Trelleborg pilot guide

Claude Cowork — a quick start for your first project.

Cowork is the agentic mode inside Claude Desktop. You describe an outcome, step away, and come back to finished work — written directly to a folder on your computer. This guide gets you from zero to a working personal-assistant workspace in about fifteen minutes. Written for Trelleborg colleagues who do not consider themselves technical.

§1 — Quick start

Four screens, fifteen minutes.

Each frame below corresponds to one screen you will see. Everything happens inside Claude Desktop — there is no separate command-line tool to install. Step-by-step instructions live on the Setup page; this is the bird's-eye view.

Step 1

Open Cowork in Claude Desktop

Screenshot placeholder

Launch the desktop app and click Cowork in the left sidebar. If you do not see it, update to the latest version — Cowork went generally available in April 2026.

Step 2

Connect an empty folder

Screenshot placeholder

Click Use existing folder and pick a fresh, dedicated folder — never your Documents root, your Desktop, or iCloud Drive root.

Step 3

Run the Trelleborg starter skill

Screenshot placeholder

In the Cowork chat box, type /trelleborg-cowork-starter. The skill runs inside Cowork itself — no terminal required. It asks four short questions, then writes your INSTRUCTIONS.md, MEMORY.md, and the supporting folder shape.

Step 4

Paste INSTRUCTIONS, run a task

Screenshot placeholder

Open INSTRUCTIONS.md from the file view, copy its contents, paste them into Cowork's Instructions field for this project. Then describe your first outcome in chat. Done.

Frames marked “Screenshot placeholder” will be replaced with real captures from Claude Desktop. Until then, the labels describe what each screen looks like.
§2 — What you will end up with

The shape of your workspace

The starter skill writes this on your behalf. You do not need to understand every file before your first task — but glance at the shape so it is not a surprise.

my-cowork/
├── INSTRUCTIONS.md         # Pasted into Cowork's field 200–300 lines
├── MEMORY.md               # Running notes Claude maintains 150 lines
├── ARCHIVE.md              # Overflow when MEMORY.md fills
├── context/                # Detail Claude reads on demand
│   ├── about-me.md
│   ├── work-preferences.md
│   └── current-focus.md
├── inbox/                  # Quarantine for anything you did not write
├── projects/               # One subfolder per active project
├── reference/              # Long-form material — read on demand only
├── output/                 # Where Claude writes finished work
├── daily-notes/            # One file per day
└── archive/                # Completed work, moved by you

Why this shape

The two files at the top — INSTRUCTIONS.md and MEMORY.md — are where most of the discipline lives. Both have explicit size caps, borrowed from the pattern Jeff Su's team formalised in May 2026 after measuring a ~25% drop in tokens-per-session when their root files were cut to size.

The remainder of the folder is a quarantine-and-output system: inbox/ for anything that arrived from outside, projects/ for active work, output/ for what Claude produces, and archive/ for completed material. Each named so that — three months from now — the folder still makes sense.

§3 — What Cowork is, and is not

Pin the words down

Anthropic's product line uses similar words for very different things. Read this once and the rest of the guide will make sense.

If you mean… It actually is… How it differs from Cowork
claude.ai Projects Cloud-based conversation folders on the claude.ai website Cloud-synced, chat-based, no agentic execution, no local file access
Claude Code Developer command-line tool for agentic coding Lives in a terminal, code-first, no desktop GUI. Cowork is the Claude Desktop equivalent for knowledge workers.
Teams (the plan) A pricing tier A billing level, not a feature. Cowork runs on the Team plan, among others
Workspaces (Console) API-level organisation units in the Anthropic developer console For developers managing API keys and billing — unrelated to desktop Cowork
Skills Packaged capabilities you can install into Claude The Trelleborg starter is a skill. Skills extend Cowork's capabilities; they are not the workspace itself.
Plugins Bundles of skills and connectors for role-specific use Adds capability to Cowork, not the container. Managed under Customize in the desktop app.
Prerequisites. Claude Desktop installed (claude.com/download), a paid plan (Pro, Max, Team, or Enterprise), and a computer that stays awake while a Cowork session runs. The desktop app must remain open for the session to complete.
§4 — Watch first (optional)

Community walkthroughs

Anthropic has not yet published an official personal-assistant Cowork tutorial. The community videos below are the strongest starting points. Tina Huang's May 2026 follow-up is the most accessible introduction; Jeff Su's piece is the closest match to the model this guide adopts.

Community content, not official Anthropic guidance. The folder shape we recommend draws from Jeff Su (May 2026) for the INSTRUCTIONS/MEMORY/ARCHIVE discipline, Ruben Hassid's revised text-file approach, and the quarantine pattern that has stabilised across the community.
§5 — Next steps

What to do, in order

  1. Open the setup walkthrough and follow the five-minute Claude Desktop sequence. You will connect a folder and run the scaffold skill from inside Cowork.
  2. Read the folder structure page so the seven subfolders and three top-level files (INSTRUCTIONS.md, MEMORY.md, ARCHIVE.md) are not a surprise the first time you open them.
  3. Fill in three small files in context/: who you are, how you like to work, what you are focused on this week. Total about half a page of writing.
  4. Run your first task. Describe an outcome — a draft, a summary, a tidy-up of the inbox/ — and let Cowork plan it. Review the plan, confirm, and watch it work.
§6 — Before you connect anything sensitive

Read the safety page once

Cowork can read and modify everything in the folder you connect to it. That makes the choice of which folder you connect a security decision. There is one documented attack class and a short list of rules to follow — none of which require any technical skill.

Safety, in one short page

How the PromptArmor attack worked, the three folders to never connect, and the manual backup that costs ten seconds. Read once before you connect a folder that already contains real files.

Read the safety page