Claude 101 · lesson 4 of 6
Claude Cowork
What is Cowork?
Claude Cowork — the claude.ai surface where Claude works directly on your files, folders, and apps instead of just chatting about them. Where a normal conversation ends with text you copy out, a Cowork session ends with edited documents, organized folders, filled spreadsheets, and completed multi-step tasks.
The mental model shift: chat is ask-and-answer; Cowork is delegate-and-review. You hand Claude a goal plus access to the relevant workspace, and it plans, acts, checks its own work, and reports back.
- Cowork task loop — the repeating cycle a Cowork session runs: understand the goal → plan the steps → act on files/apps → verify results → report or ask for input. Knowing this loop helps you write requests that give Claude everything the loop needs up front.
- Standing context — instructions and materials Claude carries into every task, so you don't repeat yourself: global instructions (preferences that apply across all work, e.g., tone, formats, "always ask before deleting") and projects (scoped workspaces bundling files, instructions, and history for one initiative).
- Skills — reusable, named capability bundles that teach Claude how your team does a specific task (e.g., "quarterly report formatting," "customer email triage"). Invoked automatically when relevant or explicitly by name. Same concept as agent-skills on the developer side.
- Plugins — packaged collections of skills, context, and connections that encode team expertise into a shareable, installable unit. A plugin lets one expert's workflow become everyone's default workflow.
Course modules (Anthropic Academy)
| Module | What it covers |
|---|---|
| Meet Claude Cowork | Setup, what Cowork can and can't do, when to use it vs chat |
| Your first tasks | Choosing well-scoped starter tasks; reviewing and iterating on results |
| Standing context | Global instructions + projects — making Claude "already briefed" |
| Skills | Creating and invoking skills for repeatable tasks |
| Plugins | Bundling skills/context to encode and distribute team expertise |
| Claude in Chrome | Claude acting inside your browser — reading pages, filling forms, multi-tab workflows |
| Claude for Microsoft 365 | Cowork across Outlook, Word, Excel, Teams, SharePoint content |
| Sharing & safety | Best practices for sharing work and validating skills before packaging them into plugins |
Extending Cowork's reach
- Claude in Chrome — browser extension that lets Cowork operate on live web apps: navigating, extracting data, filling forms. Turns "summarize this dashboard" into something Claude can actually go do.
- Claude for Microsoft 365 — integration giving Cowork access to the Microsoft ecosystem (mail, docs, spreadsheets, chat), so tasks can span the tools where enterprise work already lives.
Both integrations follow the same permission-first pattern as MCP connectors in claude-ai-features: you grant access explicitly, per surface.
Web and mobile expansion (July 2026)
Cowork launched desktop-only; on July 7, 2026 Anthropic began rolling it out to claude.ai (web), iOS, and Android in beta — starting with Max plan subscribers, then expanding to other paid plans. Chat and Cowork are being unified into the same space with access to the same files.
- Cross-device session sync — start a task at your desk, monitor progress from your phone, pick up the finished output on any device.
- Background task execution — a session can keep running with no device online at all (e.g., schedule a briefing task overnight; the finished draft is waiting in the morning).
- Human-in-the-loop notification — when Claude hits a decision that needs judgment, it pushes a notification to your phone; work pauses until you answer, then resumes.
- Anthropic doubled Cowork usage limits through Aug 5, 2026 to encourage larger delegations during the beta.
This reinforces the delegate-and-review mental model above — the review step no longer requires being at the same machine that started the task. Anthropic's own usage data (1.2M anonymized sessions, May 2026) shows software development is under 9% of Cowork use; the largest categories are business-process operations, onboarding checklists, report compilation, and spreadsheet work — evidence that CCAO-F's productivity-tool framing (not developer framing) is the right lens for this surface.
Safety and sharing best practices
- Start with reversible tasks — drafts, copies, summaries — before granting write access to originals.
- Validate skills before plugin-izing — run a skill on representative real tasks and review outputs before packaging it into a plugin others will trust implicitly. A flawed skill inside a plugin scales the flaw across the whole team.
- Keep standing context lean — stale global instructions silently distort every task; review them like you'd review code.
- Review before you rely — Cowork accelerates work, but the human remains accountable for what ships (see the Diligence pillar of ai-fluency-4d).
Why it matters for cert prep
Cowork is the flagship example of Claude as a productivity surface rather than an API — the exact scope of the CCAO-F exam (ccao-f-exam). Expect questions distinguishing chat vs Cowork use cases, and on when standing context, skills, or plugins are the right configuration tool.
See also
- claude-ai-features — the broader claude.ai feature set (Projects, Artifacts, connectors)
- agent-skills — the developer-side sibling of Cowork skills
- ai-fluency-4d — the delegation/discernment habits Cowork demands
- ccao-f-exam — the certification that tests this material