Claude Cowork System Prompt (Full Analysis): How Anthropic Engineers a Digital Coworker

This is an abridged translation of the Chinese long-form analysis. The original includes extensive excerpts and line-by-line discussion; this version keeps the key takeaways and structure.

TL;DR

Cowork is engineered to behave like a reliable “digital coworker”:

• use tools to execute tasks (with constraints)
• be honest about limitations
• stay within strict safety boundaries
• prefer traceable sources and citations for factual claims
• produce artifacts as reusable deliverables

1. Product architecture: more than a chatbot

The system prompt frames Cowork as a work-capable assistant with:

• task execution orientation
• procedural thinking (plan → act → verify)
• structured output habits

2. Product matrix: Anthropic’s broader layout

The prompt implies multiple product surfaces with different constraints and goals, balancing:

• speed vs safety
• autonomy vs governance
• general chat vs specialized workflows

3. Safety red lines: non-negotiable boundaries

The system prompt encodes:

• disallowed content refusals
• privacy preservation
• safe completion policies

4. Interaction philosophy: warm, not excessive

The assistant should:

• be helpful and direct
• avoid manipulation
• avoid empty flattery

5. User wellbeing: care boundaries

The prompt encourages supportive behavior while avoiding overconfidence in:

• medical advice
• legal advice
• high-stakes claims without verification

6. Neutrality: political and controversial topics

The prompt pushes:

• balanced framing
• transparency about uncertainty
• avoiding partisan persuasion

7. Core tools: execution primitives

Tool usage is treated as a disciplined capability:

• gather evidence
• execute steps
• validate results

8. Computer use: from conversation to execution

The assistant can:

• interact with UI environments
• inspect errors and logs
• close the loop with verification

9. Knowledge boundaries: honesty about limits

Rules of thumb:

• don’t hallucinate
• state unknowns clearly
• verify when possible

10. Citation rules: traceability

For factual claims:

• cite sources
• distinguish facts from inference
• keep citations auditable

11. Artifacts: deliverables

Artifacts make outputs reusable:

• reports, plans, docs
• structured summaries
• code and configs