Skip to content

Claude Cowork

Cert course: Introduction to Claude Cowork.

What it is

Claude Cowork is Anthropic's task-oriented workspace for knowledge work. Instead of pasting text into chat and getting text back, you delegate an outcome: Cowork proposes a plan, works through multi-step tasks, and delivers a real output file to your drive or connected tools. It's the right tool when the work needs your files, connected systems, or a finished artifact — not just a conversation.

When Cowork vs Chat

Use Cowork when the task:

  • Reads, creates, or edits files in a folder
  • Touches connected tools (Slack, Drive, Gmail, Calendar)
  • Needs multiple steps, is long-running, or runs parallel analysis
  • Should end in a real output file

Use Chat when the task:

  • Fits in one conversation or paste
  • Is a quick answer, explanation, or draft
  • You want to refine turn by turn

How it works

  1. Describe the outcome — name the source material, the transformation, and the output.
  2. Review the plan — Cowork proposes an approach and waits for your approval before acting.
  3. Let it run — watch progress, redirect mid-task, or step away.
  4. Review the result — open the finished file and verify it before sharing or acting on it.

Core concepts

  • Working folder — point Cowork at a folder and it reads/creates/edits files there directly. No upload/download loop.
  • Shared working folder — for multi-person collaboration, point Cowork at a synced, shared Google Drive folder instead of a purely local one; teammates can use the same folder as their own Cowork project, and you can schedule tasks that update it. Chats themselves still aren't shareable — only the folder and its files are.
  • Projects — each task starts fresh, but context carries between tasks through a named project backed by a local folder. Project Instructions are read in every task from that project; project memory is scoped to the project. Global Instructions apply across all tasks.
  • Skills / Plugins / Subagents — Skills are reusable markdown instructions that work across Claude surfaces. Plugins are Cowork-specific bundles that package skills and connectors for a role or domain. Subagents split large tasks into independent parallel pieces, each with focused context.
  • Scheduled tasks — once a task works reliably, schedule it on a cadence. If your computer is asleep or the app is closed, it runs when you're back and tells you it was delayed.

Prompting

Strong prompts include three parts:

  • Input — where the source material lives
  • Transformation — what to do with it
  • Output — the format, and where it should go

Example: "Turn these meeting notes into a slide deck and save the PowerPoint in this folder."

Good use cases

  • Organize a folder
  • Create a spreadsheet, report, or presentation from notes and sources
  • Synthesize many files into one brief
  • Analyze a batch of transcripts or datasets in parallel
  • Recurring workflows (weekly reports, daily checks) on a schedule

Best practices

  • Start with bounded tasks
  • Ask for the signal — contradictions, outliers, risks — not a generic summary
  • Batch related work
  • Choose the model to fit the task: Opus for complex work, Sonnet as the default, Haiku for light tasks
  • Review every output — polished work can still be wrong

Safety & permissions

Cowork runs in an isolated environment on your computer. It only accesses folders and tools you grant. It respects network restrictions. Permanently deleting files needs your explicit approval. Conversation history is stored locally.

Troubleshooting

  • "Preparing your workspace" — expected on first run or after an update.
  • Task stopped mid-run — the app was quit or the connection dropped. Sleeping your computer is fine; quitting the app pauses the task.
  • Hitting usage limits — Cowork is more compute-intensive than chat.
  • Missing output — check the working folder, verify permissions, or ask Cowork where it wrote the file.
  • Can't share a Cowork project — Cowork projects are local-only, tied to your computer, so there's no share option. Move the work to a Claude.ai chat Project (cloud, shareable) instead, or point Cowork at a shared Drive folder so teammates use the same folder (see Core concepts above).
  1. Point Cowork at a real working folder.
  2. Run one task that produces a real output file.
  3. Install a plugin for your role and inspect one skill file.
  4. Save a repeatable workflow as a skill or pluginized process.
  5. Schedule it once reliable.

See also: Claude Code for the developer/CLI surface, or For Builders if you want to author your own skills and plugins.