Introduction to Claude Cowork

A walkthrough of the Claude Academy course Introduction to Claude Cowork (Start your journey track, 14 lessons + quiz). Each lesson is summarised with my own tables and examples; take the course for the full lessons. All the courses are listed in the Claude Academy course index.

Note: Detailed content expanded from course outline and public lesson pages; verify against lessons. UI labels (e.g., “Customize”, “Work in a project or folder”) vary by app version.

Prerequisites: A paid Claude plan; the Claude desktop app (Mac/Windows) recommended for local-file work (also available on web and mobile). Helpful: some experience chatting with Claude.

Audience: Knowledge workers who want to go from first launch to confident daily use — delegating real, multi-step work to Claude rather than just chatting.

---

Module 1 — Meet Claude Cowork (4 lessons)

What is Claude Cowork?

• Cowork = Claude working with your files, folders, and apps, producing real outputs. “Claude Code for the rest of your work.”
• Core shift: delegating, not just chatting. Describe an outcome → Claude plans, executes, delivers; you steer.
• Built for multi-step, longer-running work that spans tools and ends in an artifact.
• You stay in control: Claude shows its plan, asks before consequential actions (send/delete/share), and accepts steering anytime.
• Works on your files, in your apps (connectors), in your browser (Claude in Chrome), with your tools (takes action).

Surface	Shape of work	One-liner
Chat	Turn-by-turn, bounded, quick thinking	Think with Claude
Cowork	Sustained session across many tools → finished output	Delegate to Claude
Claude Code	Dev environment in a codebase (terminal, git)	Build software with Claude

Setting up Claude Cowork

• Install desktop app (claude.com/download), sign in (paid plan). Web/mobile tasks reach connectors only, not local folders.
• Pick a folder (the most important per-task setup choice): Claude gets read + write access, reads all files (Word, Excel, PDF, PPT…) and saves outputs back. Scope it to one project.
• Connectors: set up once (Customize area), toggle per task. Common first picks: email/calendar (Outlook/Gmail), messaging (Slack/Teams), cloud storage (SharePoint/OneDrive/Drive/Box), CRM/PM tools (Notion, HubSpot, Salesforce, Asana, Linear). Some cloud connectors are read/search only — check descriptions.
• No connector? Use Claude in Chrome (may be unavailable on some enterprise plans).
• Permission modes:

Mode	Behavior
Ask before acting (default)	Pauses before outward actions — send, post, share
Act without asking	No pause for those; use only for trusted tools/tasks
Both modes	Always asks before permanently deleting a file (can’t skip)

• Try-it: point at a real folder, connect one app, ask for a brief summarizing the folder; check it against your own knowledge.

What Claude Cowork can do for you

Three patterns that signal a good Cowork task:

Pattern	Example
Multi-step	Triage a week of feedback emails into themes with quotes
Lives in real files (outputs a real artifact)	Monthly metrics report with charts from raw spreadsheets
Spans multiple tools	Draft Outlook follow-ups from calendar invite + notes; synthesize Slack launch chatter

• Scheduled tasks: type /schedule (fresh) or — better — run a task once, verify, then /schedule it. Cadence: hourly, daily, weekdays, manual.
  • Connector-only tasks run remotely (computer can be off); tasks needing local files/apps run on your computer (desktop app open; delayed runs catch up and notify you).
  • Ideas: Friday weekly review, first-of-month metrics roll-up, morning meeting briefings.
• Work away from your desk: start from phone/web; runs remotely; reaches connectors not local folders; a task stays where it started. Admin may control availability.

Hand Claude Cowork your first task

• A good Cowork prompt names: the deliverable (format/length), the inputs (folder, channels, date range, app), the nuances only you know.
• Expect clarifying questions — they close context gaps up front (pick an option or answer freely).
• Steer mid-task — interrupt when off-track; don’t wait and regenerate. Stop and restart if badly off.
• Review like a trusted-but-new colleague’s draft: meets objective? facts accurate (ask which docs it used, then check)? anything untraceable/made up?
• Mostly right → ask for edits. Wrong in a load-bearing way → the prompt lacked key context; supply it.

---

Module 2 — Make Claude Cowork yours (4 lessons)

Get better results faster — four building blocks

Block	Scope	Adopt when…
Global instructions	You, every session	You notice corrections you keep repeating
Projects	A stream of work	Recurring workstreams emerge
Skills	A specific process	You explain the same workflow repeatedly
Plugins	A packaged toolkit for a job	Processes are polished enough to share

Standing context: global instructions and projects

• Global instructions (Settings → Global instructions / “Instructions for Claude”): who you are, your role, acronyms/shorthand, preferred output format/length/tone. Applies to every chat and scheduled task.

  I’m a [role] at [company], working on [streams]. Common shorthand: … Most deliverables are [docs/decks]. Lead with the recommendation; keep background to one paragraph.

• Projects = scoped workspace for a stream of work. Contains: Instructions, Scheduled tasks, Context (folders/links), and Memory (Claude builds automatically from project conversations).
• Good project candidates: a client/account, a recurring deliverable, a launch/initiative.
• Create via Projects → New project: from scratch, from an existing folder, or from an existing project.

Skills: teach Claude Cowork your way

• A skill = reusable playbook (folder) Claude loads automatically when a task matches (or by name).

Component	Purpose
Instructions (SKILL.md)	What, when, how — a runbook for a new colleague
Assets	Logos, templates, slide masters, fonts
References	Examples of good output, style guides, clause libraries
Scripts	Code for steps that must run identically every time

• Build with Claude: “I want to build a skill for [process]. Walk me through what you need to know.” Point at real examples/templates.
• Find installed skills in Customize; update by telling Claude the correction. Skills work inside any conversation, including projects.

