R
RESEARCH HUB
Intelligence Brief
AI Tools Setup Guide Claude

Claude Cowork: The Definitive Setup Guide

Chat is prompt-response. Cowork is task delegation. Here's exactly how to set it up so you walk away and come back to finished work — from folder structure to context files to writing prompts that actually get results.

Claude Cowork launched January 12, 2026. By February 24, Anthropic had already shipped a major enterprise update. In six weeks, it went from research preview to infrastructure.

The core distinction is simple — but most people still don't get it:

"Chat is prompt-response. Cowork is task delegation. You describe an outcome, Claude makes a plan, breaks it into subtasks, executes in a sandboxed VM, and delivers finished files to your folder."

You can step away. Come back to completed work. That's new.

What You Need Before You Start

✅ macOS or Windows x64 (no mobile, no web)
✅ Any paid Claude plan — Pro ($20/mo) minimum
✅ Active internet throughout the session
⚠️ The desktop app must stay open — closing kills the session. Sleep is fine. Quitting is not.

Step 1 — Install & Access

  1. Download the Claude Desktop app from claude.com/download
  2. Sign in with your paid account
  3. At the top of the app, find the mode selector — click the "Cowork" tab to switch from Chat to Tasks mode

That's it. You're in Cowork.

Step 2 — Set Up Your Working Folder

Cowork runs on a folder-permission model. You grant Claude access to a specific directory and it can read, edit, and create files within it. Do not point it at your entire home directory.

~/Claude-Workspace/
├── context/          # Your standing context files
├── projects/         # Active project folders
│   ├── project-a/
│   └── project-b/
└── outputs/          # Where Claude delivers finished work
⚠️ Safety note: Cowork has real read/write/delete access to any folder you share. Anthropic explicitly warns: "Claude can take potentially destructive actions (such as deleting local files) if instructed to." A dedicated workspace limits the blast radius. Back up anything important before your first real task.

If you work across multiple machines, put your workspace in iCloud Drive, Dropbox, or OneDrive. It won't sync sessions — but your files will be consistent.

Step 3 — Write Your Context Files

This is the highest-leverage step. The quality of Cowork's output is directly proportional to the quality of context you provide. Create these as .md files in your context/ folder:

about-me.md

# About Me
## Role & Responsibilities
- [Your name, title, what you do day-to-day]
- [Key stakeholders you work with]
- [What success looks like in your role]

## Domain Context
- [Industry/sector specifics]
- [Key terminology or frameworks you use]

## Example Work
[Paste 1-2 examples of output you're proud of.
This gives Claude a concrete reference for quality and style.]

brand-voice.md

# Communication Style
## Tone
- [e.g., Direct and concise. No filler. Technical when warranted.]
- [Phrases you use naturally]
- [Phrases that sound wrong to you]

## Writing Samples
[Paste 2-3 short examples of your actual writing.
More useful than abstract descriptions of your style.]

## Anti-patterns
- [e.g., Never use "leverage" as a verb]
- [e.g., Don't open with "I hope this email finds you well"]

working-preferences.md

# How I Want Claude to Work
## Output Style
- Save outputs as [.docx / .xlsx / .md — your preference]
- [Short vs. detailed outputs]
- [File naming conventions]

## Guardrails
- [Things Claude should always ask before doing]
- [Things Claude should never do without confirmation]

The compounding effect: These files get better over time. After every session where Claude's output missed the mark, update the relevant context file. Over weeks, you'll have a setup that needs almost no correction.

Step 4 — Write Tasks That Actually Work

Most people write weak prompts and wonder why they get weak results. Here's the formula:

Weak: "Write a report on our Q4 metrics."

Strong: "Read the three CSV files in projects/q4-data/. Produce a 2-page executive summary in outputs/q4-summary.docx covering: revenue vs target, top 3 anomalies, and 3 recommended actions. Format as a Word document with a table for the metrics. Use the tone from brand-voice.md."

The formula: What to read + what to produce + where to save + format + constraints + which context file to reference.

Step 5 — Power Moves

Run parallel sub-agents

Cowork can run multiple sub-agents simultaneously. One researches while another writes while another formats. Describe a complex multi-part task and let it self-organize.

Iterative refinement

Don't redo the whole task. Point Claude to the output and say: "In section 2, make the tone more direct." It edits in place.

Improve your own context files

Ask Claude to help you update your context files based on what it's learned in your sessions together. It'll write better instructions than you will.

How This Compares to What You Already Have

If you're already using an OpenClaw setup or another AI agent system, note the key differences:

Cowork → desktop app, sandboxed VM, best for document-heavy tasks and file-based output
OpenClaw → background daemon, Telegram/WhatsApp interface, full terminal access, works autonomously 24/7

They're complementary. Use Cowork for heavy single-session document work. Use OpenClaw for scheduled autonomous operations and real-time research.

The context file philosophy is identical. about-me.md, brand-voice.md, working-preferences.md in Cowork are exactly what SOUL.md, USER.md, and IDENTITY.md are in OpenClaw. If you've built one, you've already built the mental model for the other.

Source: @witcheer on X · 1.9M views · Expanded and annotated by Research Hub · Mar 2026

Back to Intelligence Brief

Continue Reading

The Fuckening. 6 min → Stability Is the New Status. 6 min →