Claude Cowork System Prompt: How Anthropic Builds a Digital Coworker

This post is an abridged translation of the Chinese original. It summarizes how Cowork-mode system prompts shape Claude into a “digital coworker” rather than a pure chatbot.

TL;DR

• Cowork is positioned as an execution-capable assistant with guardrails, not just conversation.
• The system prompt encodes safety, tone, and reliability defaults.
• Tool usage and artifacts enable “do work” flows, but are constrained by policy.

1. Product architecture: more than a chatbot

Cowork is described as a system that can:

• interpret tasks
• use tools to execute steps
• produce artifacts as deliverables
• remain policy-compliant under pressure

2. Product matrix: Anthropic’s broader positioning

The prompt reflects an ecosystem view:

• different surfaces (chat vs coding vs tool-enabled experiences)
• different user needs (speed, safety, governance)

3. Safety red lines: boundaries that must not be crossed

Key themes:

• refuse disallowed content
• avoid unsafe instructions
• prioritize user safety and privacy

4. Interaction philosophy: warm, but not overbearing

The prompt drives:

• clarity and helpfulness
• a professional tone
• no excessive flattery

5. User wellbeing: the care boundary

The system prompt encourages:

• supportive responses
• but avoids overstepping into medical/legal certainty

6. Neutral stance: politics and controversial topics

The prompt aims for:

• neutrality and balance
• avoiding manipulation

7. Core tools: the “three engines” of execution

Cowork uses tool patterns to:

• retrieve and verify information
• execute actions
• produce structured outputs

8. Computer use: from talking to doing

The prompt supports:

• navigating UI environments
• taking screenshots/reading logs
• verifying outcomes

9. Knowledge boundaries: be honest about limitations

Principles:

• don’t fabricate
• state uncertainty clearly
• prefer verification

10. Citation norms: traceable sources

The system prompt encourages:

• citing sources when making factual claims
• keeping references clear and attributable

11. Artifacts: renderable deliverables

Artifacts serve as:

• structured outputs (docs, code, reports)
• deliverables the user can reuse