branch
  classDef trunk fill:#14b8a6,stroke:#14b8a6,color:#141414
  classDef branch fill:none,stroke:#ca8a04,color:#f4f4f5

When you want to change how the workspace handles a particular kind of work, you open the file at the end of that branch and edit it. The trunk does not move.

Every line in the Instructions follows one template: ALWAYS read [some file you named] for [the kind of work that file covers]. Call this the ALWAYS-read template.

How to run it

There are two real moves in Part 1.

1. Create a Cowork Project pointed at the folder you want this workspace to use. Top-left, Projects → New Project, point at an existing folder or make a new one, and turn project memory off.
2. Open the prompts panel and paste the first prompt. The prompts panel is the companion bar on this article — the right sidebar on desktop, the floating button on mobile. The model takes it from there.

Project memory off matters for a specific reason.

Files at the root are different. They travel with the folder: anyone you share the folder with sees them, and you can open them yourself.

What happens when you paste the prompt

The walkthrough has a shape worth previewing so you can follow along while you run it.

It starts by reading what's already in your folder so neither of you is guessing. If there's already a CLAUDE.md at the root, the prompt asks whether to extend it or start over.

Then it asks you, in one sentence, what this workspace is for. That single answer seeds everything that follows.

If your sentence is short or generic, the prompt asks a few short multiple-choice questions to thicken it. Who's on the other side of this work. What ends up in someone's hands when you're done. What the voice should sound like. You're picking from short option lists; you don't have to know the right answer in advance.

From your answer, the prompt comes back with a plain-language list of the things this workspace is going to want to remember. Three to seven items, each one a sentence. No filenames, no folder names. Items read like "the kind of thing that explains how this should sound" or "the kind of thing that says when something is ready to send."

You walk through the list with the prompt: drop one, rename one, add something it missed, or say it looks right. The list does not move forward until you say it does.

Once the list is right, the prompt proposes a layout — a few branches sitting alongside the Instructions at the root. You can ask for it flatter, more nested, or a different shape entirely. When the shape looks right, you name each branch. Every name is yours.

Then it goes one item at a time. For each thing on the list, the prompt asks where it should live and what the file should be called. You name it. The prompt drafts the contents from your one-sentence answer plus the item's purpose, you read the draft, and you either accept it, edit it, or ask for a different angle. Nothing lands on disk until you say go.

When all the files are written, the prompt writes CLAUDE.md at the root. Your one-sentence answer at the top, then one line per file you wrote, all of them in the ALWAYS read X for Y shape. That's the whole file. No template sections, no boilerplate. Just the lines you named. (CLAUDE.md is the only filename you don't choose — it's Cowork's fixed convention for the Instructions file. Everything else is yours.)

Last, it shows you the resulting layout on disk and reads CLAUDE.md back to you so you can see what you have. If anything's wrong, you can adjust placement, rename a file, edit CLAUDE.md directly, or restart from a different one-sentence answer.

By the end of Part 1 CLAUDE.md at the root of your folder reads roughly like this:

This workspace handles inbound product feedback for the support team.
Voice is plain, calm, and concrete. Replies stay short.

ALWAYS read voice.md for how this workspace should sound.
ALWAYS read audience.md for who's on the other side of the work.
ALWAYS read ready-check.md for what counts as done before sending.

The first lines are the one-sentence purpose you gave (cleaned up). The ALWAYS read lines are the pointers — one per file you named in the walkthrough. Your filenames will be different. Your sentences will be different. The shape is the same.

When you're done

You have CLAUDE.md at the root of your folder and the files it points to sitting wherever you put them. That's the whole maintenance story: open the files, read them, edit them, move them and update the matching line in CLAUDE.md.

Each later part adds one or two more ALWAYS read lines to the same CLAUDE.md. The trunk stays short. New branches go in beside the old ones. Nothing gets restructured.

Open the prompts panel and paste the first prompt to start.

Next: Part 2 — The Playbook extends the same workspace with the standards your work runs through.