Plugins: encode your team’s expertise

• Plugin = package of skills + connectors + subagents built around a job. Expertise travels with the install.
• Two shapes: end-to-end process (e.g., monthly close: actuals → variance table → board memo) or bundle of most-used skills (e.g., finance toolkit).
• Anthropic publishes role plugins (finance, legal, sales, marketing, support, PM…): Customize → Plugins → Install → approve connectors.
• Customize an installed plugin (opens a Cowork task), e.g., “Here are our last three red-lined NDAs. Update the /nda-triage skill… to match.”
• Build your own; check your org’s Directory first. /setup-claude interviews you and suggests a plugin.

---

Module 3 — Use Claude wherever you work (2 lessons)

Claude in Chrome

• Browser extension — the bridge for tools without a connector: internal dashboards (Tableau, Looker), vendor portals, SSO’d CRMs, internal web apps, multi-tab research.
• Works with Cowork in one conversation: browser steps → data back → deliverable built in your folder.
• Asks before sensitive actions by default.
• Watch-outs: you must already be signed in (Claude can’t log in for you); Claude sees what you see — narrow access on sensitive sites and review actions.

Claude for Microsoft 365

• Claude as an add-in inside Word, Excel, PowerPoint, Outlook (Outlook is a separate beta).

App	Strengths
Excel	Analysis, formulas, debug #REF!/circular refs, scenarios, cite cells
PowerPoint	Respects slide master/fonts/colors; native editable charts
Word	Draft/revise in place, comments + tracked changes, grounded drafts
Outlook	Triage with context; replies reflecting threads/calendar

• Cross-app flow in one conversation: Outlook → Word → Excel → PowerPoint → Outlook invite.
• Rule: Cowork for many sources → deliverable (or scheduled); M365 add-in for editing/refining the open file. Most real work uses both.

---

Module 4 — Sharing and safety in Claude Cowork (4 lessons)

Best practices for working safely

• Setup: dedicated working folder (not Documents/Desktop/Downloads); back up irreplaceable files outside Claude’s reach; test new workflows on copies.
• Prompts: be explicit with destructive verbs (“remove the section from the draft, but keep the file”); name bounds (“only Q3 contracts”, “draft only — don’t message anyone”); scheduled tasks should draft, not send, until trusted.
• In the moment: read the plan; watch for scope creep/unexpected files or sites; read confirmation prompts — most mistakes come from clicking through.
• Not the right tool for: regulated workflows needing audit trails (Cowork activity isn’t in audit logs/data exports; Compliance API returns transcripts only for Enterprise); anything you wouldn’t let a smart colleague do unsupervised; highly sensitive personal data outside IT-approved boundaries.

Validating skills for plugins

• Evals = try-outs, no code. skill-creator generates 2+ realistic prompts and produces with-skill vs without-skill output pairs.
• For each pair ask: would I use the skill version? If not, what specifically is off? Plain-English feedback → Claude revises the skill.
• Change one thing at a time, re-run. Ship when cases you care about beat baseline meaningfully and known gaps are named.

Share what you build with your team

• Distribute via your org’s private marketplace (admin-managed). Publishing options:

Option	Effect
Available	In the Directory; opt-in install
Installed by default	Pre-installed; users can turn off
Required	Installed and locked on (e.g., compliance)

• Teammates can use/disable (unless required) but not edit; updates come from the maintainer.
• Habits: one owner, evals before every publish, specific names (sales-customer-renewal-prep, not meeting-prep), quarterly review to retire stale items.

Wrap up and next steps

• Throughline: Cowork moves beyond chat to delegating real work. Next moves: set up memory, schedule a task, install a plugin, try Chrome/M365, share something you built.

---

Quick-Review Flashcards

1. Q: Chat vs Cowork vs Code in one phrase each? → A: Think with / delegate to / build software with Claude.
2. Q: Most important per-task setup choice? → A: The folder you point Claude at (read + write boundary).
3. Q: What does Claude always ask about, regardless of mode? → A: Permanently deleting a file.
4. Q: Three patterns of a good Cowork task? → A: Multi-step, lives in real files, spans multiple tools.
5. Q: Can a task started on your phone use local folders? → A: No — only connectors; start from desktop for local files.
6. Q: Three things a good Cowork prompt names? → A: The deliverable, the inputs, the nuances.
7. Q: What does a project contain? → A: Instructions, scheduled tasks, context (folders/links), and auto-built memory.
8. Q: Four file types a skill may include? → A: SKILL.md instructions, assets, references, scripts.
9. Q: Skill vs plugin? → A: A skill is one playbook; a plugin bundles skills + connectors + subagents around a job.
10. Q: When use Claude in Chrome? → A: For web tools with no connector (you must already be logged in).
11. Q: How does skill-creator evaluate a skill? → A: Side-by-side with-skill vs without-skill outputs on realistic prompts, then your feedback.
12. Q: Three marketplace publishing options? → A: Available, installed by default, required.

Practice Ideas

• Create a scoped working folder for a real project and ask Cowork for a folder brief; grade its accuracy.
• List this week’s tasks and tag each as multi-step / file-based / multi-tool.
• Delegate one task with a deliverable-inputs-nuance prompt; practice steering mid-run.
• Write a 5-sentence global instructions block; create one project for a recurring deliverable.
• Run a weekly report once, then /schedule it as draft-only.
• Build a skill with Claude for a process you keep re-explaining; run the eval loop.
• Install a role plugin (or run /setup-claude) and customize it with your team’s examples.
• Use Claude in Chrome on a login-protected dashboard and turn the data into a summary.
• In Excel, ask Claude to write variance commentary, then turn it into slides in PowerPoint.
• Draft a team-plugin proposal: owner, evals, name, review cadence, marketplace contact.