Delegating, not just chatting
Cowork is a mode of Claude in the desktop app where Claude works on tasks alongside you: you point it at a folder, connect your apps, describe an outcome, and it plans, executes, uses the tools it needs, and saves a real deliverable back. The mental shift from Chat is delegation — hand over the whole arc, not one question at a time. You can also run Cowork in the cloud (beta) from the web and mobile.
Key concept · Chat vs. Cowork vs. Code
Three shapes of work in the desktop app: Chat — turn-by-turn thinking, drafting, quick exchanges, one tool at a time. Cowork — a workspace where Claude reaches across many tools, sustains multi-step work, and produces finished deliverables. Code — a full development environment inside a codebase, with terminal and git. Cowork works in your environment: on your files, in your connected apps, in your browser (via Claude in Chrome), taking action rather than describing it. You stay in control — Claude shows its plan first and, by default, asks before it sends, shares, or deletes.
The three patterns — and running on a cadence
Not every task fits Cowork. Three patterns cover most of what it's built for, and they often overlap in one task.
| Pattern | What it means |
|---|---|
| Multi-step | Gather, compare, research, draft, format — hand the whole arc off in one prompt. |
| File-based | The inputs and the output are real files in your folder; Claude reads and edits them in place. |
| Multi-tool | The work spans your connected apps (Gmail, Slack, Drive, calendar, CRM); Claude plans across them. |
Key concept · Scheduled tasks and Cowork in the cloud
Scheduled tasks run a task on a cadence — set one up fresh with /schedule, or do a task once and then /schedule to make it recurring. Where it runs matters: a desktop scheduled task runs when your computer is on (and catches up if it was closed); a cloud task runs in the cloud on time, using your connectors rather than local folders. Cowork in the cloud lets you start a task from your phone and pick up the deliverable at your desk — but a cloud session reaches connectors, not local files, and a task runs where you started it.
Hand Claude a task, end to end
A good Cowork prompt is a brief, not a question. It's less like prompting a chatbot and more like briefing a smart colleague about to disappear into a task and come back with something done.
Key concept · Name the deliverable, inputs, and nuance — then steer
A strong delegation does three things: names the deliverable (format and length), names the inputs (which folder, channels, date range, app), and names the nuance only you know (constraints, what to lead with, what to flag). Leave a row empty and Cowork usually asks a clarifying question — options to click, or answer in your own words. As it runs, steer mid-task: watch the plan and progress, and redirect the moment it drifts (don't wait and regenerate). At the end, review with your Discernment on: does it meet the objective, are the facts traceable to sources, does anything sound made up? The finished file — not the chat — is the artifact.
Standing context, skills, and plugins
Cowork doesn't build memory automatically the way Chat does — the context that carries over is mostly what you set up.
Key concept · Two layers of context, then skills, then plugins
Global instructions are a standing brief (who you are, your shorthand, how you like output) applied to every session. Projects scope a stream of work — instructions, scheduled tasks, context folders/links, and memory Claude builds from conversations inside them. Skills are reusable playbooks (a folder with a SKILL.md plus optional assets, references, and scripts) that Claude loads automatically when a task matches — build one by describing a recurring process to Claude. Plugins bundle several skills (plus connectors and subagents) around a job, in two shapes: an end-to-end pipeline or a team's most-used skills. Install Anthropic's from the marketplace, customize them, or run /setup-cowork to find the right one.
Working safely with an autonomous tool
Claude asks before it deletes (and, by default, before it sends or shares) — that's the floor. The rest is what you bring on top.
Key concept · Set up, phrase, and check
Set up so mistakes can't reach what matters: point Claude at a dedicated working folder (not Documents/Downloads), back up anything irreplaceable, and test new workflows on copies. Phrase prompts to leave no room for the wrong action: be specific about destructive verbs ("remove from the draft but keep the file"), and name the bounds ("only Q3 contracts", "draft only — don't send"). In the moment, three checks: read the plan, watch for unexpected patterns (scope creep is a real signal), and read confirmation prompts before approving. And know when Cowork isn't the tool — regulated workflows needing audit trails, anything you wouldn't hand a quick colleague unsupervised, or sensitive data outside IT's approved boundary.
Chapter 13 quiz
Sixteen questions on Cowork's model, patterns, delegation, customization, and safety. Answer first, then expand Show answer.
Part A · Multiple choice
The core mental shift from Chat to Cowork is:
- A Delegating a whole piece of work vs. chatting turn by turn
- B Faster typing
- C Cheaper pricing
- D Nothing changes
Show answer
A. Describe an outcome; Claude plans, executes, and delivers.
Which mode is a full development environment with terminal and git?
- A Code
- B Chat
- C Cowork
- D Enterprise Search
Show answer
A — Code. Chat is turn-by-turn; Cowork delegates document-shaped work.
Which is NOT one of the three patterns that suit Cowork?
- A Single quick factual question
- B Multi-step
- C File-based
- D Multi-tool
Show answer
A. A quick question is Chat's job.
A cloud Cowork session reaches:
- A Your connectors, not your local folders
- B Your local folders only
- C Nothing
- D Other users' files
Show answer
A. For local files, start from the desktop app.
A strong Cowork prompt names:
- A The deliverable, the inputs, and the nuance only you know
- B Just the topic
- C The model version
- D Nothing specific
Show answer
A.
When Claude drifts mid-task, you should:
- A Steer immediately — don't wait for it to finish and regenerate
- B Always start a brand-new session
- C Let it finish wrong
- D Delete the folder
Show answer
A. Cowork is built for low-cost course corrections.
What are the four building blocks a skill can include?
- A Instructions (SKILL.md), assets, references, scripts
- B HTML, CSS, JS, images
- C Name, price, author, date
- D Chat, Cowork, Code, API
Show answer
A.
A plugin is:
- A A packaged set of skills (plus connectors and subagents) built around a job
- B A single prompt
- C A model
- D A folder of images
Show answer
A. Two shapes: end-to-end pipeline, or a team's most-used skills.
The single highest-leverage safety move is:
- A Pointing Claude at a dedicated working folder, not a catch-all
- B Turning off all connectors
- C Never reviewing output
- D Using bypass permissions
Show answer
A. The folder is the boundary for what Claude can touch.
Part B · True or false
Cowork works on the files you already have, editing them in place; Chat produces new documents.
Show answer
True.
Cowork builds memory automatically the way Chat does.
Show answer
False. Carried context is mostly what you set up (global instructions); projects build memory.
Skills load automatically when a task matches; you don't have to invoke them by name.
Show answer
True. Though you can be explicit if you want.
Cowork is a good fit for regulated workflows that need an audit trail.
Show answer
False. Cowork activity isn't captured in audit logs/compliance exports.
Reading a confirmation prompt before approving is part of working safely.
Show answer
True. Most mistakes are clicking through a dialog that wasn't quite the intended action.
Part C · Short answer
Name the two layers of standing context in Cowork.
Show answer
Global instructions (every session) and projects (scoped to a stream of work, with their own memory).
Give the three in-the-moment safety checks.
Show answer
Read the plan; watch for unexpected patterns/scope creep; read confirmation prompts before approving.