Claude Cowork
Claude Cowork shifts the mental model from question-and-answer chat to delegated work. The user describes an outcome, names the inputs and scope, and lets Claude plan and execute across local files, connectors, browser context, and tools. It is built on the same agent runtime as ClaudeAgentSDK.
Key components
- Projects — scoped workspaces with standing instructions, context, scheduled tasks, and accumulated memory
- Connectors — integrations that let Claude search or use external apps (email, calendar, Slack, Drive, CRM)
- Skills — reusable playbooks for recurring processes; avoid re-explaining the same workflow repeatedly
- Plugins — packaged sets of skills, connectors, and subagents around a role or process
- Claude in Chrome — bridges tools and web apps without dedicated connectors
- Scheduled tasks — recurring automations that run at defined intervals
Effective delegation pattern
A strong Cowork task specifies:
- Deliverable — what the output is and what format
- Inputs — which files, folders, or connectors to use (smallest set that contains the work)
- Nuance — domain context, constraints, and preferences
- Review boundary — where human judgment is required before proceeding
- Definition of done — what counts as complete
Use plan mode for complex or risky work before execution begins.
Context discipline
Context is an operating budget. Long sessions increase cost and reasoning noise. Connect only the apps and files the task requires. Clear or compact context when work changes. See ContextBudget.
Human review requirement
Cowork still requires human review. The file may be finished, but facts, business judgment, permanent actions, and sensitive outputs need human inspection. Default to asking before unfamiliar or irreversible actions.
Related
- ContextBudget — managing context across Cowork sessions
- BoundedAgent — the agent design principles behind Cowork tasks
- ClaudeAgentSDK — the SDK runtime Cowork is built on
- ClaudeSDKAndCowork — synthesis of both