You do not need any of this to start. Bookmark it for the second week. The reading list below is the short list — the writers and posts that materially shaped this guide. The glossary disambiguates the words around Cowork that get mixed up in conversation. And the feedback channel is where you tell us what is missing.
Most Cowork material on the open web is shallow — “ten things to try” lists and screenshot tours. These six are the ones worth your attention.
Jeff Su's May 2026 piece — the source of the 300-line
INSTRUCTIONS cap and the 150-line
MEMORY.md + ARCHIVE.md pattern
this guide adopts. He measured ~25% fewer tokens per
session after cutting his root file to size. Read this if
you want the rationale, not just the rules.
Ruben Hassid's May 2026 evolution of his earlier Cowork writing. Argues that a single compressed self-profile — voice-dictated to keep it natural — outperforms a tidy multi-file rig. We disagree on the multi-file part but the profile-writing advice is the strongest published on this.
Simon Willison's running thread covering the class of attack Cowork is exposed to. The three-leg framing on the Safety page is his. Continues to cover new incidents — including the May 2026 Microsoft Copilot Cowork disclosure that echoed the Claude one.
PromptArmor's January 2026 disclosure. The exact mechanism — hidden 1pt white-on-white text in a .docx, the Anthropic API domain on the outbound allowlist, the attacker's own account as the drop point. Their companion piece “Implement Claude Cowork securely” gives positive guidance.
Anthropic's own engineering write-up of the VM architecture that hosts Cowork's code execution. Acknowledges the exfiltration-through-allowed-domain class of risk and describes the design trade-offs they made. Technical but clearly written.
Anthropic's own setup article. The single source of truth for UI labels and current product behaviour. Every claim in this guide that describes how Cowork actually works traces back to this page or its sibling articles.
When pilot participants disagree about something, it is usually because two of the words below are being used for the same thing, or one of them is being used for two different things. Pin these down once.
INSTRUCTIONS.md contents. Claude reads it at the
start of every session in that project. The full template
lives on the Folder page.
ARCHIVE.md.
MEMORY.md. Claude does not
read it unless you ask. Lets you keep the running whiteboard
small without losing the history.
context/ subfolder inside your Cowork
project. Contains three short files Claude reads on demand
(not every session). The routing map in
INSTRUCTIONS.md points to them.
MEMORY.md. Stored locally on your machine, not
in Anthropic's cloud. View, edit, or reset from
Project settings > Memory.
/schedule command.
Runs at a fixed time, writes results back to your folder.
Requires Claude Desktop to be open. Skips and retries when
the machine wakes.
output/.
trelleborg-cowork-starter skill that scaffolds
your folder. Skills run inside Cowork and extend what it can
do. They are not the workspace itself. Managed under
Customize.
This guide is a starting point. Every Trelleborg pilot participant is expected to share what worked, what did not, and what was confusingly worded. If a section on this site needed three reads before it made sense, that is a bug — let us know which one.
The Trelleborg-internal channel for the Claude pilot. James confirms the exact channel name before the site goes live.
channel name pending — ask JamesWhat to share. A screenshot of friction, the literal sentence that confused you, or the workflow you wish was covered. Specific is better than detailed.